/usr/share/doc/sphinx3/s3_description.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 | <!DOCTYPE HTML PUBLIC "-//w3c//DTD HTML 4.01//EN">
<html>
<head>
<title>Sphinx-3 s3.X Decoder (X=7)</title>
<style type="text/css">
body { margin-left: 6%; margin-right: 3%; background: white; }
h1 { color: green; align: center; }
h2 { margin-left: -6%; color: green; margin-top: 6em; }
h3 { margin-left: -3%; margin-top: 1em; }
h4,h5,h6 { margin-top: 1em; }
pre { margin-left: 3%; font-family: monospace; }
code { font-family: monospace; font-weight: bold; }
div.wheatbox { background: wheat; padding: 0.5em; border: solid; border-width: thin; }
div.silverbox { background: silver; padding: 0.5em; border: solid; border-width: thin; }
div.endsec { background: wheat; border: solid; border-width: thin; }
</style>
</head>
<BODY>
<H1><U>Sphinx-3 s3.X Decoder (X=6)</U></H1>
<center>
Mosur K. Ravishankar (<em>aka</em> Ravi Mosur)<br>
Sphinx Speech Group<br>
School of Computer Science<br>
Carnegie Mellon University<br>
Pittsburgh, PA 15213<br>
</center>
<H2><U>Contents</U></H2>
<div class="wheatbox">
<UL>
<LI><A HREF="#sec_intro">Introduction</A></LI>
<LI><A HREF="#sec_decoverview">Overview of the s3.X Decoder</A>
<UL>
<LI><A HREF="#sec_dec_input">Inputs</A></LI>
<LI><A HREF="#sec_dec_output">Outputs</A></LI>
</UL>
</LI>
<LI><A HREF="#sec_compile">Compiling s3.X</A></LI>
<LI><A HREF="#sec_exec">Running s3.X</A>
<UL>
<LI><A HREF="#sec_args_overview">Configuration Arguments Overview</A></LI>
<LI><A HREF="#sec_dec_op">Decoder Operation</A></LI>
<LI><A HREF="#sec_dec_tune">Performance Tuning</A>
<UL>
<LI><A HREF="#sec_tune_prune">Tuning the Pruning Behaviour</a></LI>
<LI><A HREF="#sec_tune_lw">Tuning Language Model Related Parameters</a></LI>
</UL>
</LI>
<LI><A HREF="#sec_dec_errors">Some Common Errors and Failure Modes</A></LI>
</UL>
</LI>
<LI><A HREF="#sec_dict">Pronunciation Lexicon</A></LI>
<LI><A HREF="#sec_am">Acoustic Model</A></LI>
<LI><A HREF="#sec_lm">Language Model</A></LI>
<LI><A HREF="#sec_ctl">Speech Input Control File</A></LI>
<LI><A HREF="#sec_hypseg">Recognition Hypothesis Output</A></LI>
<LI><A HREF="#sec_wordlat">Word Lattice Output</A></LI>
<LI><A HREF="#sec_utilpgm">Other Utilities</A>
<UL>
<LI><A HREF="#sec_gausubvq">Gaussian Sub-Vector Quantization Utility</a></LI>
</UL>
</LI>
<LI><A HREF="#sec_src">Source Code</A></LI>
</UL>
</div>
<H2><A NAME="sec_intro"><U>Introduction</U></A></H2>
<P>Sphinx-3 is the successor to the Sphinx-II speech recognition
system from Carnegie Mellon University. It includes both an
acoustic <em>trainer</em> and various <em>decoders</em>,
<em>i.e.</em>, text recognition, phoneme recognition, N-best list
generation, etc. In this document, "Sphinx-3" refers to any
version of the Sphinx-3 decoder, and "s3.X" refers to the version
available in this distribution. Notice that s3.X is in fact a
branch from Sphinx-3, not a more recent release.
</P>
<P>The s3.X decoder is a recent implementation for speech-to-text
recognition, its main goal being speed improvements over the
original Sphinx-3 decoder. It runs about 10 times faster than the
latter on large vocabulary tasks. The following is a brief
summary of its main features and limitations:
</P>
<UL>
<LI>5-10x real-time recognition time on large vocabulary tasks</LI>
<LI>Limited to fully continuous acoustic models</LI>
<LI>Limited to 3 or 5-state left-to-right HMM topologies</LI>
<LI>Bigram or trigram language model</LI>
<LI>Batch-mode or live operation from pre-recorded speech</LI>
</UL>
<P> After s3.5, s3.X decoder also starts to integrate both the
flat-lexicon decoder search, tree decoder search by wrapping them
under the same interface. An implementation of finite-state
transducer search is also available under the same condition.
</P>
<P> All of the decoding routines could be accessible under the
executable <code> sphinx3_decode </code> through using the <code>
-op_mode </mode> options. (-op_mode 2: FST, -op_mode 3: Flat
Lexicon Decoder, -op_mode: Tree Lexicon Decoder) The original
flat-lexicon decoder interface still exists for backward
compatibility purpose.
</P>
<P>This package contains the following programs:
</P>
<OL>
<LI><code>sphinx3_decode</code>: The Sphinx-3 s3.2/s3.3/s3.X decoder processing cesptra files</li>
<LI><code>sphinx3_decode_anytopo</code>: The Sphinx-3 s3.0 decoder processing cesptra files (for backward compatibility purpose) </li>
<LI><code>sphinx3_continuous</code>: The Sphinx-3 live mode demo and ready for simple speak-and-decode application. </li>
<LI><code>sphinx3_gausubvq</code>: Sub-vector clustered acoustic model building</LI>
<LI><code>sphinx3_livedecode</code>: The Sphinx-3 s3.X decoder in live mode</LI>
<LI><code>sphinx3_livepretend</code>: The Sphinx-3 s3.X decoder in batch mode</LI>
<LI><code>sphinx3_align</code>: The Sphinx-3 aligner</LI>
<LI><code>sphinx3_allphone</code>: The Sphinx-3 phoneme recognizer</LI>
<LI><code>sphinx3_astar</code>: The Sphinx-3 N-best generator</LI>
<LI><code>sphinx3_dag</code>: The Sphinx-3 application for best-path searching</LI>
<LI><code>lm_convert</code>: A program that could convert the DMP and TXT-format of LM </LI>
</OL>
<P>This distribution has been prepared for Unix platforms. Port to
MS Windows (MS Visual C++ 6.0 workspace and project files) has
been provided.
</P>
<P>This document is a brief user's manual for the above programs.
It is <em>not</em> meant to be a detailed description of the
decoding algorithm, or an in-depth tutorial on speech recognition
technology. However, a set of Microsoft PowerPoint <a
href="s3-2.ppt">slides</a> are available that give additional
information about the decoder. Even though the slides refer to
s3.2, keep in mind that the basic search structure remains te same
in s3.X (where x=3 to 6).
</P>
<P>The initial part of this document provides an overview of the
decoder. It is followed by descriptions of the main input and
output databases; <em>i.e.</em>, the lexicon, language model, acoustic
model, etc.
</P>
<div class="endsec">
¤
<a href="#sec_intro">Back to top of this section</a>
</div>
<H2><A NAME="sec_decoverview"><U>Overview of the s3.X Decoder</U></A></H2>
<P>The s3.X decoder is based on the conventional <em>Viterbi
search</em> algorithm and <em>beam search</em> heuristics. It
uses a <em>lexical-tree</em> search structure somewhat like the
Sphinx-II decoder, but with some improvements for greater accuracy
than the latter. It takes its input from pre-recorded speech in
raw PCM format and writes its recognition results to output files.
</P>
<H3><A NAME="sec_dec_input"><U>Inputs</U></A></H3>
<P>We first give a brief outline of the input and output
characteristics of the decoder. More detailed information is
available in later sections. The decoder needs the following
inputs:
</P>
<UL>
<LI><a href="#sec_dict"><em>Lexical model</em></a>: The
lexical or pronunciation model contains pronunciations for all
the words of interest to the decoder. Like most modern speech
recognition systems, Sphinx-3 uses <em>phonetic units</em> to
build word pronunciations. Currently, the pronunciation lexicon
is almost entirely hand-crafted.
</LI>
<P></P>
<LI><a href="#sec_am"><em>Acoustic model</em></a>: Sphinx
uses acoustic models based on statistical <em>hidden Markov
models</em> (HMMs). The acoustic model is trained from acoustic
training data using the Sphinx-3 trainer. The trainer is
capable of building acoustic models with a wide range of
structures, such as <em>discrete</em>, <em>semi-continuous</em>,
or <em>continuous</em>. However, the s3.X decoder is only
capable of handling continuous acoustic models.
</LI>
<P></P>
<LI><a href="#sec_lm"><em>Language model (LM)</em></a>:
Sphinx-3 uses a conventional backoff bigram or trigram language
model.
</LI>
<P></P>
<LI><a href="#sec_ctl"><em>Speech input specification</em></a>:
This distribution contains four executable files, three of which
perform recognition. <code>sphinx3_livedecode</code> decodes live
speech, that is, speech incoming from your audio
card. <code>sphinx3_livepretend</code> decodes in batch mode using a
<em>control file</em> that describes the input to be decoded
into text. <code>sphinx3_decode</code> decodes also uses a control file
for batch mode processing. In the latter, the entire input to be
processed must be available beforehand, <em>i.e.</em>, the raw
audio samples must have been preprocessed into cepstrum
files. Also note that the decoder cannot handle arbitrary
lengths of speech input. Each separate piece (or
<em>utterance</em>) to be processed by the decoder must be no
more than 300 sec. long. Typically, one uses a
<em>segmenter</em> to chop up a cepstrum stream into manageable
segments of up to 20 or 30 sec. duration.
