/usr/share/doc/sphinx3/sphinxman_manual.html is in sphinx3-doc 0.8-0ubuntu1.
This file is owned by root:root, with mode 0o644.
The actual contents of the file can be viewed below.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090 2091 2092 2093 2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106 2107 2108 2109 2110 2111 2112 2113 2114 2115 2116 2117 2118 2119 2120 2121 2122 2123 2124 2125 2126 2127 2128 2129 2130 2131 2132 2133 2134 2135 2136 2137 2138 2139 2140 2141 2142 2143 2144 2145 2146 2147 2148 2149 2150 2151 2152 2153 2154 2155 2156 2157 2158 2159 2160 2161 2162 2163 2164 2165 2166 2167 2168 2169 2170 2171 2172 2173 2174 2175 2176 2177 2178 2179 2180 2181 2182 2183 2184 2185 2186 2187 2188 2189 2190 2191 2192 2193 2194 2195 2196 2197 2198 2199 2200 2201 2202 2203 2204 2205 2206 2207 2208 2209 2210 2211 2212 2213 2214 2215 2216 2217 2218 2219 2220 2221 2222 2223 2224 2225 2226 2227 2228 2229 2230 2231 2232 2233 2234 2235 2236 2237 2238 2239 2240 2241 2242 2243 2244 2245 2246 2247 2248 2249 2250 2251 2252 2253 2254 2255 2256 2257 2258 2259 2260 2261 2262 2263 2264 2265 2266 2267 2268 2269 2270 2271 2272 2273 2274 2275 2276 2277 2278 2279 2280 2281 2282 2283 2284 2285 2286 2287 2288 2289 2290 2291 2292 2293 2294 2295 2296 2297 2298 2299 2300 2301 2302 2303 2304 2305 2306 2307 2308 2309 2310 2311 2312 2313 2314 2315 2316 2317 2318 2319 2320 2321 2322 2323 2324 2325 2326 2327 2328 2329 2330 2331 2332 2333 2334 2335 2336 2337 2338 2339 2340 2341 2342 2343 2344 2345 2346 2347 2348 2349 2350 2351 2352 2353 2354 2355 2356 2357 2358 2359 2360 2361 2362 2363 2364 2365 2366 2367 2368 2369 2370 2371 2372 2373 2374 2375 2376 2377 2378 2379 2380 2381 2382 2383 2384 2385 2386 2387 2388 2389 2390 2391 2392 2393 2394 2395 2396 2397 2398 2399 2400 2401 2402 2403 2404 2405 2406 2407 2408 2409 2410 2411 2412 2413 2414 2415 2416 2417 2418 2419 2420 2421 2422 2423 2424 2425 2426 2427 2428 2429 2430 2431 2432 2433 2434 2435 2436 2437 2438 2439 2440 2441 2442 2443 2444 2445 2446 2447 2448 2449 2450 2451 2452 2453 2454 2455 2456 2457 2458 2459 2460 2461 2462 2463 2464 2465 2466 2467 2468 2469 2470 2471 2472 2473 2474 2475 2476 2477 2478 2479 2480 2481 2482 2483 2484 2485 2486 2487 2488 2489 2490 2491 2492 2493 2494 2495 2496 2497 2498 2499 2500 2501 2502 2503 2504 2505 2506 2507 2508 2509 2510 2511 2512 2513 2514 2515 2516 2517 2518 2519 2520 2521 2522 2523 2524 2525 2526 2527 2528 2529 2530 2531 2532 2533 2534 2535 2536 2537 2538 2539 2540 2541 2542 2543 2544 2545 2546 2547 2548 2549 2550 2551 2552 2553 2554 2555 2556 2557 2558 2559 2560 2561 2562 2563 2564 2565 2566 2567 2568 2569 2570 2571 2572 2573 2574 2575 2576 2577 2578 2579 2580 2581 2582 2583 2584 2585 2586 2587 2588 2589 2590 2591 2592 2593 2594 2595 2596 2597 2598 2599 2600 2601 2602 2603 2604 2605 2606 2607 2608 2609 2610 2611 2612 2613 2614 2615 2616 2617 2618 2619 2620 2621 2622 2623 2624 2625 2626 2627 2628 2629 2630 2631 2632 2633 2634 2635 2636 2637 2638 2639 2640 2641 2642 2643 2644 2645 2646 2647 2648 2649 2650 2651 2652 2653 2654 2655 2656 2657 2658 2659 2660 2661 2662 2663 2664 2665 2666 2667 2668 2669 2670 2671 2672 2673 2674 2675 2676 2677 2678 2679 2680 2681 2682 2683 2684 2685 2686 2687 2688 2689 2690 2691 2692 2693 2694 2695 2696 2697 2698 2699 2700 2701 2702 2703 2704 2705 2706 2707 2708 2709 2710 2711 2712 2713 2714 2715 2716 2717 2718 2719 2720 2721 2722 2723 2724 2725 2726 2727 2728 2729 2730 2731 2732 2733 2734 2735 2736 2737 2738 2739 2740 2741 2742 2743 2744 2745 2746 2747 2748 2749 2750 2751 2752 2753 2754 2755 2756 2757 2758 2759 2760 2761 2762 2763 2764 2765 2766 2767 2768 2769 2770 2771 2772 2773 2774 2775 2776 2777 2778 2779 2780 2781 2782 2783 2784 2785 2786 2787 2788 2789 2790 2791 2792 2793 2794 2795 2796 2797 2798 2799 2800 2801 2802 2803 2804 2805 2806 2807 2808 2809 2810 2811 2812 2813 2814 2815 2816 2817 2818 2819 2820 2821 2822 2823 2824 2825 2826 2827 2828 2829 2830 2831 2832 2833 2834 2835 2836 2837 2838 2839 2840 2841 2842 2843 2844 2845 2846 2847 2848 2849 2850 2851 2852 2853 2854 2855 2856 2857 2858 2859 2860 2861 2862 2863 2864 2865 2866 2867 2868 2869 2870 2871 2872 2873 2874 2875 2876 2877 2878 2879 2880 2881 2882 2883 2884 2885 2886 2887 2888 2889 2890 2891 2892 2893 2894 2895 2896 2897 2898 2899 2900 2901 2902 2903 2904 2905 2906 2907 2908 2909 2910 2911 2912 2913 2914 2915 2916 2917 2918 2919 2920 2921 2922 2923 2924 2925 2926 2927 2928 2929 2930 2931 2932 2933 2934 2935 2936 2937 2938 2939 2940 2941 2942 2943 2944 2945 2946 2947 2948 2949 2950 2951 2952 2953 2954 2955 2956 2957 2958 2959 2960 2961 2962 2963 2964 2965 2966 2967 2968 2969 2970 2971 2972 2973 2974 2975 2976 2977 2978 2979 2980 2981 2982 2983 2984 2985 2986 2987 2988 2989 2990 2991 2992 2993 2994 2995 2996 2997 2998 2999 3000 3001 3002 3003 3004 3005 3006 3007 3008 3009 3010 3011 3012 3013 3014 3015 3016 3017 3018 3019 3020 3021 3022 3023 3024 3025 3026 3027 3028 3029 3030 3031 3032 3033 3034 3035 3036 3037 | <!DOCTYPE html PUBLIC "-//W3C//DTD HTML 4.01 Transitional//EN">
<html>
<head>
<title>SphinxTrain Documentation</title>
<style type="text/css">
pre { font-size: medium; background: #f0f8ff; padding: 2mm; border-style: ridge ; color: teal}
code {font-size: medium; color: teal}
</style>
</head>
<body>
<a name="top">INDEX</a>
<p>(This is under construction.)
<!======================================================================>
<ol>
<li><a href="#0"><font color="red">Before you train</font></a>
<ul>
<li><a href="#00">The general-procedure chart</a>
<li><a href="#01">Modeling context-dependent phones with
untied states: some memory requirements</a>
<li><a href="#02">Data preparation</a>
<ul>
<li><a href="#02a">When you have a very small closed vocabulary</a>
</ul>
<li><a href="#03">The set of base and higher order feature vectors</a>
<ul>
<li><a href="#031">Feature streams</a>
</ul>
<li><a href="#04">Force-alignment</a>
</ul>
<!======================================================================>
<li><a href="#2"><font color="red">Training continuous models</font></a>
<ul>
<li><a href="#20">Creating the CI model definition file</a>
<li><a href="#21">Creating the HMM topology file</a>
<li><a href="#22">Flat initialization of CI model parameters</a>
<li><a href="#23">Training CI models</a>
<li><a href="#24">Creating the CD untied model definition file</a>
<li><a href="#25">Flat initialization of CD untied model parameters</a>
<li><a href="#26">Training CD untied models</a>
<li><a href="#27">Building decision trees for parameter sharing</a>
<ul>
<li><a href="#28">Generating the linguistic questions</a>
</ul>
<li><a href="#29">Pruning the decision trees</a>
<li><a href="#30">Creating the CD tied model definition file</a>
<li><a href="#31">Initializing and training cd tied gaussian
mixture models </a>
</ul>
<!=======================================================================>
<li><a href="#3"><font color="red">Training semi-continuous models</font></a>
<ul>
<li><a href="#3b">Vector quantization</a>
<li><a href="#3d">Creating the CI model definition file</a>
<li><a href="#3e">Creating the HMM topology file</a>
<li><a href="#3c">Flat initialization of CI model parameters</a>
<li><a href="#3f">Training CI models</a>
<li><a href="#3g">Creating the CD untied model definition file</a>
<li><a href="#3h">Flat initialization of CD untied model parameters</a>
<li><a href="#3i">Training CD untied models</a>
<li><a href="#3j">Building decision trees for parameter sharing</a>
<ul>
<li><a href="#3k">Generating the linguistic questions</a>
</ul>
<li><a href="#3l">Pruning the decision trees</a>
<li><a href="#3m">Creating the CD tied model definition file</a>
<li><a href="#3n">Initializing and training cd tied models </a>
<li><a href="#3a">Deleted interpolation</a>
</ul>
<li><a href="#4">SPHINX2 data and model formats</a>
<li><a href="#4b">SPHINX3 data and model formats</a>
<li><a href="#5">Training multilingual models</a>
<li><a href="#6">The training lexicon</a>
<li><a href="#7">Converting SPHINX3 format models to SPHINX2 format</a>
<li><a href="#8">Updating or adapting existing model sets</a>
<li><a href="#9">Using the SPHINX-III decoder with semi-continuous and
continuous models</a>
</ol>
<hr>
<!=========================================================================>
This part of the manual describes the procedure(s) for training acoustic
models using the Sphinx3 trainer. General training procedures are described
first, and followed by more detailed descriptions of the programs and scripts
used, and the analysis of their logs and other outputs.
<p>
<a name="0"></a>
<a name="00"></a>
<!------------------------------------------------------------------------->
<center><h4><font color="red">BEFORE YOU TRAIN</font></h4></center>
<TABLE width="100%" bgcolor="#ffffff">
<td>THE GENERAL-PROCEDURE CHART</td>
</table>
<!------------------------------------------------------------------------->
<pre>
Training chart for the
sphinx2 trainer
=========================
OBSOLETE
(The sphinx2 trainer is no longer used in CMU)
Training chart for the
sphinx3 trainer
=========================
type of model
|
----------------------------------
| |
CONTINUOUS SEMI-CONTINUOUS
| |
| vector-qunatization
| |
----------------------------------
|...make ci mdef
|...flat_initialize CI models
training CI models
|...make cd untied mdef
|...initialize
|
training CD untied models
|
|
|
decision tree building
|...prune trees
|...tie states
|...make cd tied mdef
training CD tied models
|
|
recursive ----------------------------------
gaussian splitting.. | |
continuous models semi-continuous models
| |
| |
----------- |
| | deleted interpolation
decode with ADAPT |
sphinx3 | |---ADAPT
decoder <------- | |
----------------
make cd tied mdef ... | .............|
with decode dict and | convert to
pruned trees | sphinx2
decode with |
sphinx3 |
decoder |
|
decode with
sphinx2
decoder
(currently opensource
and restricted to
working with sampling
rates 8khz and 16khz.
Once the s3 trainer is
released, this will have
to change to allow
people who train with
different sampling rates
to use this decoder)
</pre>
<p>
<a href="#top">back to index</a>
<hr>
<a name="01"></a>
<!------------------------------------------------------------------------->
<center><h4><font color="red">BEFORE YOU TRAIN</font></h4></center>
<TABLE width="100%" bgcolor="#ffffff">
<td>MODELING CONTEXT-DEPENDENT PHONES WITH UNTIED STATES: SOME MEMORY REQUIREMENTS</td>
</table>
<!------------------------------------------------------------------------->
<p>
Modeling Context-dependent phones (ex. triphones) with untied states requires
the largest amount of hardware resources. Take a moment to check if you
have enough. The resources required depend on the type of model you are
going to train, the dimensionality and configuration of your feature vectors,
and the number of states in the HMMs.
<p>
<b><u>Semi-continuous models</u></b>
<p>
To train 5-state/HMM models for 10,000 triphones:
<pre>
5 states/triphone = 50,000 states
For a 4-stream feature-set, each = 1024 floating point numbers/state
state has a total of 4*256 mixture
weights
= 205Mb buffer for 50,000 states
</pre>
Corresponding to each of the four feature streams, there are 256 means
and 256 variances in the codebook. ALL these, and ALL the mixture weights
and transition matrices are loaded in into the RAM, and during training
an additional buffer of equal size is allocated to store intermediate
results. These are later written out into the hard disk when the calculations
for the current training iteration are complete. Note that
there are as many transition matrices as you have phones (40-50
for the English language, depending on your dictionary) All this amounts
to allocating well over 400 Mb of RAM.
<p>
This is a bottleneck for machines with smaller memory. No matter how large
your training corpus is, you can actually train only about 10,000 triphones
at the cd-untied stage if you have ~400 Mb of RAM (A 100 hour broadcast
news corpus typically has 40,000 triphones). You could train more if
your machine is capable of handling the memory demands effectively (this
could be done, for example, by having a large amount of swap space). If
you are training on multiple machines, *each* will require this much memory.
In addition, at the end of each iteration, you have to transmit all buffers
to a single machine that performs the norm. Networking issues need to
be considered here.
<p>
The cd-untied models are used to build trees. The number of triphones you
train at this stage directly affects the quality of the trees, which would
have to be built using fewer triphones than are actually present in the
training set if you do not have enough memory.
<p>
<b><u>Continuous models</u></b>
<p>
For 10,000 triphones:
<pre>
5 states/triphone = 50,000 states
39 means (assuming a
39-component feature
vector) and 39
variances per state = 79 floating points per state
= 15.8Mb buffer for 50,000 states
</pre>
Thus we can train 12 times as many triphones as we can when we have
semicontinuous models for the same amount of memory. Since we can use
more triphones to train (and hence more information) the decision trees
are better, and eventually result in better recognition performance.
<p>
<a href="#top">back to index</a>
<hr>
<a name="02"></a>
<!------------------------------------------------------------------------->
<center><h4><font color="red">BEFORE YOU TRAIN</font></h4></center>
<TABLE width="100%" bgcolor="#ffffff">
<td>DATA PREPARATION</td>
</table>
<!------------------------------------------------------------------------->
<p>
You will need the following files to begin training:
<ol>
<li>A set of <b>feature files</b> computed from the audio training data, one each for every
recording you have in the training corpus. Each recording can be transformed
into a sequence of feature vectors using a front-end executable provided with
the SPHIN-III training package. Each front-end executable provided performs
a different analysis of the speech signals and computes a different type
of feature.
<p>
<li>A <b> control file</b> containing the list of feature-set filenames with
full paths to them. An example of the entries in this file:
<pre>
dir/subdir1/utt1
dir/subdir1/utt2
dir/subdir2/utt3
</pre>
Note that the extensions are not given. They will be provided separately
to the trainer. It is a good idea to give unique names to all feature
files, even if
including the full paths seems to make each entry in the control file
unique. You will find later that this provides a lot of flexibility
for doing many things.
<p>
<li>A <b>transcript file</b> in which the transcripts corresponding to the
feature files are listed in exactly the same order as the feature
filenames in the control file.
<p>
<li>A <b>main dictionary</b> which has all acoustic events and words in
the transcripts mapped onto the acoustic units you want to train.
Redundancy in the form of extra words is permitted. The dictionary
must have all alternate pronunciations marked with paranthesized serial
numbers starting from (2) for the second pronunciation. The marker
(1) is omitted. Here's an example:
<pre>
DIRECTING D AY R EH K T I ng
DIRECTING(2) D ER EH K T I ng
DIRECTING(3) D I R EH K T I ng
</pre>
<p>
<li>A <b>filler dictionary</b>, which usually lists the
non-speech events as "words" and maps them to user_defined phones.
This dictionary must at least have the entries
<pre>
<s> SIL
<sil> SIL
</s> SIL
</pre>
The entries stand for
<pre>
<s> : begining-utterance silence
<sil> : within-utterance silence
</s> : end-utterance silence
</pre>
Note that the words <s>, </s> and <sil> are treated as special
words and are required to be present in the filler dictionary. At least
one of these must be mapped on to a phone called "SIL". The phone
SIL is treated in a special manner and is required to be present.