</LI>
</UL>
<H3><A NAME="sec_dec_output"><U>Outputs</U></A></H3>
<P>The decoder can produce two types of recognition output:
</P>
<UL>
<LI><a href="#sec_hypseg"><em>Recognition hypothesis</em></a>: A
single best recognition result (or <em>hypothesis</em>) for each
utterance processed. It is a linear word sequence, with
additional attributes such as their time segmentation and
scores.
</LI>
<P></P>
<LI><a href="#sec_wordlat"><em>Word lattice</em></a>: A
word-graph of all possible candidate words recognized during the
decoding of an utterance, including other attributes such as
their time segmentation and acoustic likelihood scores.
</LI>
</UL>
<P>In addition, the decoder also produces a detailed log to
stdout/stderr that can be useful in debugging, gathering
statistics, etc.
</P>
<div class="endsec">
¤
<a href="#sec_decoverview">Back to top of this section</a>
</div>
<H2><A NAME="sec_compile"><u>Compiling s3.X</u></A></H2>
<P>The current distribution has been set up for Unix platforms.
The following steps are needed to compile the decoder:
</P>
<P> For Users: </P>
<OL>
<LI><code>./configure [--prefix=/my/install/directory]</code>:
the argument is optional. If not given, it will install s3.X
under /usr/local, provided you have the proper permissions. This
step is only necessary the first time you compile s3.X.</LI>
<LI><code>make clean</code>: This should remove any old object
files.</LI>
<LI><code>make</code>: This compiles the libraries and example
programs.</LI>
<LI><code>make install</code>: This will install s3.X in the
directory that you specified when you ran <code>configure</code>
and also the provided models and documentation.</LI>
</OL>
<P>Note that the Makefiles are not foolproof; they do not eliminate
the need for sometimes manually determining dependencies,
especially upon updates to header files. When in doubt, first
clean out the compilation directories entirely by running
<code>make distclean</code> and start over.
<P> For Developers </P>
<OL>
<LI> The project could be boostrap from the SVN repository
which
could be obtained by command <code> svn co https://cmusphinx.svn.sourceforge.net/svnroot/cmusphinx/trunk/sphinx3 </code> </LI>
<LI> One could boostrap the installation by typing command <code> ./autogen.sh ; ./autogen.sh </LI>
</OL>
</P>
<div class="endsec">
¤
<a href="#sec_compile">Back to top of this section</a>
</div>
<H2><A NAME="sec_exec"><u>Running s3.X</u></A></H2>
<P>Running the decoder is simply a matter of invoking the binary
(<em>i.e.</em>, <code>sphinx3_decode</code>, <code>sphinx3_livedecode</code> or
<code>sphinx3_livepretend</code>), with a number of command line arguments
specifying the various input files, as well as decoding
configuration parameters. <code>sphinx3_decode</code> and
<code>sphinx3_livepretend</code> require a control file, the directory
where the audio files are available, and a file containing the
configuration arguments. <code>sphinx3_livedecode</code>, which runs live,
requires only the file with the arguments.
</p>
<P>Invoking the binary without any argument produces a <a
href="cmdhelp.txt">help message</a> with short descriptions of all
the configuration arguments.
</P>
<H3><A name="sec_args_overview"><U>Configuration Arguments Overview</U></a></H3>
<P>This section gives a brief overview of the main configuration
arguments. They are broken down into separate groups, based on
whether they are the primary flags specifying input and output
data, arguments for optional configuration, or for performance
tuning.
</P>
<P>Note that not all the available configuration arguments are
covered below. There are a few additional and undocumented flags,
intended mainly for debugging purposes.
</P>
<H4><A name="sec_flags_primary"><U>Primary Flags</U></a></H4>
<P>Many of the flags have reasonable defaults. The ones that a
user minimally needs to provide are the input and output databases
or files, which have been discussed <a
href="#sec_decoverview">above</a>:
</P>
<table cellpadding="8">
<tr>
<td><UL><LI><code>-mdef</code></LI></UL></td>
<td><a href="#sec_am">Model definition</a> input file</td>
</tr>
<tr>
<td>
<UL>
<LI><code>-mean</code></LI>
<LI><code>-var</code></LI>
<LI><code>-mixw</code></LI>
<LI><code>-tmat</code></LI>
<LI><code>-subvq</code></LI>
<LI><code>-hmm</code></LI>
</UL>
</td>
<td><a href="#am_files">Acoustic model</a> files. One could
conveniently specify the acoustic model by simply specifying
the -hmm option. The default model file names for the
components HMMs are <code> means </code>, <code> variances
</code>, <code> mixture_weights </code>, <code>
transition_matrices </code> and <code> mdef </code> (the model
definition).
</td>
</tr>
<tr>
<td>
<UL>
<LI><code>-dict</code></LI>
<LI><code>-fdict</code></LI>
</UL>
</td>
<td><a href="#dict_main_filler">Main and filler lexicons</a></td>
</tr>
<tr>
<td>
<UL>
<LI><code>-lm</code></LI>
<LI><code>-lmctlfn</code></LI>
<LI><code>-lminmemory</code></LI>
<LI><code>-lmname</code></LI>
</td>
<td>One could specify language model <a
href="#lm_dumpfile">binary dump file</a>
or txt file using <code> -lm </code>. Set of class-based LMs
could be specified with <code> -lmctlfn </code>. By default,
the lm is accessed mainly at harddisc with a cache
mechanism. One could toggle this behavior to fully memory-mode
by specifying option <code> -lminmemory </code> </td>
</tr>
<tr>
<td>
<UL>
<LI><code>-fillpen</code></LI>
<LI><code>-fillprob</code></LI>
<LI><code>-silprob</code></LI>
</UL>
</td>
<td><a href="#lm_filler">Filler word</a> probabilities</td>
</tr>
<tr>
<td><UL>
<LI><code>-hypseg</code></LI>
<LI><code>-hypsegfmt</code></LI>
<LI><code>-hypsegscore_unscale</code></LI>
</UL></td>
<td>Output <a href="#sec_hypseg">hypotheses file with detail scores and timing. <code> -hypsegfmt </code> could be used to specify format for sphinx 2, Sphinx 3 and NIST CTM format. </a></td>
</tr>
<tr>
<td><UL><LI><code>-hyp</code></LI></UL></td>
<td>Output <a href="#sec_hypseg">hypotheses file without detail scores and timing </a></td>
</tr>
</table>
<H4><A name="sec_flags_config"><U>Additional Configuration Flags</U></a></H4>
<P>It may often be necessary to provide additional parameters to
obtain the right decoder configuration:
</P>
<table cellpadding="8">
<tr>
<td>
<UL>
<LI><code>-cmn</code></LI>
<LI><code>-agc</code></LI>
<LI><code>-varnorm</code></LI>
<LI><code>-lowerf</code></LI>
<LI><code>-upperf</code></LI>
<LI><code>-nfilt</code></LI>
<LI><code>-samprate</code></LI>
</UL>
</td>
<td><a href="#am_feature">Feature type</a> configuration</td>
</tr>
<tr>
<td><UL><LI><code>-cepdir</code></LI></UL></td>
<td>Directory prefix for cepstrum files specified in the
<a href="#sec_ctl">control file</a>, ignored in <code>sphinx3_livedecode</code> and <code>sphinx3_livepretend</code></td>
</tr>
<tr>
<td>
<UL>
<LI><code>-ctl</code></LI>
<LI><code>-ctl_lm</code></LI>
<LI><code>-ctl_mllr</code></LI>
</UL>
</td>
<td>Specify the file, the LM and the MLLR used in each
utterances. </td>
</tr>
<tr>
<td>
<UL>
<LI><code>-mllr</code></LI>
<LI><code>-cb2mllr</code></LI>
</UL>
</td>
<td> The regression matrix and the senone to regression matrix
mapping </td>
</tr>
<tr>
<td>
<UL>
<LI><code>-bestpath</code></LI>
<LI><code>-bestpathlw</code></LI>
</UL>
</td>
<td> Applicable for mode 3 (flat-lexicon) and mode 4
(tree-lexicon) search. In the second pass, control the
best-path search. </td>
</tr>
<tr>
<td>
<UL>
<LI><code>-ctloffset</code></LI>
<LI><code>-ctlcount</code></LI>
</UL>
</td>
<td>Selecting a portion of the <a href="#sec_ctl">control
file</a> to be processed</td>
</tr>
<tr>
<td>
<UL>
<LI><code>-outlatdir</code></LI>
<LI><code>-latext</code></LI>
</UL>
</td>
<td>Directory, file-extension for <a href="#sec_wordlat">word
lattices</a> output</td>
</tr>
</table>
<H4><A name="sec_flags_tune"><u>Performance Tuning Flags</u></a></H4>
<P>In yet other cases, it may be necessary to tune the following
parameters to obtain the optimal computational efficiency or
recognition accuracy:
</P>
<table cellpadding="8">
<tr>
<td>
<UL>
<LI><code>-beam</code></LI>
<LI><code>-pbeam</code></LI>
<LI><code>-wbeam</code></LI>
<LI><code>-subvqbeam</code></LI>
</UL>
</td>
<td><a href="#sec_dec_prune">Beam pruning</a> parameters</td>
</tr>
<tr>
<td>
<UL>
<LI><code>-maxwpf</code></LI>
<LI><code>-maxhistpf</code></LI>
<LI><code>-maxhmmpf</code></LI>
</UL>
</td>
<td><a href="#sec_dec_prune">Absolute pruning</a> parameters</td>
</tr>
<tr>
<td>
<UL>
<LI><code>-ci_pbeam</code></LI>
<LI><code>-max_cdsenpf</code></LI>
<LI><code>-ds</code></LI>
</UL>
</td>
<td><a href="#sec_gmm_compute">Fast GMM Computation</a> parameters</td>
</tr>
<tr>
<td>
<UL>
<LI><code>-lw</code></LI>
<LI><code>-wip</code></LI>
</UL>
</td>
<td><a href="#lm_lw_wip">Language weight, word insertion penalty</a></td>
</tr>
<tr>
<td><UL><LI><code>-Nlextree</code></LI></UL></td>
<td>Number of lexical tree instances</td>
</tr>
</table>
<H3><A name="sec_dec_op"><U>Decoder Operation</U></a></H3>
<P>This section is a bit of a mish-mash; its contents probably
belong in an FAQ section. But, hopefully, through this section a
newcomer to Sphinx can get an idea of the structure, capabilities,
and limitations of the s3.X decoder.
</P>
<H4><A name="sec_dec_init"><U>Initialization</U></a></H4>
<P>The decoder is configured during the initialization step, and
the configuration holds for the entire run. This means, for
example, that the decoder does not dynamically reconfigure the
acoustic models to <em>adapt</em> to the input. To choose another
example, there is no mechanism in this decoder to switch language
models from utterance to utterance, unlike in Sphinx-II.
The main initialization steps are outlined below.
</P>
<P><b>Log-Base Initialization.</b> Sphinx performs all likelihood
computations in the log-domain. Furthermore, for computational
efficiency, the <em>base</em> of the logarithm is chosen such that
the likelihoods can be maintained as 32-bit integer values. Thus,
all the scores reported by the decoder are <em>log-likelihood</em>
values in this peculiar log-base. The default base is typically
1.0003, and can be changed using the <code>-logbase</code>
configuration argument. The main reason for modifying the
log-base would be to control the length (duration) of an input
utterance before the accumulated log-likelihood values overflow
the 32-bit representation, causing the decoder to fail
catastrophically. The log-base can be changed over a wide range
without affecting the recognition.
</P>
<P><b>Models Initialization.</b> The lexical, acoustic, and
language models specified via the configuration arguments are
loaded during initialization. This set of models is used to
decode all the utterances in the input. (The language model is
actually only partly loaded, since s3.X uses a <a
href="#lm_dumpfile">disk-based LM</a> strategy.)
</P>
<P><b>Effective Vocabulary.</b> After the models are loaded,
the <em>effective vocabulary</em> is determined. It is the set of
words that the decoder is capable of recognizing. Recall that the
decoder is initialized with three sources of words: the <a
href="#dict_main_filler">main and filler lexicon</a> files, and
the <a href="#sec_lm">language model</a>. The effective
vocabulary is determined from them as follows:
</P>
<OL>
<LI>Find the intersection of the words in the LM and the main
pronunciation lexicon</LI>
<LI>Include all the alternative pronunciations to the set
derived above (using the main lexicon)</LI>
<LI>Include all the filler words from the filler lexicon, but
excluding the distinguished beginning and end of sentence words:
<code><s></code> and <code></s></code>.</LI>
</OL>
<P>The effective vocabulary remains in effect throughout the batch
run. It is not possible to add to or remove from this vocabulary
dynamically, unlike in the Sphinx-II system.
</P>
<P><b>Lexical Tree Construction.</b> The decoder constructs
<em>lexical trees</em> from the effective vocabulary described
above. Separate trees are constructed for words in the <a
href="#dict_main_filler">main and filler lexicons</a>.
Furthermore, several copies may be instantiated for the two,
depending on the <code>-Nlextree</code> configuration argument.
Further details of the lexical tree construction are available on
the PowerPoint <a href="s3-2.ppt">slides</a>.
</P>
<H4><A name="sec_ctl_process"><U>Control File Processing</U></a></H4>
<P>Following initialization, <code>sphinx3_decode</code> and
<code>sphinx3_livepretend</code> processes the entries in the control file
sequentially, one at a time. It is possible to process a
contiguous subset of the control file, using the
<code>-ctloffset</code> and <code>-ctlcount</code> flags, as
mentioned earlier. There is no learning or <em>adaptation</em>
capability as decoding progresses. Since <code>sphinx3_livepretend</code>
behaves as if the files were being spoken at the time of
processing, rearranging the order of the entries in the control
file may affect the individual results, but this change may be
imperceptible if the environment in which the files were recorded
remains constant. The order of entries in the control file does
not affect <code>sphinx3_decode</code>.
</P>
<H4><A name="sec_dec_prune"><U>Pruning</U></a></H4>
<P>Each entry in the control file, or utterance, is processed
using the given input models, and using the <em>Viterbi search
algorithm</em>. In order to constrain the active search space to
computationally manageable limits, <em>pruning</em> is employed,
which means that the less promising hypotheses are continually
discarded during the recognition process. There are two kinds of
pruning in s3.X, <em>beam pruning</em> and <em>absolute
pruning</em>.
</P>
<P><b>Beam Pruning.</b> Each utterance is processed in a
<em>time-synchronous</em> manner, one frame at a time. At each
frame the decoder has a number of currently <em>active</em> HMMs
to match with the next frame of input speech. But it first
discards or deactivates those whose state likelihoods are below
some <em>threshold</em>, relative to the best HMM state likelihood at
that time. The threshold value is obtained by
<em>multiplying</em> the best state likelihood by a fixed
<em>beamwidth</em>. The beamwidth is a value between 0 and 1, the
former permitting all HMMs to survive, and the latter permitting
only the best scoring HMMs to survive.
</P>
<P>Similar beam pruning is also used in a number of other
situations in the decoder, e.g., to determine the candidate words
recognized at any time, or to determine the component densities in
a mixture Gaussian that are closest to a given speech feature
vector. The various beamwidths have to be determined empirically
and are set using <a href="#sec_flags_tune">configuration
arguments</a>.
</P>
<P><b>Absolute Pruning.</b> Even with beam pruning, the number of
active entities can sometimes become computationally overwhelming.
If there are a large number of HMMs that fall within the pruning
threshold, the decoder will keep all of them active. However,
when the number of active HMMs grows beyond certain limits, the
chances of detecting the correct word among the many candidates
are considerably reduced. Such situations can occur, for example,
if the input speech is noisy or quite mismatched to the acoustic
models. In such cases, there is no point in allowing the active
search space to grow to arbitrary extents. It can be contained
using pruning parameters that limit the <em>absolute number</em>
of active entities at any instant. These parameters are also
determined empirically, and set using <a
href="#sec_flags_tune">configuration arguments</a>.
</P>
<H4><A name="sec_gmm_compute"><U>(After s3.4) Fast GMM
Computation</U></a></H4>
<p> The computation of likelihood Gaussian distribution can be one
of the dominating factor of the GMM computation. Tuning the
following parameters can control the amount of time required. </p>
<ul>
<LI><code>-ci_pbeam</code>: Enable a two-pass computation where
CI models were computed first and the CD models were then
computed. If this beam is used, only CD models, which correspond
to CI models within the beam (relative to the max CI scores), are
computed </LI>
<LI><code>-maxcdsenpf</code>: Similar to <code> -ci_pbeam
</code> but the beam is decided by an absolute number of senones
computd. </LI>
<LI><code>-ds </code> : Enable frame down-sampling. Only 1
another N frames were computed. </LI>
</ul>
<H4><U>Output Generation</U></H4>
<P>During recognition, the decoder builds an internal
<em>backpointer table</em> data structure, from which the final
outputs are generated. This table records all the candidate words
recognized during decoding, and their attributes such as their
time segmentation, acoustic and LM likelihoods, as well as their
predecessor entries in the table. When an utterance has been
fully processed, the best <a href="#sec_hypseg">recognition
hypothesis</a> is extracted from this table. Optionally, the
table is also converted into a <a
href="#sec_wordlat">word-lattice</a> and written out to a file.
</P>
<P>More information on the backpointer table is available in the
PowerPoint <a href="s3-2.ppt">slides</a>.
</P>
<H4><U>Miscellaneous Issues</U></H4>
<P><b>Role of <code><s></code> and
<code></s></code>.</b> The distinguished
<em>beginning-of-sentence</em> and <em>end-of-sentence</em> tokens
<code><s></code> and <code></s></code> are not in the
effective vocabulary, and no part of the input speech is decoded
into either of them. They are merely anchors at the ends of each
utterance, and provide context for the LM. This is in contrast to
earlier versions of Sphinx, which required some silence at either
end of each speech utterance, to be decoded into these tokens.
</P>
<H3><A name="sec_dec_tune"><U>Performance Tuning</U></a></H3>
<P>To obtain the best recognition performance, it is necessary to
select the appropriate front-end and feature type computation,
train the various models, as well as tune the decoder
configuration parameters. This section deals with the last issue.
There are mainly two groups of parameters to be tuned, pertaining
to <a href="#sec_tune_prune">pruning</a> and <a
href="#sec_tune_lw">LM</a>. Unfortunately, there are no automatic
methods for determining the values of these parameters; it is
necessary to derive them by trial and error. Additionally, the
following points should be kept in mind with regard to the pruning
parameters:
</P>
<UL>
<LI>The pruning parameters need to be tuned whenever the
acoustic model is changed.</LI> <P></P>
<LI>Changing the LM related parameters affects the effective
pruning behaviour. The pruning parameters ought to be re-tuned
after the former have been tuned (although this step is often
skipped in practice).</LI> <P></P>
<LI>For computational efficiency, the beamwidth parameter values
should be as narrow as possible (values closer to 1.0 are
narrower), and the absolute pruning parameter values should be
as small as possible.</LI> <P></P>
<LI>But, for recognition accuracy, the pruning parameters should
be as relaxed as possible. (However, relaxing the beamwidth
parameters too much can actually <em>worsen</em> recognition
accuracy. The reasons for such perverse behaviour are not quite
understood.)</LI>
</UL>
<H4><A name="sec_tune_prune"><U>Tuning the Pruning Behaviour</U></a></H4>
<P>The pruning parameters are the following:</P>
<UL>
<LI><code>-beam</code>: Determines which HMMs remain active at
any given point (frame) during recognition. (Based on the best
state score within each HMM.)