The sphinx expects you to name the acoustic events corresponding to your
general background condition as SIL. For clean speech these events may
actually be silences, but for noisy speech these may be the most general
kind of background noise that prevails in the database.
Other noises can then be modelled by phones defined by the user.
<p>
During training SIL replaces every phone flanked by "+" as the context
for adjacent phones. The phones flanked by "+" are only modeled as CI
phones and are not used as contexts for triphones. If you do not want
this to happen you may map your fillers to phones that are not flanked
by "+".
<p>
<li>A <b>phonelist</b>, which is a list of all acoustic units that you want to
train models for. The SPHINX does not permit you to have units
other than those in your dictionaries. All units in your
two dictionaries must be listed here. In other words, your phonelist
must have exactly the same units used in your dictionaries, no more and no
less. Each phone must be listed on a separate line in the file, begining from
the left, with no extra spaces after the phone. an example:
<pre>
AA
AE
OW
B
CH
</pre>
</ol>
Here's a quick checklist to verify your data preparation before you train:
<ol>
<li> Are all the transcript words in the dictionary/filler dictionary?
<li> Make sure that the size of transcript matches the .ctl file.
<li> Check the boundaries defined in the .ctl file to make sure they exist
ie, you have all the frames that are listed in the control file
<li> Verify the phonelist against the dictionary and fillerdict
</ol>
<p>
<a name="02a"></a>
<b><u>When you have a very small closed vocabulary (50-60 words)</u></b>
<p>
If you have only about 50-60 words in your vocabulary, and if your entire
test data vocabulary is covered by the training data, then you are probably
better off training word models rather than phone models. To do this,
simply define the phoneset as your set of words themselves and
have a dictionary that maps each word to itself and train.
Also, use a lesser number of fillers, and if you do need to train phone models
make sure that each of your tied states has enough counts (at least
5 or 10 instances of each).
<p>
<hr>
<a name="03"></a>
<!------------------------------------------------------------------------->
<center><h4><font color="red">BEFORE YOU TRAIN</font></h4></center>
<TABLE width="100%" bgcolor="#ffffff">
<td>THE SET OF BASE AND HIGHER ORDER FEATURE VECTORS</td>
</table>
<!------------------------------------------------------------------------->
<p>
The set of feature vectors you have computed using the Sphinx front-end
executable is called the set of <b>base</b> feature vectors. This
set of base features can be extended to include what are called <b>
higher order</b> features. Some common extensions are
<ol type="a">
<li>The set of difference vectors, where the component-wise
difference between *some* succeeding and preceding vector(s),
used to get an estimate of the slope or trend at the current
time instant, are the "extension" of the current vector.
These are called "delta" features. A more appropriate name
would be the "trend" features.
<li>The set of difference vectors of difference vectors. The
component-wise difference between the succeeding and preceding
"delta" vectors are the "extension" of the current vector. These
are called "double delta" features
<li>The set of difference vectors, where the component-wise
difference between the n^th succeeding and n^th preceding vector
are the "extension" of the current vector. These are called
"long-term delta" features, differing from the "delta" features
in just that they capture trends over a longer window of time.
<li>The vector composed of the first elements of the current vector
and the first elements of some of the above "extension" vectors.
This is called the "power" feature, and its dimensionality is
less than or equal to the total number of feature types you
consider.
</ol>
<p>
<a name="031"></a>
<!------------------------------------------------------------------------->
<b><u>Feature streams</u></b>
<!------------------------------------------------------------------------->
<p>
In semi-continuous models, it is a usual practice to keep the identities of
the base vectors and their "extension" vectors separate. Each such set is
called a "feature stream". You must specify how many feature
streams you want to use in your semi-continuous models and how you want them
arranged.
The feature-set options currently supported by the Sphinx are:
<p>
c/1..L-1/,d/1..L-1/,c/0/d/0/dd/0/,dd/1..L-1/ : read this as
cepstra/second to last component,<br>
deltacepstra/second to last component,<br>
cepstra/first component deltacepstra/first component doubledeltacepstra/first component, <br>
doubledeltacepstra/second to last component
<p>
This is a 4-stream feature vector used mostly in semi-continuous models.
There is no particular advantage to this arrangement - any permutation
would give you the same models, with parameters written in different
orders.
<p>
Here's something that's not obvious from the notation used for the
4-stream feature set: the dimensionality of the 4-stream feature vector
is 12cepstra+24deltas+3powerterms+12doubledeltas
<p>
the deltas are computed as the difference between the cepstra two frames
removed on either side of the current frame (12 of these), followed by
the difference between the cepstra four frames
removed on either side of the current frame (12 of these). The power stream
uses the first component of the two-frames-removed deltas, computed using C<sub>0</sub>.
<p>
(more to come....)
<hr>
<a name="2"></a>
<a name="20"></a>
<!------------------------------------------------------------------------->
<center><h4><font color="red">TRAINING CONTINUOUS MODELS</font></h4></center>
<TABLE width="100%" bgcolor="#ffffff">
<td>CREATING THE CI MODEL DEFINITION FILE</td>
</table>
<!------------------------------------------------------------------------->
<p>
The first step is to prepare a <b>model definition</b> file for the context
independent (CI) phones. The function of a model definition file is to
define or provide a unique numerical identity to every state of every HMM
that you are going to train, and to provide an order which will be followed
in writing out the model parameters in the model parameter files. During
the training, the states are referenced only by these numbers. The model
definition file thus partly specifies your <b>model architecture</b> and
is thus usually stored in a directory named "model_architecture". You are of
course free to store it where you please, unless you are running the
training scripts provided with the SPHINX-III package.
<p>
To generate this <b> CI model definition file</b>, use the executable <b><font
color="green">mk_model_def</font></b> with the following flag settings:
<p>
<table border="1">
<tr><td valign="top"> FLAG </td><td> DESCRIPTION </td></tr>
<tr><td valign="top"> -phonelstfn </td><td> phonelist </td></tr>
<tr><td valign="top"> -moddeffn </td><td> name of the CI model definition file that you want to create. Full path must be provided</td></tr>
<tr><td valign="top"> -n_state_pm </td><td> number of states per HMM in the
models that you want to train. If you want to train 3 state HMMs, write "3"
here, without the double quotes</td></tr>
</table>
<p>
Pipe the standard output into a log file <b>ci_mdef.log</b> (say).
If you have listed only three phones in your phonelist,
and specify that you want to build three state HMMs for each
of these phones, then your model-definition file will look like this:
<pre>
# Generated by <path_to_binary>/mk_model_def on Thu Aug 10 14:57:15 2000
0.3
3 n_base
0 n_tri
12 n_state_map
9 n_tied_state
9 n_tied_ci_state
3 n_tied_tmat
#
# Columns definitions
#base lft rt p attrib tmat ...state id's ...
SIL - - - filler 0 0 1 2 N
A - - - n/a 1 3 4 5 N
B - - - n/a 2 6 7 8 N
The # lines are simply comments. The rest of the variables mean the following:
n_base : no. of phones (also called "base" phones) that you have
n_tri : no. of triphones (we will explain this later)
n_state_map : Total no. of HMM states (emitting and non-emitting)
The Sphinx appends an extra terminal non-emitting state
to every HMM, hence for 3 phones, each specified by
the user to be modeled by a 3-state HMM, this number
will be 3phones*4states = 12
n_tied_state: no. of states of all phones after state-sharing is done.
We do not share states at this stage. Hence this number is the
same as the total number of emitting states, 3*3=9
n_tied_ci_state:no. of states for your "base" phones after state-sharing
is done. At this stage, the number of "base" phones is
the same as the number of "all" phones that you are modeling.
This number is thus again the total number of emitting
states, 3*3=9
n_tied_tmat :The HMM for each CI phone has a transition probability matrix
associated it. This is the total number of transition
matrices for the given set of models. In this case, this
number is 3.
Columns definitions: The following columns are defined:
base : name of each phone
lft : left-context of the phone (- if none)
rt : right-context of the phone (- if none)
p : position of a triphone (not required at this stage)
attrib: attribute of phone. In the phone list, if the phone is "SIL",
or if the phone is enclosed by "+", as in "+BANG+", the sphinx
understands these phones to be non-speech events. These are
also called "filler" phones, and the attribute "filler" is
assigned to each such phone. The base phones have no special
attributes, and hence are labelled as "n/a", standing for
"no attribute"
tmat : the id of the transition matrix associated with the phone
state id's : the ids of the HMM states associated with any phone. This list
is terminated by an "N" which stands for a non-emitting
state. No id is assigned to it. However, it exists, and is
listed.
</pre>
<a name="21"></a>
<!------------------------------------------------------------------------->
<center><h4><font color="red">TRAINING CONTINUOUS MODELS</font></h4></center>
<TABLE width="100%" bgcolor="#ffffff">
<td>CREATING THE HMM TOPOLOGY FILE</td>
</table>
<!------------------------------------------------------------------------->
<p>
The HMM topology file consists of a matrix with boolean entries, each entry
indiactes whether a specific transition from state=row_number to
state=column_number is permitted in the HMMs or not. For example a
3-state HMM with no skips permitted beteen states would have a topology
file with the following entries:
<pre>
4
1.0 1.0 0.0 0.0
0.0 1.0 1.0 0.0
0.0 0.0 1.0 1.0
</pre>
The number 4 is total the number of sates in an HMMs. The SPHINX
automatically appends a fourth non-emitting terminating state to the 3
state HMM. The first entry of 1.0 means that a transition from state 1 to
state 1 (itself) is permitted. Accordingly, the transition matrix estimated
for any phone would have a "transition-probability" in place of this
boolean entry. Where the entry is 0.0, the corresponding transition
probability will not be estimated (will be 0).
<p>
You can either write out the topology file manually, or
use the script script make_topology.pl provided with the SPHINX package to
do this. The script needs the following arguments:
<pre>
states_per_hmm : this is merely an integer specifying the
number of states per hmm
skipstate : "yes" or "no" depending on whether you
want the HMMs to have skipped state transitions
or not.
</pre>
<p>
Note that the topology file
is common for all HMMs and is a single file containing the topology
definition matrix. This file also defines your model architecture and is
usually placed in the model_architecture directory. This is however
optional, but recommended. If you are running scripts from the SPHINX
training package, you will find the file created in the model_architecture
directory.
<p>
<a name="22"></a>
<!------------------------------------------------------------------------->
<center><h4><font color="red">TRAINING CONTINUOUS MODELS</font></h4></center>
<TABLE width="100%" bgcolor="#ffffff">
<td>FLAT INITIALIZATION OF CI MODEL PARAMETERS</td>
</table>
<!------------------------------------------------------------------------->
<p>
CI models consist of 4 parameter files :
<ul>
<li><b>mixture_weights</b>: the weights given to every Gaussian in the Gaussian mixture corresponding to a state
<li><b>transition_matrices</b>: the matrix of state transition probabilities
<li><b>means</b>: means of all Gaussians
<li><b>variances</b>: variances of all Gaussians
</ul>
<p>
To begin training the CI models, each of these files must have some initial
entries, ie, they must be "initialized". The mixture_weights and
transition_matrices are initialized using the executable <b><font
color="green">mk_flat</font></b>. It needs the following arguments:
<p>
<table border="1">
<tr><td> FLAG </td><td> DESCRIPTION </td></tr>
<tr><td> -moddeffn </td><td> CI model definition file </td></tr>
<tr><td> -topo </td><td> HMM topology file </td></tr>
<tr><td> -mixwfn </td><td> file in which you want to write the
initialized mixture weights </td></tr>
<tr><td> -tmatfn </td><td> file in which you want to write the
initialized transition matrices </td></tr>
<tr><td> -nstream </td><td> number of independent feature streams, for
continuous models this number should be set to "1", without the double quotes
</td></tr>
<tr><td> -ndensity </td><td> number of Gaussians modeling each state. For CI
models, this number should be set to "1" </td></tr>
</table>
<p>
To initialize the means and variances, global values of these parameters are
first estimated and then copied into appropriate positions in the parameter
files.
The global mean is computed using
all the vectors you have in your feature files. This is usually
a very large number, so the job is divided into many parts. At this
stage you tell the Sphinx how many parts you want it to divide this
operation into (depending on the computing facilities you have)
and the Sphinx "accumulates" or gathers up the vectors for each part
separately and writes it into an intermediate buffer on your machine.
The executable <b><font color="green">init_gau</font></b> is used
for this purpose. It needs the following arguments:
<p>
<table border="1">
<tr><td> FLAG </td><td> DESCRIPTION </td></tr>
<tr><td> -accumdir </td><td> directory in which you want to write the
intermediate buffers </td></tr>
<tr><td> -ctlfn </td><td> control file </td></tr>
<tr><td> -part </td><td> part number </td></tr>
<tr><td> -npart </td><td> total number of parts </td></tr>
<tr><td> -cepdir </td><td> path to feature files - this will be appended
before all paths given in the control file </td></tr>
<tr><td> -cepext </td><td> filename extension of feature files, eg. "mfc"
for files called a/b/c.mfc. Double quotes are not needed </td></tr>
<tr><td> -feat </td><td> type of feature </td></tr>
<tr><td> -ceplen </td><td> dimensionality of base feature vectors </td></tr>
<tr><td> -agc </td><td> automatic gain control factor(max/none) </td></tr>
<tr><td> -cmn </td><td> cepstral mean normalization(yes/no) </td></tr>
<tr><td> -varnorm </td><td> variance normalization(yes/no) </td></tr>
</table>
<p>
Once the buffers are written, the contents of the buffers are
"normalized" or used to compute a global mean value for the feature vectors.
This is done using the executable <b><font color="green">norm</font></b> with
the following flag settings:
<p>
<table border="1">
<tr><td> FLAG </td><td> DESCRIPTION </td></tr>
<tr><td> -accumdir </td><td> buffer directory </td></tr>
<tr><td> -meanfn </td><td> file in which you want to write the global mean </td></tr>
<tr><td> -feat </td><td> type of feature </td></tr>
<tr><td> -ceplen </td><td> dimensionality of base feature vector </td></tr>
</table>
<p>
The next step is to "accumulate" the vectors for computing a global variance
value. The executable <b><font color="green">init_gau</font></b>, when
called a second time around, takes the value of the global mean and
collects a set of (vector-globalmean)<sup>2</sup> values for the
entire data set. This time around, this executable needs the following
arguments:
<p>
<table border="1">
<tr><td> FLAG </td><td> DESCRIPTION </td></tr>
<tr><td> -accumdir </td><td> directory in which you want to write the
intermediate buffers </td></tr>
<tr><td> -meanfn </td><td> globalmean file </td></tr>
<tr><td> -ctlfn </td><td> control file </td></tr>
<tr><td> -part </td><td> part number </td></tr>
<tr><td> -npart </td><td> total number of parts </td></tr>
<tr><td> -cepdir </td><td> path to feature files - this will be appended
before all paths given in the control file </td></tr>
<tr><td> -cepext </td><td> filename extension of feature files, eg. "mfc"
for files called a/b/c.mfc. Double quotes are not needed </td></tr>
<tr><td> -feat </td><td> type of feature </td></tr>
<tr><td> -ceplen </td><td> dimensionality of base feature vectors </td></tr>
<tr><td> -agc </td><td> automatic gain control factor(max/none) </td></tr>
<tr><td> -cmn </td><td> cepstral mean normalization(yes/no) </td></tr>
<tr><td> -varnorm </td><td> variance normalization(yes/no) </td></tr>
</table>
<p>
Again, once the buffers are written, the contents of the buffers are
"normalized" or used to compute a global variance value for the feature
vectors. This is again done using the executable
<b><font color="green">norm</font></b> with
the following flag settings:
<p>
<table border="1">
<tr><td> FLAG </td><td> DESCRIPTION </td></tr>
<tr><td> -accumdir </td><td> buffer directory </td></tr>
<tr><td> -varfn </td><td> file in which you want to write the global variance
</td></tr>
<tr><td> -feat </td><td> type of feature </td></tr>
<tr><td> -ceplen </td><td> dimensionality of base feature vector </td></tr>
</table>
<p>
Once the global mean and global variance are computed, they have to be
copied into the means and variances of every
state of each of the HMMs. The global mean is written
into appropriate state positions in a <b>means</b> file while the
global variance is written into appropriate state positions in a <b>variances</b> file. If you are using the scripts provided with the SPHINX package,
you will find these files with "flatinitial" as part of its name in
the model_parameters directory.
<p>
The flat <b>means</b> and <b>variances</b> file can be created using the
executable <b><font color="green">cp_parm</font></b>. In order to be able
to use this executable you will have to create a <b>copyoperations
map</b> file which is
a two-column file, with the left column id-ing the state *to* which
the global value has to be copied, and the right column id-ing the state
*from* which it has to be copied. If there are "nphones"
CI phones and each state has "nEstate_per_hmm" EMITTING states, there
will be ntotal_Estates = nphones * nEstate_per_hmm lines in the
copyoperations map file; the state id-s (on the left column) run from 0
thru (ntotal_Estates - 1). Here is an example
for a 3-state hmm (nEstate_per_hmm = 3) for two phones (nphones = 2)
(ntotal_Estates = 6; so, state ids would vary from 0-5):
<pre>
0 0
1 0
2 0
3 0
4 0
5 0
</pre>
<b><font color="green">cp_parm</font></b> requires the following arguments.