</LI>
<LI><code>-pbeam</code>: Determines which active HMM can
transition to its successor in the lexical tree at any point.
(Based on the exit state score of the source HMM.)
</LI>
<LI><code>-wbeam</code>: Determines which words are recognized
at any frame during decoding. (Based on the exit state scores
of leaf HMMs in the lexical trees.)</LI>
<LI><code>-maxhmmpf</code>: Determines the number of HMMs
(approx.) that can remain active at any frame.</LI>
<LI><code>-maxwpf</code>: Controls the number of distinct words
recognized at any given frame.</LI>
<LI><code>-maxhistpf</code>: Controls the number of distinct
word histories recorded in the backpointer table at any given
frame.</LI>
<LI><code>-subvqbeam</code>: For each <a
href="#sec_am">senone</a> and its underlying acoustic model,
determines its active mixture components at any frame.</LI>
</UL>
<P>In order to determine the pruning parameter values empirically,
it is first necessary to obtain a <em>test set</em>,
<em>i.e.</em>, a collection of test sentences not used in any
training data. The test set should be sufficiently large to
ensure statistically reliable results. For example, a
large-vocabulary task might require a test set that includes a
half-hour of speech, or more.
</P>
<P>It is difficult to tune a handful of parameters simultaneously,
especially when the input models are completely new. The
following steps may be followed to deal with this complex problem.
</P>
<OL>
<LI>To begin with, set the absolute pruning parameters to large
values, making them essentially ineffective. Set both
<code>-beam</code> and <code>-pbeam</code> to
<code>1e-60</code>, and <code>-wbeam</code> to
<code>1e-30</code>. Set <code>-subvqbeam</code> to a small
value (e.g., the same as <code>-beam</code>). Run the decoder
on the chosen test set and obtain accuracy results. (Use
default values for the <a href="#sec_tune_lw">LM related
parameters</a> when tuning the pruning parameters for the first
time.)
</LI>
<P></P>
<LI>Repeat the decoder runs, varying <code>-beam</code> up and
down, until the setting for best accuracy is identified. (Keep
<code>-pbeam</code> the same as <code>-beam</code> every time.)
</LI>
<P></P>
<LI>Now vary <code>-wbeam</code> up and down and identify its
best possible setting (keeping <code>-beam</code> and
<code>-pbeam</code> fixed at their most recently obtained
value).
</LI>
<P></P>
<LI>Repeat the above two steps, alternately optimizing
<code>-beam</code> and <code>-wbeam</code>, until convergence.
Note that during these iterations <code>-pbeam</code> should
always be the same as <code>-beam</code>. (This step can be
omitted if the accuracy attained after the first iteration is
acceptable.)
</LI>
<P></P>
<LI>Gradually increase <code>-subvqbeam</code> (<em>i.e.</em>,
towards 1.0 for a narrower setting), stopping when recognition
accuracy begins to drop noticeably. Values near the default are
reasonable. (This step is needed only if a <a
href="#am_subvq">sub-vector quantized</a> model is available for
speeding up acoustic model evaluation.)
</LI>
<P></P>
<LI>Now gradually increase <code>-pbeam</code> (<em>i.e.</em>,
towards 1.0), stopping when recognition accuracy begins to drop
noticeably. (This step is optional; it mainly optimizes the
computational effort a little more.)
</LI>
<P></P>
<LI>Reduce <code>-maxhmmpf</code> gradually until accuracy
begins to be affected. Repeat the process with
<code>-maxwpf</code>, and then with <code>-maxhistpf</code>.
(However, in some situations, especially when the vocabulary
size is small, it may not be necessary to tune these absolute
pruning parameters.)
</LI>
</OL>
<P>In practice, it may not always be possible to follow the above
steps strictly. For example, considerations of computational cost
might dictate that the absolute pruning parameters or the
<code>-subvqbeam</code> parameter be tuned earlier in the
sequence.
</P>
<H4><A name="sec_tune_lw"><U>Tuning Language Model Related
Parameters</U></a></H4>
<P>The parameters needed to be tuned are the following:</P>
<UL>
<LI><code>-lw</code>: The <a href="#lm_lw_wip">language weight</a>.
<LI><code>-wip</code>: The <a href="#lm_lw_wip">word insertion
penalty</a>.
</LI>
</UL>
<P>Like the pruning parameters, the above two are tuned on a test
set. Since the decoder is much more sensitive to the language
weight, that is typically tuned first, using the default word
insertion penalty. The latter is then tuned. It is usually not
necessary to repeat the process.
</P>
<H3><A name="sec_dec_errors"><U>Some Common Errors and Failure
Modes</U></a></H3>
<P>To be completed.</P>
<div class="endsec">
¤
<a href="#sec_exec">Back to top of this section</a>
</div>
<H2><a name="sec_dict"><U>Pronunciation Lexicon</U></a></H2>
<div class="wheatbox">
<UL>
<LI><A HREF="#dict_struct">Lexicon Structure</A>
<UL>
<LI><A HREF="#dict_multipron">Multiple Pronunciations</a></LI>
<LI><A HREF="#dict_compwd">Compound Words</a></LI>
</UL>
</LI>
<LI><A HREF="#dict_main_filler">Main and Filler Lexicons</A></LI>
</UL>
</div>
<H3><a name="dict_struct"><U>Lexicon Structure</U></a></H3>
<P>A pronunciation lexicon (or dictionary) file specifies word
pronunciations. In Sphinx, pronunciations are specified as a
linear sequence of <em>phonemes</em>. Each line in the file
contains one pronunciation specification, except that any line
that begins with a "#" character <u>in the first column</u> is
treated as a comment and is ignored. Example dictionary for
digits:
</P>
<pre>
ZERO Z IH R OW
ONE W AH N
TWO T UW
THREE TH R IY
FOUR F AO R
FIVE F AY V
SIX S IH K S
SEVEN S EH V AX N
EIGHT EY TD
NINE N AY N</pre>
<P>The lexicon is completely <em>case-insensitive</em>
(unfortunately). For example, it's not possible to have two
different entries <code>Brown</code> and <code>brown</code> in the
dictionary.
</P>
<H4><a name="dict_multipron"><U>Multiple Pronunciations</U></a></H4>
<P>A word may have more than one pronunciation, each one on a
separate line. They are distinguished by a unique parenthesized
suffix for the word string. For example:
</P>
<pre>
ACTUALLY AE K CH AX W AX L IY
ACTUALLY(2nd) AE K SH AX L IY
ACTUALLY(3rd) AE K SH L IY</pre>
<P>If a word has more than one pronunciation, its first appearance
must be the unparenthesized form. For the rest, the parenthesized
suffix may be any string, as long as it is unique for that word.
There is no other significance to the order of the alternatives;
each one is considered to be equally likely.
</P>
<H4><a name="dict_compwd"><U>Compound Words</U></a></H4>
<P>In Sphinx-3, the lexicon may also contain <em>compound
words</em>. A compound word is usually a short phrase whose
pronunciation happens to differ significantly from the mere
concatenation of the pronunciations of its constituent words.
Compound word tokens are formed by concatenating the component
word strings with an underscore character; e.g.:
</P>
<pre>WANT_TO W AA N AX</pre>
<P>(The s3.X decoder, however, treats a compound word as just
another word in the language, and does not do anything special
with it.)
</P>
<H3><a name="dict_main_filler"><U>Main and Filler Lexicons</U></a></H3>
<P>The Sphinx-3 decoders actually need two separate lexicons: a
"regular" one containing the words in the language of interest,
and also a <em>filler</em> or <em>noise</em> lexicon. The latter
defines "words" not in the language. More specifically, it
defines legal "words" that do not appear in the language model
used by the decoder, but are nevertheless encountered in normal
speech. This lexicon must include the <em>silence word</em>
<code><sil></code>, as well as the special
<em>beginning-of-sentence</em> and <em>end-of-sentence</em> tokens
<code><s></code>, and <code></s></code>, respectively.
All of them usually have the silence-phone <code>SIL</code> as
their pronunciation. In addition, this lexicon may also contain
"pronunciations" for other noise event words such as breath noise,
"UM" and "UH" sounds made during spontaneous speech, etc.
</P>
<div class="endsec">
¤
<a href="#sec_dict">Back to top of this section</a>
</div>
<H2><a name="sec_am"><U>Acoustic Model</U></a></H2>
<div class="wheatbox">
<UL>
<LI><A HREF="#am_intro">Introduction</A></LI>
<LI><A HREF="#am_feature">Acoustic Features Computation</A></LI>
<LI><A HREF="#am_training">Acoustic Model Training</A></LI>
<LI><A HREF="#am_struct">Model Structures</A></LI>
<LI><A HREF="#am_subvq">Sub-Vector Quantized Models</A></LI>
<LI><A HREF="#am_files">Model Files</A></LI>
</UL>
</div>
<H3><a name="am_intro"><U>Introduction</U></a></H3>
<P>Sphinx-3 is based on <em>subphonetic acoustic models</em>.
First, the basic sounds in the language are classified into
phonemes or <em>phones</em>. There are roughly 50 phones in the
English language. For example, here is a pronunciation for the
word <code>LANDSAT</code>:
</P>
<pre>L AE N D S AE TD</pre>
<P>Phones are then further refined into context-dependent
<em>triphones</em>, <em>i.e.</em>, phones occurring in given left
and right phonetic contexts. The reason is that the same phone
within different contexts can have widely different acoustic
manifestations, requiring separate acoustic models. For example,
the two occurrences of the <code>AE</code> phone above have
different contexts, only the first of which is nasal.
</P>
<P>In contrast to triphones, a phone considered without any
specific context is referred to as a <em>context-independent</em>
phone or <em>basephone</em>. Note also that context-dependency
gives rise to the notion of <em>cross-word</em> triphones. That
is, the left context for the leftmost basephone of a word depends
on what was the previous word spoken.
</P>
<P>Phones are also distinguished according to their position
within the word: beginning, end, internal, or single (abbreviated
<code>b</code>, <code>e</code>, <code>i</code> and <code>s</code>,
respectively). For example, in the word <code>MINIMUM</code> with
the following pronunciation:
</P>
<pre>M IH N AX M AX M</pre>
<P>the three occurrences of the phone <code>M</code> have three
different position attributes. The <code>s</code> attribute
applies if a word has just a single phone as its pronunciation.
</P>
<P>For most applications, one builds acoustic models for
triphones, qualified by the four position attributes. (This
provides far greater modelling detail and accuracy than if one
relies on just basephone models.) Each triphone is modelled by a
<em>hidden Markov model</em> or <em>HMM</em>. Typically, 3 or 5
state HMMs are used, where each state has a statistical model for
its underlying acoustics. But if we have 50 basephones, with 4
position qualifiers and 3-state HMMs, we end up with a total of
50<sup>3</sup>*4*3 distinct HMM states! Such a model set would be
too large and impractical to train. To keep things manageable,
HMM states are <em>clustered</em> into a much smaller number of
groups. Each such group is called a <em>senone</em> (in Sphinx
terminology), and all the states mapped into one senone share the
same underlying statistical model. (The clustering of HMM states
into senones is described in Mei-Yuh Hwang's PhD Thesis.)
</P>
<P>Each triphone also has a <em>state transition probability
matrix</em> that defines the topology of its HMM. Once again, to
conserve resources, there is a considerable amount of sharing.
Typically, there is one such matrix per basephone, and all
triphones derived from the same parent basephone share its state
transition matrix.
</P>
<P>The information regarding triphones and mapping from triphone
states to senones and transition matrices is captured in a
<em>model definition</em>, or <em>mdef</em> input file.
</P>
<H3><a name="am_feature"><U>Acoustic Features Computation</U></a></H3>
<P>For various reasons, it is undesirable to build acoustic models
directly in terms of the raw audio samples. Instead, the audio is
processed to extract a vector of relevant features. All acoustic
modelling is carried out in terms of such feature vectors.
</P>
<P>In Sphinx, feature vector computation is a two-stage process.
An off-line <a href="./s3_fe_spec.pdf"><em>front-end</em></a>
module is first responsible for processing the raw audio sample
stream into a <em>cepstral</em> stream, which can then be input to
the Sphinx software. The input audio stream consists of 16-bit
samples, at a sampling rate of 8 or 16 KHz depending on whether
the input is narrow or wide-band speech. The input is windowed,
resulting in <em>frames</em> of duration 25.625 ms. The number of
samples in a frame depends on the sampling rate. The output is a
stream of 13-dimensional real-valued <em>cepstrum
vectors</em>. The frames overlap, thus resulting in a rate of 100
vectors/sec.
</P>
<P>In the second stage, the Sphinx software (both trainer and
decoder) internally converts the stream of cepstrum vectors into a
<em>feature stream</em>. This process consists of the following
steps:
</P>
<OL>
<LI>An optional <em>cepstrum mean-normalization</em> (CMN) step,
which itself includes an optional <em>variance
normalization</em> (VN) step.
</LI>
<P></P>
<LI>An optional <em>automatic gain control</em> (AGC) step, in
which the signal power component of the cepstral vectors is
normalized.
</LI>
<P></P>
<LI><em>Feature vector generation</em>. The final speech
feature vector is created by typically augmenting the cepstrum
vector (after CMN and AGC, if any) with one or more time
derivatives. In s3.X, the feature vector in each frame is
computed by concatenating first and second derivatives to the
cepstrum vector, giving a 39-dimensional vector:
<div class="silverbox">
<center><IMG ALT="* " SRC="./s3/feat.gif"></center>
</div>
</LI>
</OL>
<H3><a name="am_training"><U>Acoustic Model Training</U></a></H3>
<P>This refers to the computation of a (statistical) model for
each senone in the model. As a <u>very rough approximation</u>,
this process can be described by the following <u>conceptual</u>
steps:
<P>
<OL>
<LI>Obtain a corpus of training data. This may include
thousands of sentences (or <em>utterances</em>, in Sphinx
jargon), consisting of the spoken text and corresponding audio
sample stream.
</LI>
<P></P>
<LI>For each utterance, convert the audio data to a stream of
feature vectors as described above.
</LI>
<P></P>
<LI>For each utterance, convert the text into a linear sequence
of triphone HMMs using the <a href="#sec_dict">pronunciation
lexicon</a>. (This is usually called the <em>sentence
HMM</em>.)
</LI>
<P></P>
<LI>For each utterance, find the best <em>state sequence</em> or
<em>state alignment</em> through the sentence HMM, for the
corresponding feature vector sequence. For example, the figure
below shows a single HMM with 3 states (using senones 0, 1, 2),
and an utterance of 14-frames of feature vectors. The figure
also shows a sample HMM-state (senone) sequence: each feature
frame is labelled with a senone ID.
<div class="silverbox">
<center><IMG ALT="* " SRC="./s3/falign.gif"></center>
</div>
The best state sequence is one with the <em>smallest
mismatch</em> between the input feature vectors and the labelled
senones' underlying statistical models.
</LI>
<P></P>
<LI>For each senone, gather all the frames in the training
corpus that mapped to that senone in the above step, and build a
suitable statistical model for the corresponding collection of
feature vectors.
</LI>
</OL>
<P>Note that there is a circularity in the above description. We
wish to train the senone models, but in the penultimate step, we
need the senone models to compute the best possible state
alignment. This circularity is resolved by using the iterative
<em>Baum-Welch</em> or <em>forward-backward</em> training
algorithm. The algorithm begins with some initial set of models,
which could be completely flat, for the senones. It then repeats
the last two steps several times. Each iteration uses the model
computed at the end of the previous iteration.
</P>
<P>Although not mentioned above, the HMM state-transition
probability matrices are also trained from the state alignments.
Acoustic modelling is described in greater detail in the Sphinx-3
trainer module.
</P>
<H3><a name="am_struct"><U>Model Structures</U></a></H3>
<P>The acoustic models trained as described above can be of
different degrees of sophistication. Two forms are commonly used:
</P>
<UL>
<LI><em>continuous</em>, and
<LI><em>semi-continuous</em> or <em>tied-mixture</em>.
</UL>
<P>In a continuous model, each senone has its own, private
<em>mixture-Gaussian</em> distribution that describes the
statistics of its underlying speech feature space. In a
semi-continuous model, all the senones share a single
<em>codebook</em> of Gaussian distributions, but each senone has
its own set of <em>mixture weights</em> applied to the codebook
components. Sphinx-3 supports both models, and other,
intermediate degrees of state-tying as well. (The s3.X decoder,
however, can only handle continuous density acoustic models.)
</P>
<P>Similarly, Sphinx-3 in general supports "arbitrary" HMM
topologies, unlike Sphinx-II, which is restricted to a specific
5-state topology. However, for efficiency's sake, the s3.X
decoder is hardwired to deal with only two types of HMM
topologies: 3-state and 5-state, described briefly in <a
href="../src/libs3decoder/hmm.h">hmm.h</a>.
</P>
<H3><a name="am_subvq"><U>Sub-Vector Quantized Models</U></a></H3>
<P>Continuous density acoustic models are computationally
expensive to deal with, since they can contain hundreds of
thousands of Gaussian densities that must be evaluated in each
frame. To reduce this cost, one can use an approximate model that
efficiently identifies the top scoring candidate densities in each
Gaussian mixture in any given frame. The remaining densities can
be ignored during that frame.
</P>
<P>In Sphinx-3, such an approximate model is built by
<em>sub-vector quantizing</em> the acoustic model densities. The
utility that performs this conversion is included in this
distribution and is called <code>gausubvq</code>, which stands for
Gaussian Sub-Vector Quantization.
</P>
<P>Note that if the original model consists of mixture Gaussians
that only contain a few component densities (say, 4 or fewer per
mixture), a sub-vector quantized model may not be effective in
reducing the computational load.
</P>
<H3><a name="am_files"><U>Model Files</U></a></H3>
<P>An acoustic model is represented by the following collection of
files:
</P>
<UL>
<LI>A <em>model definition</em> (or <em>mdef</em>) file. It
defines the set of basephone and triphone HMMs, the mapping of
each HMM state to a senone, and the mapping of each HMM to a
state transition matrix.
</LI>
<LI>Gaussian <em>mean</em> and <em>variance</em> (or
<em>mean</em> and <em>var</em>) files. These files contain all
the Gaussian codebooks in the model. The Gaussian means and
corresponding variance vectors are separated into the two files.
</LI>
<LI>A <em>mixture weights</em> (or <em>mixw</em>) file
containing the Gaussian mixture weights for all the senones in
the model.
</LI>
<LI>A <em>state transition matrix</em> (or <em>tmat</em>) file
containing all the HMM state transition topologies and their
transition probabilities in the model.
</LI>
<LI>An optional <em>sub-vector quantized model</em> (or
<em>subvq</em>) file containing an approximation of the acoustic
model, for efficient evaluation.
</LI>
</UL>
<P>The <em>mean</em>, <em>var</em>, <em>mixw</em>, and
<em>tmat</em> files are produced by the Sphinx-3 trainer, and
their file formats should be documented there.