<p>
<table border="1">
<tr><td> FLAG </td><td> DESCRIPTION </td></tr>
<tr><td> -cpopsfn </td><td> copyoperations map file </td></tr>
<tr><td> -igaufn </td><td> input global mean (or variance) file </td></tr>
<tr><td> -ncbout </td><td> number of phones times the number of states per
HMM (ie, total number of states) </td></tr>
<tr><td> -ogaufn </td><td> output initialized means (or variances) file </td></tr>
</table>
<p>
<b><font color="green">cp_parm</font></b> has to be run twice, once for
copying the means, and once for copying the variances. This completes the
initialization process for CI training.
<hr>
<p>
<a name="23"></a>
<!------------------------------------------------------------------------->
<center><h4><font color="red">TRAINING CONTINUOUS MODELS</font></h4></center>
<TABLE width="100%" bgcolor="#ffffff">
<td>TRAINING CONTEXT INDEPENDENT MODELS</td>
</table>
<!------------------------------------------------------------------------->
<p>
Once the flat initialization is done, you are ready to begin training the acoustic models
for the base or "context-independent" or CI phones. This step is
called CI-training. In CI-training, the flat-initialized models
are re-estimated through the forward-backward re-estimation algorithm
called the Baum-Welch algorithm. This is an iterative re-estimation
process, so you have to run many "passes" of the Baum-Welch re-estimation
over your training data. Each of these passes, or iterations, results in a
slightly better set of models for the CI phones. However, since the
objective function maximized in each of theses passes is the likelihood,
too many iterations would ultimately result in models which fit very
closely to the training data. you might not want this to happen for many
reasons. Typically, 5-8 iterations of Baum-Welch are sufficient for
getting good estimates of the CI models. You can automatically determine the
number of iterations that you need by looking at the total likelihood of the
training data at the end of the first iteration and deciding on a
"convergence ratio" of likelihoods. This is simply the ratio of the
total likelihood in the current iteration to that of the previous iteration.
As the models get more and more fitted to the training data in each
iteration, the training data likelihoods typically increase monotonically.
The convergence ratio is therefore a small positive number. The convergence
ratio becomes smaller and smaller as the iterations progress, since each
time the current models are a little less different from the previous ones.
Convergence ratios are data and task specific, but typical values at which
you may stop the Baum-Welch iterations for your CI training may
range from 0.1-0.001. When the models are variance-normalized, the convergence ratios are much smaller.
<p>
The executable used to run a Buam-Welch iteration is called "bw", and takes the
following example arguments for training continuous CI models:
<table border="1", noshade>
<tr><td align="center"> FLAG </td> <td align="center"> DESCRIPTION </td> </tr>
<tr><td valign="top"> -moddeffn</td> <td> model definition file for CI phones </td></tr>
<tr><td valign="top"> -ts2cbfn </td> <td> this flag should be set to ".cont." if
you are training continuous models, and to
".semi." if you are <a href="#3f">training semi-continuous
models</a>, without the double quotes </td></tr>
<tr><td valign="top"> -mixwfn </td> <td> name of the file in which the
mixture-weights from the previous iteration are stored. Full path must be
provided</td></tr>
<tr><td valign="top"> -mwfloor </td> <td> Floor value for the mixture weights. Any number
below the floor value is set to the floor
value.</td></tr>
<tr><td valign="top"> -tmatfn </td> <td> name of the file in which
the transition matrices from the previous iteration are stored.
Full path must be provided</td></tr>
<tr><td valign="top"> -meanfn </td> <td>name of the file in which
the means from the previous iteration are stored. Full path must be
provided</td></tr>
<tr><td valign="top"> -varfn </td> <td>name of the file in which
the variances fromt he previous iteration are stored.
Full path must be provided</td></tr>
<tr><td valign="top"> -dictfn </td> <td> Dictionary </td></tr>
<tr><td valign="top"> -fdictfn </td> <td> Filler dictionary</td></tr>
<tr><td valign="top"> -ctlfn </td> <td> control file </td></tr>
<tr><td valign="top"> -part </td> <td> You can split the training into N equal parts by
setting a flag. If there are M utterances in your control file, then this will
enable you to run the training separately on each (M/N)<sup>th</sup> part. This
flag may be set to specify which of these parts you want to currently train
on. As an example, if your total number of parts is 3, this flag can take
one of the values 1,2 or 3</td></tr>
<tr><td valign="top"> -npart </td><td> number of parts in which you have split
the training </td></tr>
<tr><td valign="top"> -cepdir </td><td> directory where your feature files are
stored</td></tr>
<tr><td valign="top"> -cepext </td><td> the extension that comes after the name listed
in the control file. For example, you may have a file called a/b/c.d and
may have listed a/b/c in your control file. Then this flag must be given the
argument "d", without the double quotes or the dot before it </td></tr>
<tr><td valign="top"> -lsnfn </td><td> name of the transcript file </td></tr>
<tr><td valign="top"> -accumdir </td><td> Intermediate results from each part of your training will be written in this directory. If you have T means to estimate, then
the size of the mean buffer from the current part of your training will
be T*4 bytes (say). There will likewise be a variance buffer, a buffer for
mixture weights, and a buffer for transition matrices</td></tr>
<tr><td valign="top"> -varfloor </td><td> minimum variance value allowed </td></tr>
<tr><td valign="top"> -topn </td><td> no. of gaussians to consider for computing the likelihood of each state. For example, if you have 8 gaussians/state models and topn is 4, then the 4 most
likely gaussian are used. </td></tr>
<tr><td valign="top"> -abeam </td><td> forward beamwidth</td></tr>
<tr><td valign="top"> -bbeam </td><td> backward beamwidth</td></tr>
<tr><td valign="top"> -agc </td><td> automatic gain control</td></tr>
<tr><td valign="top"> -cmn </td><td> cepstral mean normalization</td></tr>
<tr><td valign="top"> -varnorm </td><td> variance normalization</td></tr>
<tr><td valign="top"> -meanreest </td><td> mean re-estimation</td></tr>
<tr><td valign="top"> -varreest </td><td> variance re-estimation</td></tr>
<tr><td valign="top"> -2passvar </td><td> Setting this flag to "yes" lets bw
use the previous means in the estimation of the variance. The current variance
is then estimated as E[(x - prev_mean)<sup>2</sup>]. If this flag is set to
"no" the current estimate of the means are used to estimate variances. This
requires the estimation of variance as E[x<sup>2</sup>] - (E[x])<sup>2</sup>,
an unstable estimator that sometimes results in negative estimates of the
variance due to arithmetic imprecision</td></tr>
<tr><td valign="top"> -tmatreest </td><td> re-estimate transition matrices or not</td></tr>
<tr><td valign="top"> -feat </td><td> feature configuration</td></tr>
<tr><td valign="top"> -ceplen </td><td> length of basic feature vector</td></tr>
</table>
<p>
If you have run the training in many parts, or even if you have run the
training in one part, the executable for Baum-Welch described above generates
only intermediate buffer(s). The final model parameters, namely the
means, variances, mixture-weights and transition matrices, have to be
estimated using the values stored in these buffers. This is done by the
executable called "norm", which takes the following arguments:
<table border="1", noshade>
<tr><td align="center"> FLAG </td> <td align="center"> DESCRIPTION </td></tr>
<tr><td valign="top"> -accumdir </td><td> Intermediate buffer directory</td></tr>
<tr><td valign="top"> -feat </td><td> feature configuration</td></tr>
<tr><td valign="top"> -mixwfn </td> <td> name of the file in which you want
to write the mixture weights.
Full path must be provided</td></tr>
<tr><td valign="top"> -tmatfn </td> <td> name of the file in which you want to write
the transition matrices. Full path must be
provided</td></tr>
<tr><td valign="top"> -meanfn </td> <td>name of the file in which you want to write
the means. Full path must be
provided</td></tr>
<tr><td valign="top"> -varfn </td> <td>name of the file in which you want to write
the variances. Full path must be
provided</td></tr>
<tr><td valign="top"> -ceplen </td><td> length of basic feature vector</td></tr>
</table>
If you have not re-estimated any of the model parameters in the bw step, then
the corresponding flag must be omitted from the argument given to the
norm executable. The executable will otherwise try to read a non-existent
buffer from the buffer directory and will not go through. Thus if you have
set -meanreest to be "no" in the argument for bw, then the flag -meanfn must
not be given in the argument for norm. This is useful mostly during adaptation.
<p>
Iterations of baum-welch and norm finally result CI models. The iterations
can be stopped once the likelihood on the training data converges. The
model parameters computed by norm in the final iteration are now used
to initialize the models for context-dependent phones (triphones) with
untied states. This is the next major step of the training process. We
refer to the process of training triphones HMMs with untied states as the
"CD untied training".
<hr>
<p>
<a name="24"></a>
<!------------------------------------------------------------------------->
<center><h4><font color="red">TRAINING CONTINUOUS MODELS</font></h4></center>
<TABLE width="100%" bgcolor="#ffffff">
<td>CREATING THE CD UNTIED MODEL DEFINITION FILE</td>
</table>
<!------------------------------------------------------------------------->
<p>
The next step is the CD-untied training, in which HMMs are trained
for all context-dependent phones (usually triphones) that are seen in the
training corpus. For the CD-untied training, we first need to
to generate a model definition file for all the triphones
occuring in the training set. This is done in several steps:
<ul>
First, a list of all triphones possible in the vocabulary is generated
from the dictionary. To get this complete list of triphones from the
dictionary, it is first necessary to write the list of phones in the
following format:
<pre>
phone1 0 0 0 0
phone2 0 0 0 0
phone3 0 0 0 0
phone4 0 0 0 0
...
</pre>
The phonelist used for the CI training must be used to generate this, and
the order in which the phones are listed must be the same.
<p>
Next, a temporary dictionary is generated, which has all words except
the filler words (words enclosed in ++()++ ). The entry
<pre>
SIL SIL
</pre>
must be
added to this temporary dictionary, and the dictionary must be sorted in
alphabetical order. The program "quick_count" provided with the SPHINX-III
package can now be used to generate the list of all possible triphones from
the temporary dictionary. It takes the following arguments:
<p>
<table border="1">
<tr><td> FLAG </td><td> DESCRIPTION </td></tr>
<tr><td valign="top"> -q </td><td> mandatory flag to tell quick_count
to consider all word pairs while constructing triphone list </td></tr>
<tr><td> -p </td><td> formatted phonelist </td></tr>
<tr><td> -b </td><td> temporary dictionary </td></tr>
<tr><td> -o </td><td> output triphone list </td></tr>
</table>
<p>
Here is a typical output from quick_count
<pre>
AA(AA,AA)s 1
AA(AA,AE)b 1
AA(AA,AO)1 1
AA(AA,AW)e 1
</pre>
<p>
The "1" in AA(AA,AO)1 indicates that this is a word-internal triphone. This
is a carry over from Sphinx-II. The output from quick_count has to be
now written into the following format:
<pre>
AA AA AA s
AA AA AE b
AA AA AO i
AA AA AW e
</pre>
This can be done by simply replacing "(", ",", and ")" in the
output of quick_count by a space and printing only the first four
columns.
While
doing so, all instances of " 1" must be replaced by " i". To the top of
the resulting file the list of CI phones must be appened in the following
format
<pre>
AA - - -
AE - - -
AO - - -
AW - - -
..
..
AA AA AA s
AA AA AE b
AA AA AO i
AA AA AW e
</pre>
<br>
<em>For example, if the output of the quick_count is stored in
a file named "quick_count.out", the following perl command will
generate the phone list in the desired form.
perl -nae '$F[0] =~ s/\(|\)|\,/ /g; $F[0] =~ s/1/i/g; print $F[0]; if ($F[0] =~ /\s+$/){print "i"}; print "\n"' quick_count.out</em>
<p>
The above list of triphones (and phones) is converted to the model
definition file that lists all possible triphones from the dictionary. The
program used from this is "mk_model_def" with the following arguments
<table border="1">
<tr><td> FLAG </td><td> DESCRIPTION </td></tr>
<tr><td> -moddeffn </td><td> model definition file with all possible triphones(alltriphones_mdef)to be written</td></tr>
<tr><td> -phonelstfn </td><td> list of all triphones </td></tr>
<tr><td> -n_state_pm </td></tr> number of states per HMM </td></tr>
</table>
In the next step we find the number of times each of the triphones
listed in the alltriphones_mdef occured in the training corpus
To do this we call the program "param_cnt" which takes the following
arguments:
<table border="1">
<tr><td> FLAG </td><td> DESCRIPTION </td></tr>
<tr><td> -moddeffn </td><td> model definition file with all possible triphones(alltriphones_mdef)</td></tr>
<tr><td> -ts2cbfn </td><td> takes the value ".cont." if you are building continuous models</td></tr>
<tr><td> -ctlfn </td><td> control file corresponding to your training
transcripts</td></tr>
<tr><td> -lsnfn </td><td> transcript file for training </td></tr>
<tr><td> -dictfn </td><td> training dictionary </td></tr>
<tr><td> -fdictfn </td><td> filler dictionary </td></tr>
<tr><td> -paramtype </td><td> write "phone" here, without the double
quotes</td></tr>
<tr><td> -segdir </td><td> /dev/null </td></tr>
</table>
<p>
param_cnt writes out the counts for each of the triphones onto stdout.
All other messages are sent to stderr. The stdout therefore has to
be directed into a file. If you are using csh or tcsh it would be done
in the following manner:
<pre>
(param_cnt [arguments] > triphone_count_file) >&! LOG
</pre>
Here's an example of the output of this program
<pre>
+GARBAGE+ - - - 98
+LAUGH+ - - - 29
SIL - - - 31694
AA - - - 0
AE - - - 0
...
AA AA AA s 1
AA AA AE s 0
AA AA AO s 4
</pre>
The final number in each row shows the number of times that particular
triphone (or filler phone) has occured in the training corpus. Not that
if all possible triphones of a CI phone are listed in the all_triphones.mdef
the CI phone itself will have 0 counts since all instances of it would have
been mapped to a triphone.
<p>
This list of counted triphones is used to shortlist the triphones that
have occured a minimum number (threshold) of times. The shortlisted
triphones appear in the same format as the file from which they have been
selected.
The shortlisted triphone list has the same format as the triphone list used
to generate the all_triphones.mdef. The formatted list of CI phones has to
be included in this as before. So, in the earlier example, if a threshold
of 4 were used, we would obtain the shortlisted triphone list as
<pre>
AA - - -
AE - - -
AO - - -
AW - - -
..
..
AA AA AO s
..
</pre>
The threshold is adjusted such that the total number of triphones
above the threshold is less that the maximum number of triphones that the
system can train (or that you wish to train). It is good to train as many
triphones as possible. The maximum number of triphones may however be
dependent on the memory available on your machine. The logistics related
to this are described in the beginning of this manual.
<p>
Note that thresholding is usually done so to reduce the number of
triphones, in order that the resulting models will be small enough to fit
in the computer's memory. If this is not a problem, then the threshold can
be set to a smaller number. If the triphone occurs too few times, however,
(ie, if the threshold is too small), there will not be enough data to train
the HMM state distributions properly. This would lead to poorly
estimated CD untied models, which in turn may affect the decision trees
which are to be built using these models in the next major step of the
training.
<p>
A model definition file is now created to include only these shortlisted
triphones. This is the final model definition file to be used for the
CD untied training. The reduced triphone list is then to the model
definition file using mk_model_def with the following arguments:
<table border="1">
<tr><td> FLAG </td><td> DESCRIPTION </td></tr>
<tr><td> -moddeffn </td><td> model definition file for CD untied
training</td></tr>
<tr><td> -phonelstfn </td><td> list of shortlisted triphones </td></tr>
<tr><td> -n_state_pm </td></tr> number of states per HMM </td></tr>
</table>
</ul>
<p>
Finally, therefore, a model definition file which
lists all CI phones and seen triphones is constructed. This file, like
the CI model-definition file, assigns unique id's to each HMM state
and serves as a reference file for handling and identifying the CD-untied
model parameters. Here is an example of the CD-untied model-definition file:
If you have listed five phones in your phones.list file,
<p>
SIL
B
AE
T
<p>
and specify that you want to build three state HMMs for each
of these phones, and if you have one utterance listed in your
transcript file:
<p> <s> BAT A TAB </s>
for which your dictionary and fillerdict entries are:
<pre>
Fillerdict:
<s> SIL
</s> SIL
</pre>
<pre>
Dictionary:
A AX
BAT B AE T
TAB T AE B
</pre>
<p>
then your CD-untied model-definition file will look like this:
<pre>
# Generated by <path_to_binary>/mk_model_def on Thu Aug 10 14:57:15 2000
0.3
5 n_base
7 n_tri
48 n_state_map
36 n_tied_state
15 n_tied_ci_state
5 n_tied_tmat
#
# Columns definitions
#base lft rt p attrib tmat ...state id's ...