</P>
<div class="endsec">
¤
<a href="#sec_am">Back to top of this section</a>
</div>
<H2><a name="sec_lm"><U>Language Model</U></a></H2>
<div class="wheatbox">
<UL>
<LI><A HREF="#lm_intro">Introduction</a></LI>
<LI><A HREF="#lm_ngrams">Unigrams, Bigrams, Trigrams, LM
Vocabulary</a></LI>
<LI><A HREF="#lm_pron_case">Pronunciation and Case
Considerations</a></LI>
<LI><A HREF="#lm_dumpfile">Binary LM File</A></LI>
<LI><A HREF="#lm_filler">Silence and Filler Words</a></LI>
<LI><A HREF="#lm_lw_wip">Language Weight and Insertion
Penalty</A></LI>
</UL>
</div>
<H3><a name="lm_intro"><U>Introduction</U></a></H3>
<P>The main language model (LM) used by the Sphinx decoder is a
conventional bigram or trigram backoff language model. The <a
href="http://www.speech.cs.cmu.edu/SLM_info.html"><em>CMU-Cambridge
SLM toolkit</em></a> is capable of generating such a model from LM
training data. Its output is an ascii text file. But a large
text LM file can be very slow to load into memory. To speed up
this process, the LM must be compiled into a <a
href="#lm_dumpfile">binary form</a>. The code to convert from an
ascii text file to the binary format is available at <a
href="http://www.sourceforge.net/projects/cmusphinx">SourceForge</a>
in the CVS tree, in a module named <em>share</em>.
</P>
<H3><a name="lm_ngrams"><U>Unigrams, Bigrams, Trigrams, LM
Vocabulary</U></a></H3>
<P>A trigram LM primarily consists of the following:</P>
<UL>
<LI><em>Unigrams:</em> The entire set of words in this LM, and
their individual probabilities of occurrence in the language.
The unigrams must include the special
<em>beginning-of-sentence</em> and <em>end-of-sentence</em>
tokens: <code><s></code>, and <code></s></code>
respectively.</LI> <P></P>
<LI><em>Bigrams:</em> A <em>bigram</em> is mathematically
<em>P(word2 | word1)</em>. That is, the <em>conditional
probability</em> that <em>word2</em> immediately follows
<em>word1</em> in the language. An LM typically contains this
information for some subset of the possible word pairs. That
is, not all possible <em>word1 word2</em> pairs need be covered
by the bigrams.</LI> <P></P>
<LI><em>Trigrams:</em> Similar to a bigram, a <em>trigram</em>
is <em>P(word3 | word1, word2)</em>, or the conditional
probability that <em>word3</em> immediately follows a <em>word1
word2</em> sequence in the language. Not all possible 3-word
combinations need be covered by the trigrams.</LI> <P></P>
</UL>
<P>The <em>vocabulary</em> of the LM is the set of words covered
by the unigrams.</P>
<P>The LM probability of an entire sentence is the product of the
individual word probabilities. For example, the LM probability of
the sentence <code>"HOW ARE YOU"</code> is:</P>
<pre>
P(HOW | <s>) *
P(ARE | <s>, HOW) *
P(YOU | HOW, ARE) *
P(</s> | ARE, YOU)</pre>
<H3><a name="lm_pron_case"><U>Pronunciation and Case
Considerations</U></a></H3>
<P>In Sphinx, the LM cannot distinguish between different
pronunciations of the same word. For example, even though the
lexicon might contain two different pronunciation entries for the
word <code>READ</code> (present and past tense forms), the
language model cannot distinguish between the two. Both
pronunciations would inherit the same probability from the
language model.
</P>
<P>Secondly, the LM is <em>case-insensitive</em>. For example, it
cannot contain two different tokens <code>READ</code> and
<code>read</code>.
</P>
<P>The reasons for the above restrictions are historical. Precise
pronunciation and case information has rarely been present in LM
training data. It would certainly be desirable to do away with
the restrictions at some time in the future.
</P>
<H3><a name="lm_dumpfile"><U>Binary LM File</U></a></H3>
<P>The binary LM file (also referred to as the LM <em>dump</em>
file) is more or less a disk image of the LM data structure
constructed in memory. This data structure was originally
designed during the Sphinx-II days, when efficient memory usage
was the focus. In Sphinx-3, however, memory usage is no longer an
issue since the binary file enables the decoder to use a
<em>disk-based LM</em> strategy. That is, the LM binary file is
no longer read entirely into memory. Rather, the portions
required during decoding are read in on demand, and cached. For
large vocabulary recognition, the memory resident portion is
typically about 10-20% of the bigrams, and 5-10% of the trigrams.
</P>
<P>Since the decoder uses a <a href="#lm_dumpfile">disk-based
LM</a>, it is necessary to have efficient access to the binary LM
file. Thus, network access to an LM file at a remote location is
not recommended. It is desirable to have the LM file be resident
on the local machine.
</P>
<P>The binary dump file can be created from the ascii form using
the <code>lm3g2dmp</code> utility, which is part of the Sphinx-II
distribution, and also available as standalone code, as mentioned
before. (The header of the dump file itself contains a brief
description of the file format.)
</P>
<H3><a name="lm_filler"><U>Silence and Filler Words</U></a></H3>
<P>Language models typically do not cover acoustically significant
events such as silence, breath-noise, <em>UM</em> or <em>UH</em>
sounds made by a person hunting for the right phrase, etc. These
are known generally as <em>filler words</em>, and are excluded
from the LM vocabulary. The reason is that a language model
training corpus, which is simply a lot of text, usually does not
include such information.
</P>
<P>Since the main trigram LM ignores silence and filler words,
their "language model probability" has to be specified in a
separate file, called the <em>filler penalty file</em>. The
format of this file is very straightforward; each line contains
one word and its probability, as in the following example:
</P>
<pre>
++UH++ 0.10792
++UM++ 0.00866
++BREATH++ 0.00147</pre>
<P>The filler penalty file is not required. If it <em>is</em>
present, it does not have to contain entries for every filler
word. The decoder allows a default value to be specified for
filler word probabilities (through the <code>-fillprob</code>
configuration argument), and a default silence word probability
(through the <code>-silprob</code> argument).
</P>
<P>Like the main trigram LM, filler and silence word probabilities
are obtained from appropriate training data. However, training
them is considerably easier since they are merely unigram
probabilities.
</P>
<P>Filler words are invisible or <em>transparent</em> to the
trigram language model. For example, the LM probability of the
sentence <code>"HAVE CAR <sil> WILL TRAVEL"</code> is:</P>
<pre>
P(HAVE | <s>) *
P(CAR | <s>, HAVE) *
P(<sil>) *
P(WILL | HAVE, CAR) *
P(TRAVEL | CAR, WILL) *
P(</s> | WILL, TRAVEL)</pre>
<H3><a name="lm_lw_wip"><U>Language Weight and Word Insertion
Penalty</U></a></H3>
<P>During recognition the decoder combines both acoustic
likelihoods and language model probabilities into a single score
in order to compare various hypotheses. This combination of the
two is not just a straightforward product. In order to obtain
optimal recognition accuracy, it is usually necessary to
<em>exponentiate</em> the language model probability using a
<em>language weight</em> before combining the result with the
acoustic likelihood. (Since likelihood computations are actually
carried out in the log-domain in the Sphinx decoder, the LM weight
becomes a multiplicative factor applied to LM log-probabilities.)
</P>
<P>The language weight parameter is typically obtained through
trial and error. In the case of Sphinx, the optimum value for
this parameter has usually ranged between 6 and 13, depending on
the task at hand.
</P>
<P>Similarly, though with lesser impact, it has also been found
useful to include a <em>word insertion penalty</em> parameter
which is a fixed penalty for each new word hypothesized by the
decoder. It is effectively another multiplicative factor in the
language model probability computation (before the application of
the language weight). This parameter has usually ranged between
0.2 and 0.7, depending on the task.
</P>
<div class="endsec">
¤
<a href="#sec_lm">Back to top of this section</a>
</div>
<H2><a name="sec_ctl"><U>Speech Input Control File</U></a></H2>
<P>The Sphinx-3 decoder processes entries listed in a <em>control
file</em>. Each line in the control file identifies a separate
<em>utterance</em>. A line has the following format (the brackets
indicate a group of fields that is optional):
</P>
<pre>
AudioFile [ StartFrame EndFrame UttID ]</pre>
<P><em>AudioFile</em> is the speech input file. In this
distribution of s3.X, this file is in raw audio format. In all
other versions of Sphinx-3, this file contains cepstrum data. The
filename extension should be omitted from the specification. If
this is the only field in the line, the entire file is processed
as one utterance. In this case, an <em>utterance ID</em> string
is automatically derived from the cepstrum filename, by stripping
any leading directory name components from it. E.g.: if the
control file contains the following entries:
</P>
<pre>
/net/alf20/usr/rkm/SHARED/cep/nov94/h1_et_94/4t0/4t0c0201
/net/alf20/usr/rkm/SHARED/cep/nov94/h1_et_94/4t0/4t0c0202
/net/alf20/usr/rkm/SHARED/cep/nov94/h1_et_94/4t0/4t0c0203</pre>
<P>three utterances are processed, with IDs <code>4t0c0201</code>,
<code>4t0c0202</code>, and <code>4t0c0203</code>, respectively.
</P>
<P>If, on the other hand, a control file entry includes the
<em>StartFrame</em> and <em>EndFrame</em> fields, only that
portion of the cepstrum file is processed. This form of the
control file is frequently used if the speech input can be
arbitrarily long, such as an entire TV news show. There is one
big cepstrum file, but it is processed in smaller chunks or
segments. In this case, the final <em>UttID</em> field is the
utterance ID string for the entry.
</P>
<P>The utterance ID associated with a control file entry is used
to identify all the output from the decoder for that utterance.
For example, if the decoder is used to generate <a
href="#sec_wordlat">word lattice</a> files, they are named using
the utterance ID. Hence, each ID, whether automatically derived
or explicitly specified, should be unique over the entire control
file.