SIL - - - filler 0 0 1 2 N
AE - - - n/a 1 3 4 5 N
AX - - - n/a 2 6 7 8 N
B - - - n/a 3 9 10 11 N
T - - - n/a 4 12 13 14 N
AE B T i n/a 1 15 16 17 N
AE T B i n/a 1 18 19 20 N
AX T T s n/a 2 21 22 23 N
B SIL AE b n/a 3 24 25 26 N
B AE SIL e n/a 3 27 28 29 N
T AE AX e n/a 4 30 31 32 N
T AX AE b n/a 4 33 34 35 N
The # lines are simply comments. The rest of the variables mean the following:
n_base : no. of CI phones (also called "base" phones), 5 here
n_tri : no. of triphones , 7 in this case
n_state_map : Total no. of HMM states (emitting and non-emitting)
The Sphinx appends an extra terminal non-emitting state
to every HMM, hence for 5+7 phones, each specified by
the user to be modeled by a 3-state HMM, this number
will be 12phones*4states = 48
n_tied_state: no. of states of all phones after state-sharing is done.
We do not share states at this stage. Hence this number is the
same as the total number of emitting states, 12*3=36
n_tied_ci_state:no. of states for your CI phones after state-sharing
is done. The CI states are not shared, now or later.
This number is thus again the total number of emitting CI
states, 5*3=15
n_tied_tmat : The total number of transition matrices is always the same
as the total number of CI phones being modeled. All triphones
for a given phone share the same transition matrix. This
number is thus 5.
Columns definitions: The following columns are defined:
base : name of each phone
lft : left-context of the phone (- if none)
rt : right-context of the phone (- if none)
p : position of a triphone. Four position markers are supported:
b = word begining triphone
e = word ending triphone
i = word internal triphone
s = single word triphone
attrib: attribute of phone. In the phone list, if the phone is "SIL",
or if the phone is enclosed by "+", as in "+BANG+", these
phones are interpreted as non-speech events. These are
also called "filler" phones, and the attribute "filler" is
assigned to each such phone. The base phones and the
triphones have no special attributes, and hence are
labelled as "n/a", standing for "no attribute"
tmat : the id of the transition matrix associated with the phone
state id's : the ids of the HMM states associated with any phone. This list
is terminated by an "N" which stands for a non-emitting
state. No id is assigned to it. However, it exists, and is
listed.
</pre>
<hr>
<p>
<a name="25"></a>
<!------------------------------------------------------------------------->
<center><h4><font color="red">TRAINING CONTINUOUS MODELS</font></h4></center>
<TABLE width="100%" bgcolor="#ffffff">
<td>FLAT INITIALIZATION OF CD UNTIED MODEL PARAMETERS</td>
</table>
<!------------------------------------------------------------------------->
<p>
In the next step in CD untied training, after the CD untied model
definition file has been constructed, the model parameters are first
intialized. During this process, the model parameter files corresponding
to the CD untied model-definition file are generated. Four files are
generated: means, variances, transition matrices and mixture weights. In
each of these files, the values are first copied from the corresponding
CI model parameter file. Each state of a particular CI phone contributes
to the same state of the same CI phone in the Cd -untied model parameter file,
and also to the same state of the *all* the triphones of the same CI
phone in the CD-untied model parameter file. The CD-untied model definition
file is of course used as a reference for this mapping. This process, as
usual, is called "initialization".
<p>
Initialization for the CD-untied training is done by the executable called
"init_mixw". It need the following arguments:
<table border="1">
<tr><td valign="top"> -src_moddeffn </td><td> source (CI) model definition file </td></tr>
<tr><td valign="top"> -src_ts2cbfn </td><td> .cont. </td></tr>
<tr><td valign="top"> -src_mixwfn </td><td> source (CI) mixture-weight file </td></tr>
<tr><td valign="top"> -src_meanfn </td><td> source (CI) means file </td></tr>
<tr><td valign="top"> -src_varfn </td><td> source (CI) variances file </td></tr>
<tr><td valign="top"> -src_tmatfn </td><td> source (CI) transition-matrices file </td></tr>
<tr><td valign="top"> -dest_moddeffn </td><td> destination (CD untied) model definition file </td></tr>
<tr><td valign="top"> -dest_ts2cbfn </td><td> .cont. </td></tr>
<tr><td valign="top"> -dest_mixwfn </td><td> destination (CD untied) mixtrue weights file </td></tr>
<tr><td valign="top"> -dest_meanfn </td><td> destination (Cd untied) means file </td></tr>
<tr><td valign="top"> -dest_varfn </td><td> destination (CD untied) variances file </td></tr>
<tr><td valign="top"> -dest_tmatfn </td><td> destination (Cd untied) transition matrices file </td></tr>
<tr><td valign="top"> -feat </td><td> feature configuration </td></tr>
<tr><td valign="top"> -ceplen </td><td> dimensionality of base feature vector </td></tr>
</table>
<hr>
<p>
<a name="26"></a>
<!------------------------------------------------------------------------->
<center><h4><font color="red">TRAINING CONTINUOUS MODELS</font></h4></center>
<TABLE width="100%" bgcolor="#ffffff">
<td>TRAINING CD UNTIED MODELS</td>
</table>
<!------------------------------------------------------------------------->
<p>
Once the initialization for CD-untied training is done, the next step is to actually train the CD untied models. To do this, as in the CI training,
the Baum-Welch forward-backward algorithm is iteratively used. Each iteration
consists of generating bw buffers by running the bw executable on the
training corpus (this can be divided into many parts as explained in the
CI training), follwed by running the norm executable to compute the
final parameters at the end of the iteration. The arguments required by the
bw executable at this stage are as follows.
<p>
<table border="1">
<tr><td align="center"> FLAG </td> <td align="center"> DESCRIPTION </td> </tr>
<tr><td valign="top"> -moddeffn</td> <td> CD-untied model definition file </td></tr>
<tr><td valign="top"> -ts2cbfn </td> <td> this flag should be set to ".cont." if
you are training continuous models, and to
".semi." if you are training semi-continuous
models, without the double quotes </td></tr>
<tr><td valign="top"> -mixwfn </td> <td> name of the file in which the
mixture-weights from the previous iteration are stored. Full path must be
provided</td></tr>
<tr><td valign="top"> -mwfloor </td> <td> Floor value for the mixture weights. Any number
below the floor value is set to the floor
value.</td></tr>
<tr><td valign="top"> -tmatfn </td> <td> name of the file in which
the transition matrices from the previous iteration are stored.
Full path must be provided</td></tr>
<tr><td valign="top"> -meanfn </td> <td>name of the file in which
the means from the previous iteration are stored. Full path must be
provided</td></tr>
<tr><td valign="top"> -varfn </td> <td>name of the file in which
the variances fromt he previous iteration are stored.
Full path must be provided</td></tr>
<tr><td valign="top"> -dictfn </td> <td> Dictionary </td></tr>
<tr><td valign="top"> -fdictfn </td> <td> Filler dictionary</td></tr>
<tr><td valign="top"> -ctlfn </td> <td> control file </td></tr>
<tr><td valign="top"> -part </td> <td> You can split the training into N equal parts by
setting a flag. If there are M utterances in your control file, then this will
enable you to run the training separately on each (M/N)<sup>th</sup> part. This
flag may be set to specify which of these parts you want to currently train
on. As an example, if your total number of parts is 3, this flag can take
one of the values 1,2 or 3</td></tr>
<tr><td valign="top"> -npart </td><td> number of parts in which you have split
the training </td></tr>
<tr><td valign="top"> -cepdir </td><td> directory where your feature files are
stored</td></tr>
<tr><td valign="top"> -cepext </td><td> the extension that comes after the name listed
in the control file. For example, you may have a file called a/b/c.d and
may have listed a/b/c in your control file. Then this flag must be given the
argument "d", without the double quotes or the dot before it </td></tr>
<tr><td valign="top"> -lsnfn </td><td> name of the transcript file </td></tr>
<tr><td valign="top"> -accumdir </td><td> Intermediate results from each part of your training will be written in this directory. If you have T means to estimate, then
the size of the mean buffer from the current part of your training will
be T*4 bytes (say). There will likewise be a variance buffer, a buffer for
mixture weights, and a buffer for transition matrices</td></tr>
<tr><td valign="top"> -varfloor </td><td> minimum variance value allowed </td></tr>
<tr><td valign="top"> -topn </td><td> no. of gaussians to consider for likelihood computation</td></tr>
<tr><td valign="top"> -abeam </td><td> forward beamwidth</td></tr>
<tr><td valign="top"> -bbeam </td><td> backward beamwidth</td></tr>
<tr><td valign="top"> -agc </td><td> automatic gain control</td></tr>
<tr><td valign="top"> -cmn </td><td> cepstral mean normalization</td></tr>
<tr><td valign="top"> -varnorm </td><td> variance normalization</td></tr>
<tr><td valign="top"> -meanreest </td><td> mean re-estimation</td></tr>
<tr><td valign="top"> -varreest </td><td> variance re-estimation</td></tr>
<tr><td valign="top"> -2passvar </td><td> Setting this flag to "yes" lets bw
use the previous means in the estimation of the variance. The current variance
is then estimated as E[(x - prev_mean)<sup>2</sup>]. If this flag is set to
"no" the current estimate of the means are used to estimate variances. This
requires the estimation of variance as E[x<sup>2</sup>] - (E[x])<sup>2</sup>,
an unstable estimator that sometimes results in negative estimates of the
variance due to arithmetic imprecision</td></tr>
<tr><td valign="top"> -tmatreest </td><td> re-estimate transition matrices or not</td></tr>
<tr><td valign="top"> -feat </td><td> feature configuration</td></tr>
<tr><td valign="top"> -ceplen </td><td> length of basic feature vector</td></tr>
</table>
<p>
<li> The Baum-Welch step should be followed by the nomalization step. The
executable "norm" must be used for this. The arguments required by the norm
executable are the same as that for CI training, and are listed below:
<p>
<table border="1">
<tr><td align="center"> FLAG </td> <td align="center"> DESCRIPTION </td></tr>
<tr><td valign="top"> -accumdir </td><td> Intermediate buffer directory</td></tr>
<tr><td valign="top"> -feat </td><td> feature configuration</td></tr>
<tr><td valign="top"> -mixwfn </td> <td> name of the file in which you want
to write the mixture weights.
Full path must be provided</td></tr>
<tr><td valign="top"> -tmatfn </td> <td> name of the file in which you want to write
the transition matrices. Full path must be
provided</td></tr>
<tr><td valign="top"> -meanfn </td> <td>name of the file in which you want to write
the means. Full path must be
provided</td></tr>
<tr><td valign="top"> -varfn </td> <td>name of the file in which you want to write
the variances. Full path must be
provided</td></tr>
<tr><td valign="top"> -ceplen </td><td> length of basic feature vector</td></tr>
</table>
<p>
The iterations of Baum-Welch and norm must be run till the likelihoods
converge (ie, the convergence ratio reaches a small threshold value). Typically
this happens in 6-10 iterations.
<hr>
<p>
<a name="27"></a>
<!------------------------------------------------------------------------->
<center><h4><font color="red">TRAINING CONTINUOUS MODELS</font></h4></center>
<TABLE width="100%" bgcolor="#ffffff">
<td>BUILDING DECISION TREES FOR PARAMETER SHARING</td>
</table>
<!------------------------------------------------------------------------->
<p>
Once the CD-untied models are computed, the next major step in training
continuous models is decision tree building. Decision trees are used to
decide which of the HMM states of all the triphones (seen and unseen) are
similar to each other, so that data from all these states are collected
together and used to train one global state, which is called a "senone".
Many groups of similar states are formed, and the number of "senones"
that are finally to be trained can be user defined. A senone is also
called a tied-state and is obviously shared across the triphones which
contributed to it. The technical details of decision tree building and
state tying are explained in the technical section of this manual. It is
sufficient to understand here that for state tying, we require to build
decision trees.
<p>
<a name="28"></a>
<u><b>Generating the linguistic questions</b></u>
<p>
The decision trees require the CD-untied models and a set of predefined
phonetic classes (or classes of the acoustic units you are modeling) which
share some common property. These classes or questions are used to
partition the data at any given node of a tree. Each question results in
one partion, and the question that results in the "best" partition (maximum
increase in likelihood due to the partition) is chosen to partition the
data at that node. All linguistic questions are written in a single file
called the "linguistic questions" file. One decision tree is built for each
state of each phone.
<p>
For example, if you want to build a decision tree for the contexts (D B P
AE M IY AX OY) for any phone, then you could ask the question: does the
context belong to the class vowels? If you have defined the class vowels to
have the phones AE AX IY OY EY AA EH (in other words, if one of your
linguistic questions has the name "VOWELS" and has the elements AE AX IY OY
EY AA EH corresponding to that name), then the decision tree would branch
as follows:
<pre>
D B P AE M IY AX OY
|
question: does this context belong to the class VOWELS ?
/\
/ \
/ \
yes no
/ \
/ \
AE IY AX OY D B P M
| |
question question
/\ /\
/ \ / \
</pre>
Here is an example of a "linguistic-questions" file:
<pre>
ASPSEG HH
SIL SIL
VOWELS AE AX IY OY EY AA EH
ALVSTP D T N
DENTAL DH TH
LABSTP B P M
LIQUID L R
</pre>
The column on the left specifies the name gives to the class. This name
is user defined. The classes consist of a single phone or a cluster of phones
whaich share some common acoustic property. If your acoustic units are not
completely phonetically motivated, or if you are training models for
a language whose phonetic structure you are not completely sure about,
then the executable classed "make_quests" provided with the SPHINX-III
package can be used to generate the linguistic questions. It uses the
CI models to make the questions, and needs the following arguments:
<p>
<table border="1">
<tr><td valign="top"> FLAG </td><td> DESCRIPTION </td></tr>
<tr><td valign="top"> -moddeffn </td><td> CI model definition file </td></tr>
<tr><td valign="top"> -meanfn </td><td> CI means file </td></tr>
<tr><td valign="top"> -varfn </td><td> CI variances file </td></tr>
<tr><td valign="top"> -mixwfn </td><td> CI mixture weights file </td></tr>
<tr><td valign="top"> -npermute </td><td> A bottom-up top-down clustering
algorithm is used to group the phones into classes. Phones are clustered
using bottom-up clustering until npermute classes are obtained. The npermute
classes are exhaustively partitioned into two classes and evaluated to identify
the optimal partitioning of the entire phone set into two groups.
An identical procedure is performed recursively on each of these groups
to generate an entire tree.
npermute is typically between 8 and 12. Smaller values of npermute result
in suboptimal clustering. Larger values become computationally prohibitive.
</td></tr>
<tr><td valign="top"> -niter </td><td> The bottom-up top-down clustering
can be iterated to give more optimal clusters. niter sets the number of
iterations to run. niter is typiclly set to 1 or 2. The clustering
saturates after 2 iterations.
</td></tr>
<tr><td valign="top"> -qstperstt </td><td> The algoritm clusters state
distributions belonging to each state of the CI phone HMMs to generate
questions. Thus all 1st states are clustered to generate one subset of
questions, all 2nd states are clustered for the second subset, and so on.
qstperstt determines how many questions are to be generated by clustering
any state. Typically this is set to a number between 20 and 25.