</P>
<P>Any line in the control file beginning with a <code>#</code>
character is a comment line, and is ignored.
</P>
<div class="endsec">
¤
<a href="#sec_ctl">Back to top of this section</a>
</div>
<H2><a name="sec_hypseg"><U>Recognition Hypothesis Output</U></a></H2>
<P>The Sphinx-3 decoder produces a single recognition
<em>hypothesis</em> for each utterance it processes. The
hypotheses for all the utterances processed in a single run are
written to a single output file, one line per utterance. The line
format is as follows:
</P>
<pre><em>u</em> S <em>s</em> T <em>t</em> A <em>a</em> L <em>l</em> <em>sf</em> <em>wa</em> <em>wl</em> <em>wd</em> <em>sf</em> <em>wa</em> <em>sl</em> <em>wd</em> ... <em>nf</em></pre>
<P>The <code>S</code>, <code>T</code>, <code>A</code>, and
<code>L</code> fields are keywords and appear in the output as
shown. The remaining fields are briefly described below:
</P>
<UL>
<LI><em>u</em>: the utterance ID
<LI><em>s</em>: an acoustic score scaling done during acoustic
likelihood computation (However, this field is 0 in the s3.X
decoder output.)
<LI><em>t</em>: the total score for this hypothesis
<LI><em>a</em>: the total acoustic score for this hypothesis
<LI><em>l</em>: the total language model score for this hypothesis
</UL>
<P>The <em>l score</em> field is followed by groups of four fields,
one group for each successive word in the output hypothesis. The
four fields are:
</P>
<UL>
<LI><em>sf</em>: Start frame for the word (its end frame is just
before the start frame of the next word)
<LI><em>wa</em>: Acoustic score for the word
<LI><em>wl</em>: LM score for the word
<LI><em>wd</em>: The word string itself.
</UL>
<P>The final field, <em>nf</em>, in each hypothesis line is the
total number of frames in the utterance.</P>
<P>Note that all scores are <em>log-likelihood</em> values in the
peculiar logbase used by the decoder. Secondly, the acoustic
scores are <em>scaled</em> values; in each frame, the acoustic
scores of all active <a href="#sec_am">senones</a> are scaled such
that the best senone has a log-likelihood of 0. Finally, the
language model scores reported include the <a
href="#lm_lw_wip"><em>language weight</em> and <em>word-insertion
penalty</em></a> parameters.
</P>
<P>Here is an <a href="./s3/hypseg.txt">example hypothesis file</a>
for three utterances.
</P>
<div class="endsec">
¤
<a href="#sec_hypseg">Back to top of this section</a>
</div>
<H2><a name="sec_wordlat"><U>Word Lattice Output</U></a></H2>
<div class="wheatbox">
<UL>
<LI><A HREF="#wordlat_overview">Word Lattice Overview</A></LI>
<LI><A HREF="#wordlat_format">Word Lattice File Format</A></LI>
</UL>
</div>
<h3><a name="wordlat_overview"><U>Word Lattice Overview</U></a></H3>
<P>During recognition the decoder maintains not just the single
best hypothesis, but also a number of alternatives or candidates.
For example, <code>REED</code> is a perfectly reasonable
alternative to <code>READ</code>. The alternatives are useful in
many ways: for instance, in N-best list generation. To facilitate
such <em>post-processing</em>, the decoder can optionally produce
a <em>word lattice</em> output for each input utterance. This
output records all the candidate words recognized by the decoder
at any point in time, and their main attributes such as time
segmentation and acoustic likelihood scores.
</P>
<P>The term "lattice" is used somewhat loosely. The word-lattice
is really a <em>directed acyclic graph</em> or <em>DAG</em>. Each
node of the DAG denotes a word instance that begins at a
particular frame within the utterance. That is, it is a unique
<code><word,start-time></code> pair. (However, there could
be a number of end-times for this word instance. One of the
features of a time-synchronous Viterbi search using beam pruning
is that word candidates hypothesized by the decoder have a
well-defined start-time, but a fuzzy range of end-times. This is
because the start-time is primarily determined by <em>Viterbi
pruning</em>, while the possible end-times are determined by beam
pruning.)
</P>
<P>There is a directed edge between two nodes in the DAG if the
start-time of the destination node immediately follows one of the
end times of the source node. That is, the two nodes can be
adjacent in time. Thus, the edge determines one possible
segmentation for the source node: beginning at the source's
start-time and ending one frame before the destination's
start-time. The edge also contains an acoustic likelihood for
this particular segmentation of the source node.
</P>
<P><em>Note:</em> The beginning and end of sentence tokens,
<code><s></code> and <code></s></code>, are not decoded
as part of an utterance by the s3.X decoder. However, they have
to be included in the word lattice file, for compatibility with
the older Sphinx-3 decoder software. They are assigned 1-frame
segmentations, with log-likelihood scores of 0. To accommodate
them, the segmentations of adjacent nodes have to be "fudged" by 1
frame.
</P>
<h3><a name="wordlat_format"><U>Word Lattice File Format</U></a></H3>
<P>A word lattice file essentially contains the above information
regarding the nodes and edges in the DAG. It is structured in
several sections, as follows:
</P>
<OL>
<LI>A comment section, listing important configuration arguments
as comments</LI>
<LI><code>Frames</code> section, specifying the number of frames
in utterance</LI>
<LI><code>Nodes</code> section, listing the nodes in the DAG</LI>
<LI><code>Initial</code> and <code>Final</code> nodes (for
<code><s></code> and <code></s></code>,
respectively)</LI>
<LI><code>BestSegAscr</code> section, a historical remnant now
essentially empty</LI>
<LI><code>Edges</code> section, listing the edges in the DAG</LI>
</OL>
<P>The file is formatted as follows. Note that any line in the
file that begins with the <code>#</code> character in the first
column is considered to be a comment.
</P>
<pre>
# getcwd: <current-working-directory>
# -logbase <logbase-in-effect>
# -dict <main lexicon>
# -fdict <filler lexicon>
# ... (other arguments, written out as comment lines)
#
Frames <number-of-frames-in-utterance>
#
Nodes <number-of-nodes-in-DAG> (NODEID WORD STARTFRAME FIRST-ENDFRAME LAST-ENDFRAME)
<Node-ID> <Word-String> <Start-Time> <Earliest-End-time> <Latest-End-Time>
<Node-ID> <Word-String> <Start-Time> <Earliest-End-time> <Latest-End-Time>
<Node-ID> <Word-String> <Start-Time> <Earliest-End-time> <Latest-End-Time>
... (for all nodes in DAG)
#
Initial <Initial-Node-ID>
Final <Final-Node-ID>
#
BestSegAscr 0 (NODEID ENDFRAME ASCORE)
#
Edges (FROM-NODEID TO-NODEID ASCORE)
<Source-Node-ID> <Destination-Node-ID> <Acoustic Score>
<Source-Node-ID> <Destination-Node-ID> <Acoustic Score>
<Source-Node-ID> <Destination-Node-ID> <Acoustic Score>
... (for all edges in DAG)
End</pre>
<P>Note that the <em>node-ID</em> values for DAG nodes are
assigned sequentially, starting from 0. Furthermore, they are
sorted in <em>descending order</em> of their
<em>earliest-end-time</em> attribute.