</td></tr>
<tr><td valign="top"> -tempfn </td><td> </td></tr>
<tr><td valign="top"> -questfn </td><td> output lingustic questions file </td></tr>
</table>
<p>
Once the linguistic questions have been generated, decision trees must be
built for each state of each CI phone present in your phonelist. Decision
trees are however not built for filler phones written as +()+ in your
phonelist. They are also not built for the SIL phone. In order to build
decision trees, the executable "bldtree" must be used. It takes the following
arguments:
<table border="1">
<tr><td valign="top"> FLAG </td><td> DESCRIPTION </td></tr>
<tr><td valign="top"> -treefn </td><td> full path to the directory in which
you want the decision trees to be written </td></tr>
<tr><td valign="top"> -moddeffn </td><td> CD-untied model definition file
</td></tr>
<tr><td valign="top"> -mixwfn </td><td> Cd-untied mixture weights file
</td></tr>
<tr><td valign="top"> -ts2cbfn </td><td> .cont. </td></tr>
<tr><td valign="top"> -meanfn </td><td> CD-untied means file </td></tr>
<tr><td valign="top"> -varfn </td><td> CD-untied variances file </td></tr>
<tr><td valign="top"> -mwfloor </td><td> Floor value of the mixture
weights. Values below this are reset to this value. A typical value is
1e-8 </td></tr>
<tr><td valign="top"> -psetfn </td><td> linguistic questions file
</td></tr>
<tr><td valign="top"> -phone </td><td> CI phone for which you want to build
the decision tree </td></tr>
<tr><td valign="top"> -state </td><td> The HMM state for which you want to
build the decision tree. For a three state HMM, this value can be 0,1 or
2. For a 5 state HMM, this value can be 0,1,2,3 or 4, and so on </td></tr>
<tr><td valign="top"> -stwt </td><td> This flag needs a string of numbers
equal to the number of HMM-states, for example, if you were using 5-state
HMMs, then the flag could be given as "-stwt 1.0 0.3 0.1 0.01 0.001". Each
of these numbers specify the weights to be given to state distributions
during tree building, beginning with the *current* state. The second number
specifies the weight to be given to the states *immediately adjacent* to
the current state (if there are any), the third number specifies the weight
to be given to adjacent states *one removed* from the immediately adjacent
one (if there are any), and so on. A typical set of values for 5 state HMMs
is "1.0 0.3 0.1 0.01 0.001" </td></tr>
<tr><td valign="top"> -ssplitmin </td><td> Complex questions are built for
the decision tree by first building "pre-trees" using the linguistic
questions in the question file. The minimum number of bifurcations in this
tree is given by ssplitmin. This should not be lesser than 1. This value is
typically set to 1. </td></tr>
<tr><td valign="top"> -ssplitmax </td><td> The maximum number of
bifurcations in the simple tree before it is used to build complex
questions. This number is typically set to 7. Larger values would be more
computationally intensive. This number should not be smaller than the value
given for ssplitmin</td></tr> <tr><td valign="top"> -ssplitthr </td><td>
Minimum increase in likelihood to be considered for a bifurcation in the
simple tree. Typically set to a very small number greater than or equal to
0 </td></tr>
<tr><td valign="top"> -csplitmin </td><td> The minimum number of
bifurcations in the decision tree. This should not be less than 1
</td></tr>
<tr><td valign="top"> -csplitmax </td><td> The maximum number of
bifurcations in the decision tree. This should be as large as
computationlly feasible. This is typically set to 2000 </td></tr>
<tr><td valign="top"> -csplitthr </td><td> Minimum increase in likelihood
to be considered for a bifurcation in the decision tree. Typically set to a
very small number greater than or equal to 0. </td></tr>
<tr><td valign="top"> -cntthresh </td><td> Minimum number of observations
in a state for it to be considered in the decision tree building process.
</td></tr>
</table>
If, for example, you have a phonelist which contains the following phones
<pre>
+NOISE+
SIL
AA
AX
B
</pre>
and you are training 3 state HMMs, then you must build 9 decision trees,
one each for each state of the phones AA, AX and B.
<hr>
<p>
<a name="29"></a>
<!------------------------------------------------------------------------->
<center><h4><font color="red">TRAINING CONTINUOUS MODELS</font></h4></center>
<TABLE width="100%" bgcolor="#ffffff">
<td>PRUNING THE DECISION TREES</td>
</table>
<!------------------------------------------------------------------------->
<p>
Once the decision trees are built, they must be pruned to have as many
leaves as the number of tied states (senones) that you want to
train. Remember that the number of tied states does not include the CI
states, which are never tied. In the pruning process, the bifurcations in
the decision trees which resulted in the minimum increase in
likelihood are progressively removed and replaced by the parent node. The
selection of the branches to be pruned out is done across the entire
collection of decision trees globally. The executable to be used for
decision tree pruning is called "prunetree" and requires the following
arguments:
<table border="1">
<tr><td vlign="top"> FLAG </td><td> DESCRIPTION </td></tr>
<tr><td vlign="top"> -itreedir </td><td> directory in which the full
decision trees are stored </td></tr>
<tr><td vlign="top"> -nseno </td><td> number of senones that you want to
train </td></tr>
<tr><td vlign="top"> -otreedir </td><td> directory to store the pruned
decision trees </td></tr>
<tr><td vlign="top"> -moddeffn </td><td> CD-untied model definition file
</td></tr>
<tr><td vlign="top"> -psetfn </td><td> lingistic questions file </td></tr>
<tr><td vlign="top"> -minocc </td><td> minimum number of observations in the
given tied state. If there are fewer observations, the branches corresponding
to the tied state get pruned out by default. This value should never be 0,
otherwise you will end up having senones with no data to train
(which are seen 0 times in the training set) </td></tr>
</table>
<hr>
<p>
<a name="30"></a>
<!------------------------------------------------------------------------->
<center><h4><font color="red">TRAINING CONTINUOUS MODELS</font></h4></center>
<TABLE width="100%" bgcolor="#ffffff">
<td>CREATING THE CD TIED MODEL DEFINITION FILE</td>
</table>
<!------------------------------------------------------------------------->
<p>
Once the trees are pruned, a new model definition file must be created
which
<ul start="a">
<li>contains all the triphones which are seen during training
<li>has the states corresponding to these triphones identified with
senones from the pruned trees
</ul>
In order to do this, the model definition file which contains all possible
triphones from the current training dictionary can be used [alltriphones model
definition file]. This was
built during the process of building the CD-untied model definition file.
Remember that the CD-untied model definition file contained only
a selected number of triphones, with various thresholds used for selection.
That file, therefore, cannot be used to build the CD-tied model definition
file, except in the exceptional case where you are sure that the
CD-untied model definition file includes *all* triphones seen during training.
The executable that must be used to tie states is called "tiesate" and
needs the following arguments:
<p>
<table border="1">
<tr><td valign="top"> FLAG </td><td> DESCRIPTION </td></tr>
<tr><td valign="top"> -imoddeffn </td><td> alltriphones model definition
file </td></tr>
<tr><td valign="top"> -omoddeffn </td><td> CD-tied model definition file
</td></tr>
<tr><td valign="top"> -treedir </td><td> pruned tree directory </td></tr>
<tr><td valign="top"> -psetfn </td><td> linguistic questions file </td></tr>
</table>
<p>
Here is an example of a CD-tied model definition file, based on the earlier
example given for the CD-untied model definition file. The alltriphones model
definition file:
<pre>
# Generated by [path]/mk_model_def on Sun Nov 26 12:42:05 2000
# triphone: (null)
# seno map: (null)
#
0.3
5 n_base
34 n_tri
156 n_state_map
117 n_tied_state
15 n_tied_ci_state
5 n_tied_tmat
#
# Columns definitions
#base lft rt p attrib tmat ... state id's ...
SIL - - - filler 0 0 1 2 N
AE - - - n/a 1 3 4 5 N
AX - - - n/a 2 6 7 8 N
B - - - n/a 3 9 10 11 N
T - - - n/a 4 12 13 14 N
AE B T i n/a 1 15 16 17 N
AE T B i n/a 1 18 19 20 N
AX AX AX s n/a 2 21 22 23 N
AX AX B s n/a 2 24 25 26 N
AX AX SIL s n/a 2 27 28 29 N
AX AX T s n/a 2 30 31 32 N
AX B AX s n/a 2 33 34 35 N
AX B B s n/a 2 36 37 38 N
AX B SIL s n/a 2 39 40 41 N
AX B T s n/a 2 42 43 44 N
AX SIL AX s n/a 2 45 46 47 N
AX SIL B s n/a 2 48 49 50 N
AX SIL SIL s n/a 2 51 52 53 N
AX SIL T s n/a 2 54 55 56 N
AX T AX s n/a 2 57 58 59 N
AX T B s n/a 2 60 61 62 N
AX T SIL s n/a 2 63 64 65 N
AX T T s n/a 2 66 67 68 N
B AE AX e n/a 3 69 70 71 N
B AE B e n/a 3 72 73 74 N
B AE SIL e n/a 3 75 76 77 N
B AE T e n/a 3 78 79 80 N
B AX AE b n/a 3 81 82 83 N
B B AE b n/a 3 84 85 86 N
B SIL AE b n/a 3 87 88 89 N
B T AE b n/a 3 90 91 92 N
T AE AX e n/a 4 93 94 95 N
T AE B e n/a 4 96 97 98 N
T AE SIL e n/a 4 99 100 101 N
T AE T e n/a 4 102 103 104 N
T AX AE b n/a 4 105 106 107 N
T B AE b n/a 4 108 109 110 N
T SIL AE b n/a 4 111 112 113 N
T T AE b n/a 4 114 115 116 N
</pre>
is used as the base to give the following CD-tied model definition file
with 39 tied states (senones):
<pre>
# Generated by [path]/mk_model_def on Sun Nov 26 12:42:05 2000
# triphone: (null)
# seno map: (null)
#
0.3
5 n_base
34 n_tri
156 n_state_map
54 n_tied_state
15 n_tied_ci_state
5 n_tied_tmat
#
# Columns definitions
#base lft rt p attrib tmat ... state id's ...
SIL - - - filler 0 0 1 2 N
AE - - - n/a 1 3 4 5 N
AX - - - n/a 2 6 7 8 N
B - - - n/a 3 9 10 11 N
T - - - n/a 4 12 13 14 N
AE B T i n/a 1 15 16 17 N
AE T B i n/a 1 18 16 19 N
AX AX AX s n/a 2 20 21 22 N
AX AX B s n/a 2 23 21 22 N
AX AX SIL s n/a 2 24 21 22 N
AX AX T s n/a 2 25 21 22 N
AX B AX s n/a 2 26 21 27 N
AX B B s n/a 2 23 21 27 N
AX B SIL s n/a 2 24 21 27 N
AX B T s n/a 2 25 21 27 N
AX SIL AX s n/a 2 26 21 28 N
AX SIL B s n/a 2 23 21 28 N
AX SIL SIL s n/a 2 24 21 28 N
AX SIL T s n/a 2 25 21 28 N
AX T AX s n/a 2 26 21 29 N
AX T B s n/a 2 23 21 29 N
AX T SIL s n/a 2 24 21 29 N
AX T T s n/a 2 25 21 29 N
B AE AX e n/a 3 30 31 32 N
B AE B e n/a 3 33 31 32 N
B AE SIL e n/a 3 34 31 32 N
B AE T e n/a 3 35 31 32 N
B AX AE b n/a 3 36 37 38 N
B B AE b n/a 3 36 37 39 N
B SIL AE b n/a 3 36 37 40 N
B T AE b n/a 3 36 37 41 N
T AE AX e n/a 4 42 43 44 N
T AE B e n/a 4 45 43 44 N
T AE SIL e n/a 4 46 43 44 N
T AE T e n/a 4 47 43 44 N
T AX AE b n/a 4 48 49 50 N
T B AE b n/a 4 48 49 51 N
T SIL AE b n/a 4 48 49 52 N
T T AE b n/a 4 48 49 53 N
</pre>
<hr>
<p>
<a name="31"></a>
<!------------------------------------------------------------------------->
<center><h4><font color="red">TRAINING CONTINUOUS MODELS</font></h4></center>
<TABLE width="100%" bgcolor="#ffffff">
<td>INITIALIZING AND TRAINING CD TIED GAUSSIAN MIXTURE MODELS </td>
</table>
<!------------------------------------------------------------------------->
<p>
The next step is to train the CD-tied models. In the case of
continuous models, the HMM states can be modeled by either a single
Gaussian distribution, or a mixture of Gaussian distributions. The number
of Gaussians in a mixture-distribution must preferably be even, and a power
of two (for example, 2,4,8,16, 32,..). To model the HMM states by a mixture
of 8 Gaussians (say), we first have to train 1 Gaussian per state
models. Each Gaussian distribution is then split into two by perturbing its
mean slightly, and the resulting two distributions are used to intialize
the training for 2 Gaussian per state models. These are further perturbed
to initialize for 4 Gaussains per state models and a further split is done
to initalize for the 8 Gaussians per state models. So the CD-tied training
for models with 2<sup>N</sup> Gaussians per state is done in N+1
steps. Each of these N+1 steps consists of
<p>
<ol>
<li>initialization
<li>iterations of Baum-Welch followed by norm
<li>Gaussian splitting (not done in the N+1<sup>th</sup> stage of CD-tied training)
</ol>
<p>
The training begins with the initialization of the 1 Gaussian per state models.
During initialization, the model parameters from
the CI model parameter files are copied into appropriate positions
in the CD tied model parameter files. Four model parameter files
are created, one each for the means, variances, transition matrices
and mixture weights. During initialization, each state of a particular CI phone
contributes
to the same state of the same CI phone in the CD-tied model parameter file,
and also to the same state of the *all* the triphones of the same CI
phone in the CD-tied model parameter file. The CD-tied model definition
file is used as a reference for this mapping.
<p>
Initialization for the 1 gaussian per state models is done by the executable
called <b><font color="green">init_mixw</font></b>. It requires the following
arguments:
<p>
<table border="1">
<tr><td valign="top"> -src_moddeffn </td><td> source (CI) model definition file
</td></tr>
<tr><td valign="top"> -src_ts2cbfn </td><td> .cont. </td></tr>
<tr><td valign="top"> -src_mixwfn </td><td> source (CI) mixture-weight file </td
></tr>
<tr><td valign="top"> -src_meanfn </td><td> source (CI) means file </td></tr>
<tr><td valign="top"> -src_varfn </td><td> source (CI) variances file </td></tr>
<tr><td valign="top"> -src_tmatfn </td><td> source (CI) transition-matrices file
</td></tr>
<tr><td valign="top"> -dest_moddeffn </td><td> destination (CD tied) model def
inition file </td></tr>
<tr><td valign="top"> -dest_ts2cbfn </td><td> .cont. </td></tr>
<tr><td valign="top"> -dest_mixwfn </td><td> destination
(CD tied 1 Gau/state) mixture weights file </td></tr>
<tr><td valign="top"> -dest_meanfn </td><td> destination (CD tied 1 Gau/state) means file
</td></tr>
<tr><td valign="top"> -dest_varfn </td><td> destination (CD tied 1 Gau/state) variances file </td></tr>
<tr><td valign="top"> -dest_tmatfn </td><td> destination (CD tied 1 Gau/state) transition
matrices file </td></tr>
<tr><td valign="top"> -feat </td><td> feature configuration </td></tr>
<tr><td valign="top"> -ceplen </td><td> dimensionality of base feature vector
<td></tr>
</table>
<p>
<p>
The executables used for baum-welch, norm and Gaussaian splitting are
<b><font color="green">bw</font></b>,
<b><font color="green">norm</font></b> and
<b><font color="green">inc_comp</font></b>
<p>
The arguments needed by <b><font color="green">bw</font></b> are
<p>
<table border="1", noshade>
<tr><td align="center"> FLAG </td> <td align="center"> DESCRIPTION </td> </tr>
<tr><td valign="top"> -moddeffn</td> <td> CD tied model definition file</td></tr>
<tr><td valign="top"> -ts2cbfn </td> <td> this flag should be set to ".cont." if
you are training continuous models, and to
".semi." if you are training semi-continuous
models, without the double quotes </td></tr>
<tr><td valign="top"> -mixwfn </td> <td> name of the file in which the
mixture-weights from the previous iteration are stored. Full path must be
provided</td></tr>
<tr><td valign="top"> -mwfloor </td> <td> Floor value for the mixture weights. Any number
below the floor value is set to the floor
value.</td></tr>
<tr><td valign="top"> -tmatfn </td> <td> name of the file in which
the transition matrices from the previous iteration are stored.
Full path must be provided</td></tr>
<tr><td valign="top"> -tpfloor </td> <td> Floor value for the transition probabilities. Any number
below the floor value is set to the floor
value.</td></tr>
<tr><td valign="top"> -meanfn </td> <td>name of the file in which
the means from the previous iteration are stored. Full path must be
provided</td></tr>
<tr><td valign="top"> -varfn </td> <td>name of the file in which
the variances fromt he previous iteration are stored.
Full path must be provided</td></tr>
<tr><td valign="top"> -dictfn </td> <td> Dictionary </td></tr>
<tr><td valign="top"> -fdictfn </td> <td> Filler dictionary</td></tr>
<tr><td valign="top"> -ctlfn </td> <td> control file </td></tr>
<tr><td valign="top"> -part </td> <td> You can split the training into N equal parts by
setting a flag. If there are M utterances in your control file, then this will
enable you to run the training separately on each (M/N)<sup>th</sup> part. This
flag may be set to specify which of these parts you want to currently train
on. As an example, if your total number of parts is 3, this flag can take
one of the values 1,2 or 3</td></tr>
<tr><td valign="top"> -npart </td><td> number of parts in which you have split
the training </td></tr>
<tr><td valign="top"> -cepdir </td><td> directory where your feature files are
stored</td></tr>
<tr><td valign="top"> -cepext </td><td> the extension that comes after the name listed
in the control file. For example, you may have a file called a/b/c.d and
may have listed a/b/c in your control file. Then this flag must be given the
argument "d", without the double quotes or the dot before it </td></tr>
<tr><td valign="top"> -lsnfn </td><td> name of the transcript file </td></tr>
<tr><td valign="top"> -accumdir </td><td> Intermediate results from each part of your training will be written in this directory. If you have T means to estimate, then
the size of the mean buffer from the current part of your training will
be T*4 bytes (say). There will likewise be a variance buffer, a buffer for
mixture weights, and a buffer for transition matrices</td></tr>
<tr><td valign="top"> -varfloor </td><td> minimum variance value allowed </td></tr>
<tr><td valign="top"> -topn </td><td> no. of gaussians to consider for likelihood computation</td></tr>
<tr><td valign="top"> -abeam </td><td> forward beamwidth</td></tr>
<tr><td valign="top"> -bbeam </td><td> backward beamwidth</td></tr>
<tr><td valign="top"> -agc </td><td> automatic gain control</td></tr>
<tr><td valign="top"> -cmn </td><td> cepstral mean normalization</td></tr>
<tr><td valign="top"> -varnorm </td><td> variance normalization</td></tr>
<tr><td valign="top"> -meanreest </td><td> mean re-estimation</td></tr>
<tr><td valign="top"> -varreest </td><td> variance re-estimation</td></tr>
<tr><td valign="top"> -2passvar </td><td> Setting this flag to "yes" lets bw
use the previous means in the estimation of the variance. The current variance
is then estimated as E[(x - prev_mean)<sup>2</sup>]. If this flag is set to
"no" the current estimate of the means are used to estimate variances. This
requires the estimation of variance as E[x<sup>2</sup>] - (E[x])<sup>2</sup>,
an unstable estimator that sometimes results in negative estimates of the
variance due to arithmetic imprecision</td></tr>
<tr><td valign="top"> -tmatreest </td><td> re-estimate transition matrices or not</td></tr>
<tr><td valign="top"> -feat </td><td> feature configuration</td></tr>
<tr><td valign="top"> -ceplen </td><td> length of basic feature vector</td></tr>
</table>
<p>
The arguments needed by <b><font color="green">norm</font></b> are:
<p>
<table border="1", noshade>
<tr><td align="center"> FLAG </td> <td align="center"> DESCRIPTION </td></tr>
<tr><td valign="top"> -accumdir </td><td> Intermediate buffer directory</td></tr>
<tr><td valign="top"> -feat </td><td> feature configuration</td></tr>
<tr><td valign="top"> -mixwfn </td> <td> name of the file in which you want
to write the mixture weights.
Full path must be provided</td></tr>
<tr><td valign="top"> -tmatfn </td> <td> name of the file in which you want to write
the transition matrices. Full path must be
provided</td></tr>
<tr><td valign="top"> -meanfn </td> <td>name of the file in which you want to write
the means. Full path must be
provided</td></tr>
<tr><td valign="top"> -varfn </td> <td>name of the file in which you want to write
the variances. Full path must be
provided</td></tr>
<tr><td valign="top"> -ceplen </td><td> length of basic feature vector</td></tr>
</table>
<p>
The arguments needed by <b><font color="green">inc_comp</font></b> are:
<p>
<table border="1">
<tr><td> FLAG </td><td> DESCRIPTION </td></tr>
<tr><td> -ninc </td><td> how many gaussians (per state) to split currently. You need not
always split to double the number of Gaussians. you can specify other numbers
here, so long as they are less than the number of Gaussians you currently have.This is a positive integer like "2", given without the double quotes </td></tr>
<tr><td> -ceplen </td><td> length of the base feature vector </td></tr>
<tr><td> -dcountfn </td><td> input mixture weights file </td></tr>
<tr><td> -inmixwfn </td><td> input mixture weights file </td></tr>
<tr><td> -outmixwfn </td><td>output mixture weights file </td></tr>
<tr><td> -inmeanfn </td><td> input means file </td></tr>
<tr><td> -outmeanfn </td><td>ouput means file </td></tr>
<tr><td> -invarfn </td><td> input variances file </td></tr>
<tr><td> -outvarfn </td><td> output variances file </td></tr>
<tr><td> -feat </td><td> type of feature </td></tr>
</table>
<p>
<a href="#top">Back to index</a>
<hr>
<a name="3"></a>
<a name="3b"></a>
<center><h4><font color="red">TRAINING SEMI-CONTINUOUS MODELS</font></h4></center>
<TABLE width="100%" bgcolor="#ffffff">
<td>VECTOR QUANTIZATION</td>
</table>
<p>
This is done in two steps. In the first step, the
feature vectors are accumulated for quantizing the vector space. Not
all feature vectors are used. Rather, a sampling of the vectors available
is done by the executable "agg_seg". This executable simply "aggregates"
the vectors into a buffer. The following flag settings
must be used with this executable:
<p>
<table border="1">
<tr><td> FLAG </td><td> DESCRIPTION </td></tr>
<tr><td> -segdmpdirs </td><td> directory in which you want to put the aggregate buffer </td></tr>
<tr><td> -segdmpfn </td><td> name of the buffer (file) </td></tr>
<tr><td> -segtype </td><td> all </td></tr>
<tr><td> -ctlfn </td><td> control file </td></tr>
<tr><td> -cepdir </td><td> path to feature files </td></tr>
<tr><td> -cepext </td><td> feature vector filename extension </td></tr>
<tr><td> -ceplen </td><td> dimensionality of the base feature vector </td></tr>
<tr><td> -agc </td><td> automatic gain control factor(max/none) </td></tr>
<tr><td> -cmn </td><td> cepstral mean normalization(yes/no) </td></tr>
<tr><td> -feat </td><td> type of feature. As mentioned earlier, the 4-stream feature vector is usually given as an option here. When you specify the
4-stream feature, this program will compute and aggregate vectors
corresponding to all streams separately.</td></tr>
<tr><td> -stride </td><td> how many samples to ignore during sampling of
vectors (pick every stride'th sample)</td></tr>
</table>
<p>
In the second step of vector quantization, an Expectation-Maximization (EM)
algorithm is applied to segregate each aggregated stream of vectors into a
codebook of N Gaussians. Usually N is some power of 2, the commonly used
number is N=256. The number 256 can in principle be varied, but this option
is not provided in the SPHINX-II decoder. So if you intend to use the
SPHINX-II decoder, but are training models with SPHINX-III trainer, you
must use N=256. It has been observed that the quality of the models built with
256 codeword codebooks is sufficient for good recognition. Increasing the
number of codewords may cause data-insufficiency problems. In many instances,
the choice to train semi-continuous models (rather than continuous ones) arises
from insufficiency of training data. When this is indeed the case,
increasing the number of codebooks might aggravate the estimation
problems that might arise due to data insufficiency. Consider this
fact seriously before you decide to increase N.
<p>
In SPHINX-III, the EM-step is done through a k-means algorithm carried
out by the executable <b><font color="green">kmeans_init</font></b>.
This executable is usually used with the following flag settings:
<pre>
-grandvar yes
-gthobj single
-stride 1
-ntrial 1
-minratio 0.001
-ndensity 256
-meanfn full_path_to_codebookmeans.file
-varfn full_path_to_codebookvariances.file
-reest no
-segdmpdirs directory_in_which_you_want_to_put_aggregate.file
-segdmpfn aggregate.file
-ceplen dimensionality_of_feature_vector
-feat type_of_feature
-agc automatic_gain_control_factor(max/none)
-cmn cepstral_mean_normalization(yes/no)
</pre>
Once the vector quantization is done, you have to flat-initialize your
acoustic models to prepare for the first real step in training. The following
steps explain the flat-initialization process:
<p>
<a name="3d"></a>
<center><h4><font color="red">TRAINING SEMI-CONTINUOUS MODELS</font></h4></center>
<TABLE width="100%" bgcolor="#ffffff">
<td>CREATING THE CI MODEL DEFINITION FILE</td>
</table>
<p>
<a href="#20">
This procedure is the same as described for continuous models</a>.
<a name="3e"></a>
<center><h4><font color="red">TRAINING SEMI-CONTINUOUS MODELS</font></h4></center>
<TABLE width="100%" bgcolor="#ffffff">
<td>CREATING THE HMM TOPOLOGY FILE</td>
</table>
<p>
<a href="#21">
This procedure is the same as described for continuous models</a>.
<a name="3c"></a>
<center><h4><font color="red">TRAINING SEMI-CONTINUOUS MODELS</font></h4></center>
<TABLE width="100%" bgcolor="#ffffff">
<td>FLAT INITIALIZATION OF CI MODEL PARAMETERS</td>
</table>
<p>
In flat-initialization, all mixture weights are set to be equal for
all states, and all state transition probabilities are set to be equal.
Unlike in continuous models, the means and variances of the codebook
Gaussians are not given
global values, since they are already estimated from the data in the
vector quantization step. To flat-initialize the mixture weights, each
component of each mixture-weight distribution of each feature stream is set
to be a number equal to 1/N, where N is the codebook size.
The mixture_weights and
transition_matrices are initialized using the executable <b><font
color="green">mk_flat</font></b>. It needs the following arguments:
<p>
<table border="1">
<tr><td> FLAG </td><td> DESCRIPTION </td></tr>
<tr><td> -moddeffn </td><td> CI model definition file </td></tr>
<tr><td> -topo </td><td> HMM topology file. </td></tr>
<tr><td> -mixwfn </td><td> file in which you want to write the
initialized mixture weights </td></tr>
<tr><td> -tmatfn </td><td> file in which you want to write the
initialized transition matrices </td></tr>
<tr><td> -nstream </td><td> number of independent feature streams, for
continuous models this number should be set to "1", without the double quotes
</td></tr>
<tr><td> -ndensity </td><td> codebook size. This number is usually set
to "256", without the double quotes</td></tr>
</table>
<p>
<a name="3f"></a>
<center><h4><font color="red">TRAINING SEMI-CONTINUOUS MODELS</font></h4></center>
<TABLE width="100%" bgcolor="#ffffff">
<td>TRAINING CI MODELS</td>
</table>
<p>
<a href="#23">
This procedure is the same as described for continuous models</a>, except
<p>
<ol>
<li>For the executable <b><font color="green">bw</font></b>, the flags -tst2cbfn, -topn and -feat must be set to the values
<p>
<table border="1">
<tr><td> FLAG </td><td> VALUE </td></tr>
<tr><td> -tst2cbfn </td><td> .semi. </td></tr>
<tr><td> -topn </td><td> This value should be lower than or equal to the
codebook size. It decides how many components of each mixture weight
distribution are used to estimate likelihoods during the baum-welch passes. Itaffects the speed of training. A higher value results in slower iterations </td></tr>
<tr><td> -feat </td><td> The specific feature type you are using to train
the semi-continuous models </td></tr>
</table>
<p>
<li>For the executable <b><font color="green">norm</font></b>, the flag -feat
must be set to the value
<p>
<table border="1">
<tr><td> FLAG </td><td> VALUE </td></tr>
<tr><td> -feat </td><td> The specific feature type you are using to train
the semi-continuous models </td></tr>
</table>
</ol>
<p>
Also, it is important to remember here that the re-estimated means and
variances now correspond to <em>codebook</em> means and variances. In
semi-continuous models, the codebooks are also re-estimated during training.
The vector quantization step is therefore only an <em>initialization</em>
step for the codebooks. This fact will affect the way we do model
adaptation for the semi-continuous case.
<p>
<a name="3g"></a>
<center><h4><font color="red">TRAINING SEMI-CONTINUOUS MODELS</font></h4></center>
<TABLE width="100%" bgcolor="#ffffff">
<td>CREATING THE CD UNTIED MODEL DEFINITION FILE</td>
</table>
<p>
<a href="#24">
This procedure is the same as described for continuous models</a>.
<a name="3h"></a>
<center><h4><font color="red">TRAINING SEMI-CONTINUOUS MODELS</font></h4></center>
<TABLE width="100%" bgcolor="#ffffff">
<td>FLAT INITIALIZATION OF CD UNTIED MODEL PARAMETERS</td>
</table>
<p>
<a href="#25">
This procedure is the same as described for continuous models</a>.
<a name="3i"></a>
<center><h4><font color="red">TRAINING SEMI-CONTINUOUS MODELS</font></h4></center>
<TABLE width="100%" bgcolor="#ffffff">
<td>TRAINING CD UNTIED MODELS</td>
</table>
<p>
<a href="#26">
This procedure is the same as described for continuous models</a>.
<a name="3j"></a>
<center><h4><font color="red">TRAINING SEMI-CONTINUOUS MODELS</font></h4></center>
<TABLE width="100%" bgcolor="#ffffff">
<td>BUILDING DECISION TREES FOR PARAMETER SHARING</td>
</table>
<p>
<a href="#27">
This procedure is the same as described for continuous models</a>.
<a name="3k"></a>
<center><h4><font color="red">TRAINING SEMI-CONTINUOUS MODELS</font></h4></center>
<TABLE width="100%" bgcolor="#ffffff">
<td>GENERATING THE LINGUISTIC QUESTIONS</td>
</table>
<p>
<a href="#28">
This procedure is the same as described for continuous models</a>.
<a name="3l"></a>
<center><h4><font color="red">TRAINING SEMI-CONTINUOUS MODELS</font></h4></center>
<TABLE width="100%" bgcolor="#ffffff">
<td>PRUNING THE DECISION TREES</td>
</table>
<p>
<a href="#29">
This procedure is the same as described for continuous models</a>.
<a name="3m"></a>
<center><h4><font color="red">TRAINING SEMI-CONTINUOUS MODELS</font></h4></center>
<TABLE width="100%" bgcolor="#ffffff">
<td>CREATING THE CD TIED MODEL DEFINITION FILE</td>
</table>
<p>
<a href="#30">
This procedure is the same as described for continuous models</a>.
<a name="3n"></a>
<center><h4><font color="red">TRAINING SEMI-CONTINUOUS MODELS</font></h4></center>
<TABLE width="100%" bgcolor="#ffffff">
<td>INITIALIZING AND TRAINING CD TIED MODELS</td>
</table>
<p>
During initialization, the model parameters from
the CI model parameter files are copied into appropriate positions
in the CD tied model parameter files. Four model parameter files
are created, one each for the means, variances, transition matrices
and mixture weights. During initialization, each state of a particular CI phone contributes
to the same state of the same CI phone in the CD-tied model parameter file,
and also to the same state of the *all* the triphones of the same CI
phone in the CD-tied model parameter file. The CD-tied model definition
file is used as a reference for this mapping.
<p>
Initialization for the CD-tied training is done by the executable called
<b><font color="green">init_mixw</font></b>. It requires the following
arguments:
<p>
<table border="1">
<tr><td valign="top"> -src_moddeffn </td><td> source (CI) model definition file
</td></tr>
<tr><td valign="top"> -src_ts2cbfn </td><td> .semi. </td></tr>
<tr><td valign="top"> -src_mixwfn </td><td> source (CI) mixture-weight file </td
></tr>
<tr><td valign="top"> -src_meanfn </td><td> source (CI) means file </td></tr>
<tr><td valign="top"> -src_varfn </td><td> source (CI) variances file </td></tr>
<tr><td valign="top"> -src_tmatfn </td><td> source (CI) transition-matrices file
</td></tr>
<tr><td valign="top"> -dest_moddeffn </td><td> destination (CD tied) model def
inition file </td></tr>
<tr><td valign="top"> -dest_ts2cbfn </td><td> .semi. </td></tr>
<tr><td valign="top"> -dest_mixwfn </td><td> destination (CD tied) mixture wei
ghts file </td></tr>
<tr><td valign="top"> -dest_meanfn </td><td> destination (CD tied) means file
</td></tr>
<tr><td valign="top"> -dest_varfn </td><td> destination (CD tied) variances fi
le </td></tr>
<tr><td valign="top"> -dest_tmatfn </td><td> destination (CD tied) transition
matrices file </td></tr>
<tr><td valign="top"> -feat </td><td> feature configuration </td></tr>
<tr><td valign="top"> -ceplen </td><td> dimensionality of base feature vector </
td></tr>
</table>
<p>
@@
<a name="3a"></a>
<center><h4><font color="red">TRAINING SEMI-CONTINUOUS MODELS</font></h4>
</center>
<TABLE width="100%" bgcolor="#ffffff">
<td>DELETED INTERPOLATION</td>
</table>
<p>
Deleted interpolation is the final step in creating semi-continuous models.
The output of deleted interpolation are semi-continuous models in sphinx-3
format. These have to be further converted to sphinx-2 format, if you want
to use the SPHINX-II decoder.
<p>
Deleted interpolation is an iterative process to interpolate between
CD and CI mixture-weights to reduce the effects of overfitting. The data are
divided into two sets, and the data from one set are used to estimate
the optimal interpolation factor between CI and CD models trained
from the other set. Then the two data sets are switched and this
procedure is repeated using the last estimated interpolation factor
as an initialization for the current step. The switching is continued
until the interpolation factor converges.
<p>
To do this, we need *two* balanced data sets. Instead of the actual data,
however, we use the Bauim-Welch buffers, since the related math is convenient.
we therefore need an *even* number of buffers that can be grouped into two
sets. DI cannot be performed if you train using only one buffer. At
least in the final iteration of the training, you must perform the training
in (at least) two parts. You could also do this serially
as one final iteration of training AFTER BW has converegd, on a non-lsf
setup.