</P>
<P>Here is an <a href="4t0c020c.lat">example word lattice</a> file.</P>
<div class="endsec">
¤
<a href="#sec_wordlat">Back to top of this section</a>
</div>
<H2><a name="sec_utilpgm"><U>Other Utilities</U></a></H2>
<P>In addition to the s3.X decoders, <code>sphinx3_decode</code>,
<code>sphinx3_livedecode</code> and <code>livrepretend</code>, this
distribution also provides other utility programs.</P>
<H4><a name="sec_gausubvq"><U>Gaussian Sub-Vector Quantization
Utility</U></A></H4>
<P></P>
<div class="endsec">
¤
<a href="#sec_utilpgm">Back to top of this section</a>
</div>
<H2><a name="sec_src"><U>Source Code</U></a></H2>
<P>In alphabetical order:</P>
<table>
<tr>
<td><a href="../src/libs3decoder/libam/adaptor.c"><code>adaptor.c</code></a></td>
<td>A wrapper of adaptation routines. </td>
</tr>
<tr>
<td><a href="../src/libs3decoder/libam/approx_cont_mgau.c"><code>approx_cont_mgau.c</code></a></td>
<td>Fast Gaussian Distribution Computation</td>
</tr>
<tr>
<td><a href="../src/libs3decoder/libcep_feat/agc.c"><code>agc.c</code></a></td>
<td>Automatic gain control (on signal energy)</td>
</tr>
<tr>
<td><a href="../src/libs3decoder/libsearch/ascr.c"><code>ascr.c</code></a></td>
<td>Senone acoustic scores</td>
</tr>
<tr>
<td><a href="../src/libs3decoder/libcommon/bio.c"><code>bio.c</code></a></td>
<td>Binary file I/O support</td>
</tr>
<tr>
<td><a href="../src/libs3decoder/libcommon/blkarray.c"><code>blkarray.c</code></a></td>
<td>Block array used in FST search</td>
</tr>
<tr>
<td><a href="../src/libs3decoder/libam/cb2mllr.c"><code>cb2mllr.c</code></a></td>
<td>Codebook to MLLR mapping</td>
</tr>
<tr>
<td><a href="../src/libs3decoder/libep/classify.c"><code>classify.c</code></a></td>
<td>GMM Classifier. </td>
</tr>
<tr>
<td><a href="../src/libs3decoder/libcep_feat/cmn.c"><code>cmn.c</code></a></td>
<td>Cepstral mean normalization and variance normalization</td>
</tr>
<tr>
<td><a href="../src/libs3decoder/libdict/cmu6_lts_rules.c"><code>cmn6_lts_rules.c</code></a></td>
<td>Letter-to-sound rules from flite.</td>
</tr>
<tr>
<td><a href="../src/libs3decoder/libdict/ctxt_table.c"><code>ctxt_table.c</code></a></td>
<td>Context table for search functions. </td>
</tr>
<tr>
<td><a href="../src/libs3decoder/libconfidence/confidence.c"><code>confidence.c</code></a></td>
<td>Word-lattice based word-based confidence scoring. </td>
</tr>
<tr>
<td><a href="../src/libs3decoder/libcommon/corpus.c"><code>corpus.c</code></a></td>
<td>Control file processing</td>
</tr>
<tr>
<td><a href="../src/libs3decoder/libcommon/encoding.c"><code>encoding.c</code></a></td>
<td>Take care of text-encoding issues. </td>
</tr>
<tr>
<td><a href="../src/libs3decoder/cont_mgau.c"><code>cont_mgau.c</code></a></td>
<td>Mixture Gaussians (acoustic model)</td>
</tr>
<tr>
<td><a
href="../src/programs/decode.c"><code>decode.c</code></a></td>
<td>Main file for <code>sphinx3_decode</code></td>
</tr>
<tr>
<td><a href="../src/libs3decoder/libdict/dict.c"><code>dict.c</code></a></td>
<td>Pronunciation lexicon</td>
</tr>
<tr>
<td><a href="../src/libs3decoder/libdict/dict2pid.c"><code>dict2pid.c</code></a></td>
<td>Generation of triphones for the pronunciation dictionary</td>
</tr>
<tr>
<td><a href="../src/libs3decoder/libep/endptr.c"><code>endptr.c</code></a></td>
<td>Voting-based end-pointer.</td>
</tr>
<tr>
<td><a href="../src/libs3decoder/libam/fast_algo_struct.c"><code>fast_algo_struct.c</code></a></td>
<td>Structures of all fast algorithm. </td>
</tr>
<tr>
<td><a href="../src/libs3decoder/libfeat/feat.c"><code>feat.c</code></a></td>
<td>Feature vectors computation</td>
</tr>
<tr>
<td><a href="../src/libs3decoder/libdict/fillpen.c"><code>fillpen.c</code></a></td>
<td>Filler word probabilities</td>
</tr>
<tr>
<td><a href="../src/libs3decoder/libsearch/flat_fwd.c"><code>flat_fwd.c</code></a></td>
<td>Implementation of flat lexicon decoder. </td>
</tr>
<tr>
<td><a href="../src/libs3decoder/libsearch/fsg_history.c"><code>fsg_history.c</code></a></td>
<td>The history table used in the FSG search </td>
</tr>
<tr>
<td><a href="../src/libs3decoder/libsearch/fsg_lextree.c"><code>fsg_lextree.c</code></a></td>
<td>It is actually flat structure at this point. </td>
</tr>
<tr>
<td><a href="../src/programs/gausubvq.c"><code>gausubvq.c</code></a></td>
<td>Standalone acoustic model sub-vector quantizer</td>
</tr>
<tr>
<td><a href="../src/libs3decode/libam/gs.c"><code>gs.c</code></a></td>
<td>Gaussian selector by Bochierri</td>
</tr>
<tr>
<td><a href="../src/libs3decoder/hmm.c"><code>hmm.c</code></a></td>
<td>HMM evaluation</td>
</tr>
<tr>
<td><a href="../include/hyp.h"><code>hyp.h</code></a></td>
<td>Recognition hypotheses data type</td>
</tr>
<tr>
<td><a href="../src/libs3decoder/libam/interp.c"><code>interp.c</code></a></td>
<td>Interpolation of scoustic models</td>
</tr>
<tr>
<td><a href="../include/kb.h"><code>kb.h</code></a></td>
<td>All knowledge bases, search structures used by decoder</td>
</tr>
<tr>
<td><a href="../src/libs3decoder/libsearch/kbcore.c"><code>kbcore.c</code></a></td>
<td>Collection of core knowledge bases</td>
</tr>
<tr>
<td><a href="../src/libs3decoder/kdtree.c"><code>kdtree.c</code></a></td>
<td>KD-tree support in semi-continous HMM</td>
</tr>
<tr>
<td><a href="../src/libs3decoder/lextree.c"><code>lextree.c</code></a></td>
<td>Lexical search tree</td>
</tr>
<tr>
<td><a href="../src/libs3decoder/libAPI/live_decode_args.c"><code>live_decode_args.c</code></a></td>
<td>Argument definition of the livedecoder. </td>
</tr>
<tr>
<td><a href="../src/libs3decoder/libAPI/live_decode_API.c"><code>live_decode_API.c</code></a></td>
<td>Live decoder functions</td>
</tr>
<tr>
<td><a href="../src/libs3decoder/libdict/lts.c"><code>lts.c</code></a></td>
<td>Letter to sound rules. </td>
</tr>
<tr>
<td><a href="../src/libs3decoder/liblm/lm.c"><code>lm.c</code></a></td>
<td>Trigram language model, top-level controller module</td>
</tr>
<tr>
<td><a href="../src/libs3decoder/liblm/lm_3g.c"><code>lm_3g.c</code></a></td>
<td>Trigram language model, TXT file driver</td>
</tr>
<tr>
<td><a href="../src/libs3decoder/liblm/lm_3g_dmp.c"><code>lm_3g_dmp.c</code></a></td>
<td>Trigram language model, DMP file driver</td>
</tr>
<tr>
<td><a href="../src/libs3decoder/liblm/lm_attfsm.c"><code>lm_attfsm.c</code></a></td>
<td>Trigram language model, ATT-FSM file format driver</td>
</tr>
<tr>
<td><a href="../src/libs3decoder/liblm/lm_class.c"><code>lm_class.c</code></a></td>
<td>Handling of class-based LM </td>
</tr>
<tr>
<td><a href="../src/libs3decoder/liblm/lmset.c"><code>lmset.c</code></a></td>
<td>Handling of a set of LMs </td>
</tr>
<tr>
<td><a href="../src/libs3decoder/libcommon/logs3.c"><code>logs3.c</code></a></td>
<td>Support for log-likelihood operations</td>
</tr>
<tr>
<td><a
href="../src/programs/main_live_example.c"><code>main_live_example.c</code></a></td>
<td>Main file for <code>s3livedecode</code> showing use of
<code>live_decode_API.h</code></td>
</tr>
<tr>
<td><a
href="../src/programs/main_live_pretend.c"><code>main_live_pretend.c</code></a></td>
<td>Main file for <code>s3livepretend</code> showing use of
<code>live_decode_API.h</code></td>
</tr>
<tr>
<td><a href="../src/libs3decoder/libam/mdef.c"><code>mdef.c</code></a></td>
<td>Acoustic model definition</td>
</tr>
<tr>
<td><a href="../src/libs3decoder/libcommon/misc.c"><code>misc.c</code></a></td>
<td>Miscellaneous routines used in Sphinx 3</td>
</tr>
<tr>
<td><a href="../src/libs3decoder/libam/mllr.c"><code>mllr.c</code></a></td>
<td>transformation of mean based on a linear regression matrix. </td>
</tr>
<tr>
<td><a href="../src/libs3decoder/libam/ms_gauden.c"><code>ms_gauden.c</code></a></td>
<td>Multi-stream Gaussian computation. (Adapted from s3.0) </td>
</tr>
<tr>
<td><a href="../src/libs3decoder/libam/ms_mllr.c"><code>ms_mllr.c</code></a></td>
<td>Multi-stream MLLR. (Adapted from s3.0) </td>
</tr>
<tr>
<td><a href="../src/libs3decoder/libam/ms_senone.c"><code>ms_senone.c</code></a></td>
<td>Multi-stream Senone computation. (Adapted from s3.0) </td>
</tr>
<tr>
<td><a href="../include/s2_semi_mgau.c"><code>s2_semi_mgau.c</code></a></td>
<td>Sphinx 2 semi-continuous HMM computation. </td>
</tr>
<tr>
<td><a href="../include/s3types.h"><code>s3types.h</code></a></td>
<td>Various data types, for ease of modification</td>
</tr>
<tr>
<td><a href="../src/libs3decoder/libcommon/stat.c"><code>stat.c</code></a></td>
<td>Statistics of decoding.</td>
</tr>
<tr>
<td><a href="../src/libs3decoder/subvq.c"><code>subvq.c</code></a></td>
<td>Sub-vector quantized acoustic model</td>
</tr>
<tr>
<td><a href="../src/libs3decoder/libam/tmat.c"><code>tmat.c</code></a></td>
<td>HMM transition matrices (topology definition)</td>
</tr>
<tr>
<td><a href="../src/libs3decoder/libcommon/vector.c"><code>vector.c</code></a></td>
<td>Vector operations, quantization, etc.</td>
</tr>
<tr>
<td><a href="../src/libs3decoder/libsearch/vithist.c"><code>vithist.c</code></a></td>
<td>Backpointer table (Viterbi history)</td>
</tr>
<tr>
<td><a href="../src/libs3decoder/libdict/wid.c"><code>wid.c</code></a></td>
<td>Mapping between LM and lexicon word IDs</td>
</tr>
</table>
<P></P>
<div class="endsec">
¤
<a href="#sec_src">Back to top of this section</a>
</div>
<P></P>
<H2></H2><!-- Just to provide some space -->
<address>Maintained by <a href="mailto:egouvea+sourceforge@cs.cmu.edu">Evandro B. Gouvêa<a> and <a href="mailto:archan+sourceforge@cs.cmu.edu"> Arthur Chan <a> </address>
<!-- Created: Sun Feb 22 14:03:14 EST 1998 -->
<!-- hhmts start -->
Last modified: Thu Jul 22 09:35:27 EDT 2004
<!-- hhmts end -->
</body>
</html>
|