<p>
Note here that the norm executable used at the end of every Baum-Welch
iteration also computes models from the buffers, but it does not require an
even number of buffers. BW returns numerator terms and denominator terms
for the final estimation, and norm performs the actual division. The number
of buffers is not important, but you would need to run norm at the end
of EVERY iteration of BW, even if you did
the training in only one part. When you have multiple parts norm sums up
the numerator terms from the various buffers, and the denominator terms,
and then does the division.
<p>
The executable "delint" provided with the SPHINX-III package does the
deleted interpolation. It takes the following arguments:
<table border="1">
<tr><td valign="top">FLAG</td><td>DESCRIPTION</td></tr>
<tr><td valign="top"> -accumdirs </td><td> directory which holds the baum-welch buffers </td></tr>
<tr><td valign="top"> -moddeffn </td><td> CD-tied model-definition file </td></tr>
<tr><td valign="top"> -mixwfn </td><td> CD-tied mixture weights files </td></tr>
<tr><td valign="top"> -cilambda </td><td> initial interpolation factor between the CI
models and the Cd models. It is the weight given given to the CI models
initially. The values range from 0 to 1. This is typically set to
0.9 </td></tr>
<tr><td valign="top"> -ceplen </td><td> dimentionality of base feature vector </td></tr>
<tr><td valign="top"> -maxiter </td><td> the number of iterations of deleted-interpolation
that you want to run. DI can be slow to converge, so this number is typically
between 1000-4000 </td></tr>
</table>
</ol>
(more to come...)
<em>After the decision trees are built using semi-continuous models, it is
possible to train continuous models. ci-semicontinuous models need to be
trained for initializing the semicontinuous untied models. ci-continuous
models need to be trained for initializing the continuous tied state
models. the feature set can be changed after the decision tree building
stage.
</em>
<a href="#top">Back to index</a>
<hr>
<a name="4"></a>
<TABLE width="100%" bgcolor="#ffffff"><td>
SPHINX2 data and model formats</td></table>
<ol>
<li>Feature set: This is a binary file with all the elements in each of the
vectors stored sequentially. The header is a 4 byte integer which tells us
how many floating point numbers there are in the file. This is followed by
the actual cepstral values (usually 13 cepstral values per frame, with 10ms
skip between adjacent frames. Framesize is usually fixed and is usually
25ms).
<pre>
<4_byte_integer header>
vec 1 element 1
vec 1 element 2
.
.
vec 1 element 13
vec 2 element 1
vec 2 element 2
.
.
vec 2 element 13
</pre>
<li> Sphinx2 semi-continuous HMM (SCHMM) formats:
<br>The sphinx II SCHMM format is rather complicated. It has the following
main components (each of which has sub-components):
<ul>
<li>A set of codebooks
<li>A "sendump" file that stores state (senone) distributions
<li> A "phone" and a "map" file which map senones on to states of a triphone
<li> A set of ".chmm" files that store transition matrices
</ul>
<ol>
<p>
<li>Codebooks: There are 8 codebook files. The sphinx-2 uses a four stream feature set:
<ul>
<li>cepstral feature: [c1-c12], (12 components)
<li> delta feature: [delta_c1-delta_c12,longterm_delta_c1-longterm_delta_c12],(24 components)
<li> power feature: [c0,delta_c0,doubledelta_c0], (3 components)
<li> doubledelta feature: [doubledelta_c-doubledelta_c12] (12 components)
</ul>
The 8 codebooks files store the means and variances of all the gaussians
for each of these 4 features. The 8 codebooks are,
<ul>
<li>cep.256.vec [this is the file of means for the cepstral feature]
<li> cep.256.var [this is the file of variacnes for the cepstral feature]
<li> d2cep.256.vec [this is the file of means for the delta cepstral feature]
<li> d2cep.256.var [this is the file of variances for the delta cepstral feature]
<li> p3cep.256.vec [this is the file of means for the power feature]
<li> p3cep.256.var [this is the file of variances for the power feature]
<li> xcep.256.vec [this is the file of means for the double delta feature]
<li> xcep.256.var [this is the file of variances for the double delta feature]
</ul>
All files are binary and have the following format:
[4 byte int][4 byte float][4 byte float][4 byte float]......
The 4 byte integer header stores the number of floating point values to
follow in the file. For the cep.256.var, cep.256.vec, xcep.256.var and
xcep.256.vec this value should be 3328. For d2cep.* it should be 6400,
and for p3cep.* it should be 768.
The floating point numbers are the components of the mean vectors (or
variance vectors) laid end to end. So cep.256.[vec,var] have 256 mean
(or variance) vectors, each 13 dimensions long,
d2cep.256.[vec,var] have 256 mean/var vectors, each 25 dimensions long,
p3cep.256.[vec,var] have 256 vectors, each of dimension 3,
xcep.256.[vec,var] have 256 vectors of length 13 each.
<p>
The 0th component of the cep,d2cep and xcep distributions are not used in
likelihood computation and are part of the format for purely historical
reasons.
<p>
<li> The "sendump" file: The "sendump" file stores the mixture weights of the states associated with
each phone. (this file has a little ascii header, which might help you
a little). Except for the header, this is a binary file. The mixture weights
have all been transformed to 8 bit integer by the following operation
intmixw = (-log(float mixw) >> shift)
The log base is 1.0003. The "shift" is the number of bits the smallest
mixture weight has to be shifted right to fit in 8 bits.
The sendump file stores,
<pre>
for each feature (4 features in all)
for each codeword (256 in all)
for each ci-phone (including noise phones)
for each tied state associated with ci phone,
probability of codeword in tied state
end
for each CI state associated with ci phone, ( 5 states )
probability of codeword in CI state
end
end
end
end
</pre>
The sendump file has the following storage format (all data, except for
the header string are binary):
<pre>
Length of header as 4 byte int (including terminating '\0')
HEADER string (including terminating '\0')
0 (as 4 byte int, indicates end of header strings).
256 (codebooksize, 4 byte int)
Num senones (Total number of tied states, 4 byte int)
[lut[0], (4 byte integer, lut[i] = -(i"<<"shift))
prob_of_codeword[0]_of_feat[0]_1st_CD_sen_of_1st_ciphone (unsigned char)
prob_of_codeword[0]_of_feat[0]_2nd_CD_sen_of_1st_ciphone (unsigned char)
..
prob_of_codeword[0]_of_feat[0]_1st_CI_sen_of_1st_ciphone (unsigned char)
prob_of_codeword[0]_of_feat[0]_2nd_CI_sen_of_1st_ciphone (unsigned char)
..
prob_of_codeword[0]_of_feat[0]_1st_CD_sen_of_2nd_ciphone (unsigned char)
prob_of_codeword[0]_of_feat[0]_2nd_CD_sen_of_2nd_ciphone (unsigned char)
..
prob_of_codeword[0]_of_feat[0]_1st_CI_sen_of_2st_ciphone (unsigned char)
prob_of_codeword[0]_of_feat[0]_2nd_CI_sen_of_2st_ciphone (unsigned char)
..
]
[lut[1], (4 byte integer)
prob_of_codeword[1]_of_feat[0]_1st_CD_sen_of_1st_ciphone (unsigned char)
prob_of_codeword[1]_of_feat[0]_2nd_CD_sen_of_1st_ciphone (unsigned char)
..
prob_of_codeword[1]_of_feat[0]_1st_CD_sen_of_2nd_ciphone (unsigned char)
prob_of_codeword[1]_of_feat[0]_2nd_CD_sen_of_2nd_ciphone (unsigned char)
..
]
... 256 times ..
Above repeats for each of the 4 features
</pre>
<p>
<li> PHONE file: The phone file stores a list of phones and triphones used by
the decoder. This is an ascii file
It has 2 sections.
The first section lists the CI phones in the models
and consists of lines of the format
<pre>
AA 0 0 8 8
</pre>
"AA" is the CI phone, the first "0" indicates that it is a CI phone,
the first 8 is the index of the CI phone, and the last 8 is the
line number in the file.
The second 0 is there for historical reasons.
<p>
The second section lists TRIPHONES
and consists of lines of the format
<pre>
A(B,C)P -1 0 num num2
</pre>
"A" stands for the central phone, "B" for the left context, and
"C" for the right context phone. The "P" stands for the position of
the triphone and can take 4 values "s","b","i", and "e", standing
for single word, word beginning, word internal, and word ending triphone.
The -1 indicates that it is a triphone and not a CI phone. num
is the index of the CI phone "A", and num2 is the position of the
triphone (or ciphone) in the list, essentially the number of the
line in the file (beginning with 0).
<p>
<li> map file: The "map" file stores a mapping table to show which senones each state of
each triphone are mapped to. This is also an ascii file with lines of the form
<pre>
AA(AA,AA)s<0> 4
AA(AA,AA)s<1> 27
AA(AA,AA)s<2> 69
AA(AA,AA)s<3> 78
AA(AA,AA)s<4> 100
</pre>
The first line indicates that the 0th state of the triphone "AA" in the
context of "AA" and "AA" is modelled by th 4th senone associated
with the CI phone AA. Note that the numbering is specific to the
CI phone. So the 4th senone of "AX" would also be numbered 4 (but
this should not cause confusion)
<p>
<li> chmm FILES: There is one *.chmm file per ci phone. Each stores the transition matrix
associated with that partiular ci phone in following binary format.
(Note all triphones associated with a ci phone share its transition matrix)
(all numbers are 4 byte integers):
<ul>
<li> -10 (a header to indicate this is a tmat file)
<li> 256 (no of codewords)
<li>5 (no of emitting states)
<li>6 (total no. of states, including non-emitting state)
<li> 1 (no. of initial states. In fbs8 a state sequence can only begin
with state[0]. So there is only 1 possible initial state)
<li>0 (list of initial states. Here there is only one, namely state 0)
<li>1 (no. of terminal states. There is only one non-emitting terminal state)
<li>5 (id of terminal state. This is 5 for a 5 state HMM)
<li>14 (total no. of non-zero transitions allowed by topology)
<pre>
[0 0 (int)log(tmat[0][0]) 0] (source, dest, transition prob, source id)
[0 1 (int)log(tmat[0][1]) 0]
[1 1 (int)log(tmat[1][1]) 1]
[1 2 (int)log(tmat[1][2]) 1]
[2 2 (int)log(tmat[2][2]) 2]
[2 3 (int)log(tmat[2][3]) 2]
[3 3 (int)log(tmat[3][3]) 3]
[3 4 (int)log(tmat[3][4]) 3]
[4 4 (int)log(tmat[4][4]) 4]
[4 5 (int)log(tmat[4][5]) 4]
[0 2 (int)log(tmat[0][2]) 0]
[1 3 (int)log(tmat[1][3]) 1]
[2 4 (int)log(tmat[2][4]) 2]
[3 5 (int)log(tmat[3][5]) 3]
</pre>
There are thus 65 integers in all, and so each *.chmm file should be
65*4 = 260 bytes in size.
</ul>
</ol>
(more to come...)
<p>
<a href="#top">Back to index</a>
<hr>
<a name="4b"></a>
<TABLE width="100%" bgcolor="#ffffff"><td>
SPHINX3 data and model formats</td></table>
<ol>
<p>
All senone-ids in the model files are with reference to the corresponding
model-definition file for the model-set.
<p>
<b><u>The means file</b></u>
<p>
The ascii means file for 8 Gaussians/state 3-state HMMs
looks like this:
<pre>
param 602 1 8
mgau 0
feat 0
density 0 6.957e-01 -8.067e-01 -6.660e-01 3.402e-01 -2.786e-03 -1.655e-01
2.2 56e-02 9.964e-02 -1.237e-01 -1.829e-01 -3.777e-02 1.532e-03 -9.610e-01
-3.883e-0 1 5.229e-01 2.634e-01 -3.090e-01 4.427e-02 2.638e-01 -4.245e-02
-1.914e-01 -5.52 1e-02 8.603e-02 3.466e-03 5.120e+00 1.625e+00 -1.103e+00
1.611e-01 5.263e-01 2.4 79e-01 -4.823e-01 -1.146e-01 2.710e-01 -1.997e-05
-3.078e-01 4.220e-02 2.294e-01
1.023e-02 -9.163e-02
density 1 5.216e-01 -5.267e-01 -7.818e-01 2.534e-01 6.536e-02 -1.335e-01
-1.3 22e-01 1.195e-01 5.900e-02 -2.095e-01 -1.349e-01 -8.872e-02 -4.965e-01
-2.829e-0 1 5.302e-01 2.054e-01 -2.669e-01 -2.415e-01 2.915e-01 1.406e-01
-1.572e-01 -1.50 1e-01 2.426e-02 1.074e-01 5.301e+00 7.020e-01 -8.537e-01
1.448e-01 3.256e-01 2.7 09e-01 -3.955e-01 -1.649e-01 1.899e-01 1.983e-01
-2.093e-01 -2.231e-01 1.825e-01
1.667e-01 -2.787e-02
density 2 5.844e-01 -8.953e-01 -4.268e-01 4.602e-01 -9.874e-02 -1.040e-01
-3. 739e-02 1.566e-01 -2.034e-01 -8.387e-02 -3.551e-02 4.647e-03
-6.439e-01 -8.252e- 02 4.776e-01 2.905e-02 -4.012e-01 1.112e-01 2.325e-01
-1.245e-01 -1.147e-01 3.39 0e-02 1.048e-01 -7.266e-02 4.546e+00 8.103e-01
-4.168e-01 6.453e-02 3.621e-01 1. 821e-02 -4.503e-01 7.951e-02 2.659e-01
-1.085e-02 -3.121e-01 1.395e-01 1.340e-01
-5.995e-02 -7.188e-02
.....
.....
density 7 6.504e-01 -3.921e-01 -9.316e-01 1.085e-01 9.951e-02 7.447e-02
-2.42 3e-01 -8.710e-03 7.210e-02 -7.585e-02 -9.116e-02 -1.630e-01
-3.008e-01 -3.175e-0 1 1.687e-01 3.389e-01 -3.703e-02 -2.052e-01 -3.263e-03
1.517e-01 8.243e-02 -1.40 6e-01 -1.070e-01 4.236e-02 5.143e+00 5.469e-01
-2.331e-01 1.896e-02 8.561e-02 1. 785e-01 -1.197e-01 -1.326e-01 -6.467e-02
1.787e-01 5.523e-02 -1.403e-01 -7.172e- 02 6.666e-02 1.146e-01
mgau 1
feat 0
density 0 3.315e-01 -5.500e-01 -2.675e-01 1.672e-01 -1.785e-01 -1.421e-01
9.0 70e-02 1.192e-01 -1.153e-01 -1.702e-01 -3.114e-02 -9.050e-02 -1.247e-01
3.489e-0 1 7.102e-01 -2.001e-01 -1.191e-01 -6.647e-02 2.222e-01 -1.866e-01
-1.067e-01 1.0 52e-01 7.092e-02 -8.763e-03 5.029e+00 -1.354e+00 -2.135e+00
2.901e-01 5.646e-01 1.525e-01 -1.901e-01 4.672e-01 -3.508e-02 -2.176e-01
-2.031e-01 1.378e-01 1.029e -01 -4.655e-02 -2.512e-02
density 1 4.595e-01 -8.823e-01 -4.397e-01 4.221e-01 -2.269e-03 -6.014e-02
-7. 198e-02 9.702e-02 -1.705e-01 -6.178e-02 -4.066e-02 9.789e-03
-3.188e-01 -8.284e- 02 2.702e-01 6.192e-02 -2.077e-01 2.683e-02 1.220e-01
-4.606e-02 -1.107e-01 1.16 9e-02 8.191e-02 -2.150e-02 4.214e+00 2.322e-01
-4.732e-02 1.834e-02 8.372e-02 -7 .559e-03 -1.111e-01 -3.453e-03 5.487e-02
2.355e-02 -8.777e-02 4.309e-02 3.460e-0 2 -1.521e-02 -3.808e-02
</pre>
This is what it means, reading left to right, top to bottom:
<p>
Parameters for 602 tied-states (or senones), 1 feature stream,
8 Gaussians per state.
<p>
Means for senone no. 0, feature-stream no. 0.
Gaussian density no. 0, followed by its 39-dimensional mean vector.
(Note that each senone is a mixture of 8 gaussians, and each
feature vector consists of 13 cepstra, 13 delta cepstra and 13
double delta cepstra)
<pre>
Gaussian density no. 1, followed by its 39-dimensional mean vector.
Gaussian density no. 2, followed by its 39-dimensional mean vector.
.....
.....
Gaussian density no. 7, followed by its 39-dimensional mean vector.
Means for senone no. 1, feature-stream no. 0.
Gaussian density no. 0, followed by its 39-dimensional mean vector.
Gaussian density no. 1, followed by its 39-dimensional mean vector.
</pre>
- and so on -
<p>
<b><u>The variances file</b></u>
<pre>
param 602 1 8
mgau 0
feat 0
density 0 1.402e-01 5.048e-02 3.830e-02 4.165e-02 2.749e-02 2.846e-02
2.007e- 02 1.408e-02 1.234e-02 1.168e-02 1.215e-02 8.772e-03 8.868e-02
6.098e-02 4.579e- 02 4.383e-02 3.646e-02 3.460e-02 3.127e-02 2.336e-02
2.258e-02 2.015e-02 1.359e- 02 1.367e-02 1.626e+00 4.946e-01 3.432e-01
7.133e-02 6.372e-02 4.693e-02 6.938e- 02 3.608e-02 3.147e-02 4.044e-02
2.396e-02 2.788e-02 1.934e-02 2.164e-02 1.547e- 02
density 1 9.619e-02 4.452e-02 6.489e-02 2.388e-02 2.337e-02 1.831e-02
1.569e- 02 1.559e-02 1.082e-02 1.008e-02 6.238e-03 4.387e-03 5.294e-02
4.085e-02 3.499e- 02 2.327e-02 2.085e-02 1.766e-02 1.781e-02 1.315e-02
1.367e-02 9.409e-03 7.189e- 03 4.893e-03 1.880e+00 3.342e-01 3.835e-01
5.274e-02 4.430e-02 2.514e-02 2.516e- 02 2.863e-02 1.982e-02 1.966e-02
1.742e-02 9.935e-03 1.154e-02 8.361e-03 8.059e- 03
density 2 1.107e-01 5.627e-02 2.887e-02 2.359e-02 2.083e-02 2.143e-02
1.528e- 02 1.264e-02 1.223e-02 9.553e-03 9.660e-03 9.241e-03 3.391e-02
2.344e-02 2.220e- 02 1.873e-02 1.436e-02 1.458e-02 1.362e-02 1.350e-02
1.191e-02 1.036e-02 8.290e- 03 5.788e-03 1.226e+00 1.287e-01 1.037e-01
3.079e-02 2.692e-02 1.870e-02 2.873e- 02 1.639e-02 1.594e-02 1.453e-02
1.043e-02 1.137e-02 1.086e-02 8.870e-03 9.182e- 03
</pre>
- and so on -
The format is exactly as for the means file.
<p>
<b><u>The mixture_weights file</b></u>
<p>
The ascii mixure_weights file for 8 Gaussians/state 3-state HMMs
looks like this:
<pre>
mixw 602 1 8
mixw [0 0] 7.434275e+03
8.697e+02 9.126e+02 7.792e+02 1.149e+03 9.221e+02 9.643e+02 1.037e+03 8.002e+02
mixw [1 0] 8.172642e+03
8.931e+02 9.570e+02 1.185e+03 1.012e+03 1.185e+03 9.535e+02 7.618e+02 1.225e+03
</pre>
This is what it means, reading left to right, top to bottom:
<p>
Mixtrue weights for 602 tied-states (or senones), 1 feature stream,
8 Gaussians per state (Each mixture weight is a vector with 8 components)
<p>
Mixture weights for senone no. 0, feature-stream no. 0, number of
times this senone occured in the training corpus (instead of
writing normalized values, this number is directly recorded since it
is useful in other places during training [interpolation, adaptation,
tree building etc]).
When normalized (for example, by the decoder during decoding), the
mixture weights above would read as:
</pre>
mixw 602 1 8
mixw [0 0] 7.434275e+03
1.170e-01 1.228e-01 1.048e-01 1.546e-01 1.240e-01 1.297e-01 1.395e-01 1.076e-01
mixw [1 0] 8.172642e+03
1.093e-01 1.171e-01 1.450e-01 1.238e-01 1.450e-01 1.167e-01 9.321e-02 1.499e-01
</pre>
<p>
<b><u>The transition_matrices file</b></u>
<p>
The ascii file looks like this:
<pre>
tmat 34 4
tmat [0]
6.577e-01 3.423e-01
6.886e-01 3.114e-01
7.391e-01 2.609e-01
tmat [1]
8.344e-01 1.656e-01
7.550e-01 2.450e-01
6.564e-01 3.436e-01
tmat [2]
8.259e-01 1.741e-01
7.598e-01 2.402e-01
7.107e-01 2.893e-01
tmat [3]
4.112e-01 5.888e-01
4.371e-01 5.629e-01
5.623e-01 4.377e-01
</pre>
- and so on -
This is what it means, reading left to right, top to bottom:
<p>
Transition matrices for 34 HMMs, each with four states (3 emitting states +
1 non-emitting state)
<p>
Transition matrix for HMM no. 0 (NOTE THAT THIS IS THE HMM NUMBER, AND
NOT THE SENONE NUMBER), matrix.
<pre>
Transition matrix for HMM no 1, matrix.
Transition matrix for HMM no 2, matrix.
Transition matrix for HMM no 3, matrix.
</pre>
- and so on -
<p>
<b><u>Explanation of the feature-vector components:</b></u>
<p>
The 13 dimensional cepstra, 13 dimensional delta cepstra and
13 dimensional double-delta cepstra are arranged, in all model
files, in the following order:
1s_12c_12d_3p_12dd (you can denote this by s3_1x39 in the
decoder flags).
The format string means: 1 feature-stream, 12 cepstra, 12 deltacepstra,
3 power and 12 doubledeltacepstra.
The power part is composed of the 0th component of the cepstral
vector, 0th component of the d-cepstral vector and 0th component
of the dd-cepstral vector.
<p>
In the quantized models, you will see the string
24,0-11/25,12-23/26,27-38
In this string, the slashes are delimiters. The numbers represent various
components occuring in each of the 3 codebooks. In the above string,
for instance, the first codebook is composed of the 24th component of
the feature vector (s3_1x39) followed by components 0-11. The second
codeword has components 25, followed by components 12-23, and the
third codeword is composed of components 26 and 27-28. This basically
accounts for the odd order in s3_1x39. By constructing the codewords
in this manner, we ensure that the first codeword is composed entirely
of cepstral terms, the second codeword of delta cepstral terms and the
third codeword of double delta terms.
<p>
s3_1x39 is a historical order. It can be disposed of in any new code that
you write.
Writing the feature vector components in different
orders has no effect on recognition, provided training and
test feature formats are the same.
<p>
<a name="5"></a>
<TABLE width="100%" bgcolor="#ffffff"><td>
TRAINING MULTILINGUAL MODELS</td></table>
Once you have acoustic data and the corresponding transcriptions for any
language, and a lexicon which translates words used in the transcription
into sub-word units (or just maps them into some reasonable-looking
acoustic units), you can use the SPHINX to train acoustic models for that
language. You do not need anything else.
<p>
The linguistic questions that are needed for building the decision trees
are automatically designed by the SPHINX. Given the acoustic units you
choose to model, the SPHINX can automatically determine the best
combinations of these units to compose the questions. The hybrid algorithm
that the SPHINX uses clusters state distributions of context-independent
phones to obtain questions for triphonetic contexts. This is very useful
if you want to train models for languages whose phonetic structure you do
not know well enough to design your own phone classes (or if a phonetician
is not available to help you do it). An even greater advantage comes from
the fact that the algorithm can be effectively used in situations where the
subword units are not phonetically motivated. Hence you can comfortably
use any set of acoustic units that look reasonable to you for the
task.
<p>
If you are completely lost about the acoustic units but have enough
training data for all (or most) words used in the transcripts, then build
word models instead of subword models. You do not have to build decision
trees. Word models are usually context-independent models, so you only have
to follow through the CI training. Word models do have some limitations,
which are currently discussed in the non-technical version of this manual.
<p>
<a href="#top">Back to index</a>
<hr>
<a name="6"></a>
<TABLE width="100%" bgcolor="#ffffff"><td>
THE TRAINING LEXICON</td></table>
Inconsistencies in the training lexicon can result in bad acoustic models.
Inconsistencies stem from the usage of a phoneset with phones that are
confusible in the pattern space of our recognizer. To get an idea about the
confusibility of the phones that you are using, look at the per-frame log
likelihoods of the utterances during training. A greater number of phones
in the lexicon should ordinarily result in higher log likelihoods. If you
have a baseline to compare with, and this is *not* the case, then it means
that the phoneset is more diffuse over the pattern space (more compact, if
you observe the opposite for a smaller phone set), and the corresponding
distributions are wider (sharper in the other case). Generally, as the
number of applicable distributions decreases over a given utterance, the
variances tend to become larger and larger. The distributions flatten out
since the areas under the distributions are individually conserved (to
unity) and so the overall per frame likelihoods are expected to be lower.
<p>
The solution is to fix the phoneset, and to redo the lexicon in terms of a
phoneset of smaller size covering the acoustic space in a more compact
manner. One way to do this is to collapse the lexicon into syllables and
longer units and to expand it again using a changed and smaller
phoneset. The best way to do this is still a research problem, but if you
are a native speaker of the language and have a good ear for sounds, your
intuition will probably work. The SPHINX will, of course, be able to train
models for any new phoneset you come up with.
<p><a href="#top">Back to index</a>
<hr>
<a name="7"></a>
<TABLE width="100%" bgcolor="#ffffff"><td>
CONVERTING SPHINX3 FORMAT MODELS TO SPHINX2 FORMAT</td></table>
To convert the 5 state/HMM, 4 feature stream semi-continuous models
trained using the Sphinx3 trainer into the Sphinx2 format (compatible
with the Sphinx2 decoder), programs in the following directories
must be compiled and used:
<pre>
-----------------------------------------------------------------------
program directory corresponding function
executable of executable
-----------------------------------------------------------------------
mk_s2cb mk_s2cb makes s2 codebooks
mk_s2hmm mk_s2hmm makes s2 mixture weights
mk_s2phone mdef2phonemap makes phone and map files
mk_s2seno makesendmp makes senone dmp files
-----------------------------------------------------------------------
Variables needed:
-----------------
s2dir : sphinx_2_format directory
s3dir : sphinx_3_format directory
s3mixw : s3dir/mixture_weights
s3mean : s3dir/means
s3var : s3dir/variances
s3tmat : s3dir/transition_matrices
s3mdef : s3dir/mdef_file (MAKE SURE that this mdef file
includes all the phones/triphones needed for
the decode. It should ideally be made from
the decode dictionary, if the decode vocabulary
is fixed)
Usage:
------
mk_s2cb
-meanfn s3mean
-varfn s3var
-cbdir s2dir
-varfloor 0.00001
mk_s2hmm
-moddeffn s3mdef
-mixwfn s3mixw
-tmatfn s3tmat
-hmmdir s2dir
makesendmp
s2_4x $s3mdef .semi. $s3mixw 0.0000001 $s2dir/sendump
(the order is important)
cleanup: s2dir/*.ccode s2dir/*.d2code s2dir/*.p3code s2dir/*.xcode
mdef2phonemap
grep -v "^#" s3mdef | mdef2phonemap s2dir/phone s2dir/map
</pre>
make sure that the mdef file used in the programs above includes all the
triphones needed. The programs (especially the makesendmp program) will not
work if any tied state is missing from the mdef file. This can happen if
you ignore the dictionary provided with the models and try to make a
triphone list using another dictionary. Even though you may have the same
phones, there may be enough triphones missing to leave out some leaves in
the pruned trees altogether (since they cannot be associated with any of
the new triphones states). To avoid this, use the dictionary provided. You
may extend it by including new words.
<hr>
<a name="8"></a>
<TABLE width="100%" bgcolor="#ffffff"><td>
UPDATING OR ADAPTING EXISTING MODELS SETS</td></table>
In general one is better off training speaker specific models if sufficient
data (at least 8-10 hours) are available. If you have less data for
a speaker or a domain, then the better option is to adapt any existing
models you have to the data. Exactly how you adapt would depend on the kind
of acoustic models you're using. If you're using semi-continuous
models, adaptation could be performed by interpolating speaker specific
models with speaker-independent models. For continuous HMMs you would have
to use MLLR, or one of its variants.
To adapt or update existing semicontinuous models, follow these steps:
<p>
<ol>
<li> Compute features for the new training data. The features must be
computed in the same manner as your old training features. In fact, the
feature computation in the two cases must be identical as far as possible.
<li> Prepare transcripts and dictionary for the new data. The dictionary must
have the same phoneset as was used for training the models. The transcripts
must also be prepared in the same manner. If you have new filler phones
then the fillerdict must map them to the old filler phones.
<li> The new training transcript and the corresponding ctl file can include
the old training data IF all you are doing is using additional data
from the SAME domain that you might have recently acquired. If you
are adapting to a slightly different domain or slightly different acoustic
conditions, then use only the new data.
<li> Starting with the existing deleted-interpolated models, and using the
same tied mdef file used for training the base models and the same
training parameters like the difference features, number of streams etc.,
run through one
or two passes of Baum-Welch. However, this must be done without
re-estimating the means and variances. Only the mixture-weights must be
re-estimated. If you are running the norm after the Baum-Welch, then
make sure that the norm executable is set to normalize only the mixture
weights.
<li> Once the mixture weights are re-estimated, the new mixture weights must be
interpolated with the ones you started with. The executable "mixw_interp"
provided with the SPHINX package may be used for this. You can experiment
with various mixing weights to select the optimal one.
This is of course the simplest update/adaptation technique. There are more
sophisticated techniques which will be explained here later.
</ol>
<p>
<p>
<b>The <font color="green">mixw_interp</font> executable</b>:
<p>
This is used in model adaptation for interpolating between two mixture weight
files. It requires the following flags:
<table border="1">
<tr valign="top"><td>FLAG</td><td>DESCRIPTION</td></tr>
<tr valign="top"><td>-SImixwfn</td><td>The original Speaker-Independent mixture weights file</td></tr>
<tr valign="top"><td>-SDmixwfn</td><td>The Speaker Dependent mixture weight file that you have after the bw iterations for adaptation </td></tr>
<tr valign="top"><td>-tokencntfn</td><td>The token count file</td</tr>
<tr valign="top"><td>-outmixwfn</td><td>The output interpolated mixture weight parameter file name</td></tr>
<tr valign="top"><td>-SIlambda</td><td>Weight given to SI mixing weights</td></tr>
</table>
<hr>
<a name="9"></a>
<TABLE width="100%" bgcolor="#ffffff"><td>
USING THE SPHINX-III DECODER WITH SEMI-CONTINUOUS AND CONTINUOUS MODELS</td></table>
There are two flags which are specific to the type of model being used, the
rest of the flags are independent of model type. The flags you need to change
to switch from continuous models to semi-continuous ones are:
<p>
<ul>
<li>the -senmgaufn flag would change from ".cont." to ".semi."
<li>the -feat flag would change from the feature you are using with continuous
models to the feature you are using with
the semicontinuous models (usually it is s3_1x39 for continuous models
and s2_4x for semi-continuous models)
</ul>
<p>
Some of the other decoder flags and their usual settings are as follows:
<pre>
-logbase 1.0001 \
-bestpath 0 \
-mdeffn $mdef \
-senmgaufn .cont. \
-meanfn $ACMODDIR/means \
-varfn $ACMODDIR/variances \
-mixwfn $ACMODDIR/mixture_weights \
-tmatfn $ACMODDIR/transition_matrices \
-langwt 10.5 \
-feat s3_1x39 \
-topn 32 \
-beam 1e-80 \
-nwbeam 1e-40 \
-dictfn $dictfn \
-fdictfn $fdictfn \
-fillpenfn $fillpenfn \
-lmfn $lmfile \
-inspen 0.2 \
-ctlfn $ctlfn \
-ctloffset $ctloffset \
-ctlcount $ctlcount \
-cepdir $cepdir \
-bptblsize 400000 \
-matchsegfn $matchfile \
-outlatdir $outlatdir \
-agc none \
-varnorm yes \
</pre>
<p>
<a name="04"></a>
<!------------------------------------------------------------------------->
<center><h4><font color="red">BEFORE YOU TRAIN</font></h4></center>
<TABLE width="100%" bgcolor="#ffffff">
<td>FORCE-ALIGNMENT</td>
</table>
<!------------------------------------------------------------------------->
Multiple pronunciations are not automatically considered in
the SPHINX. You have to mark the right pronunciations in the
transcripts and insert the interword silences. For this
<p>
a) remove the non-silence fillers from your filler dictionary and
put them in your regular dictionary
<p>
b) Remove *all* silence markers (<s>, <sil> and </s>) from your
training transcripts
<p>
For faligning with semi-continuous models, use the binary s3align provided
with the trainer package with the following flag settings. For faligning with
continuous models, change the settings of the flags -senmgaufn (.cont.),
-topn (no. of Gaussians in the Gaussian mixture modeling each HMM state),
-feat (the correct feature set):
<pre>
-outsentfn <faligned transcripts filename>
-insentfn <transcript filename>
-ctlfn <ctl file corresponding to trascript file>
-ctloffset 0
-ctlcount <no. of entries in ctl file >
-cepdir <feature files directory>
-dictfn <dictionary>
-fdictfn < filler dictionary>
-mdeffn <mdef file name>
-senmgaufn .semi.
-meanfn <model directory/means>
-varfn <model directory/variances>
-mixwfn <model directory/mixture_weights>
-tmatfn <model directory/transition_matrices>
-topn 4
-feat s2_4x
-beam 1e-90
-agc <max or none>
-cmn <none or current>
-logfn <logfile name>
</pre>
<p>
<em> last modified: 22 Nov. 2000 </em>
</body>
</html>
|