File size: 81,261 Bytes
6fa4bc9 |
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 |
{
"paper_id": "2021",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T07:33:41.820049Z"
},
"title": "Low-Resource ASR with an Augmented Language Model",
"authors": [
{
"first": "Timofey",
"middle": [],
"last": "Arkhangelskiy",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Universit\u00e4t Hamburg",
"location": {}
},
"email": "timarkh@gmail.com"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "It is widely known that a good language model (LM) can dramatically improve the quality of automatic speech recognition (ASR). However, when dealing with a lowresource language, it is often the case that not only aligned audio data is scarce, but there are also not enough texts to train a good LM. This is the case of Beserman, an unwritten dialect of Udmurt (Uralic > Permic). With about 10 hours of aligned audio and about 164K words of texts available for training, the word error rate of a Deepspeech model with the best set of parameters equals 56.4%. However, there are other linguistic resources available for Beserman, namely a bilingual Beserman-Russian dictionary and a rule-based morphological analyzer. The goal of this paper is to explore whether and how these additional resources can be exploited to improve the ASR quality. Specifically, I attempt to use them in order to expand the existing LM by generating a large number of fake sentences that in some way look like genuine Beserman text. It turns out that a sophisticated enough augmented LM generator can indeed improve the ASR quality. Nevertheless, the improvement is far from dramatic, with about 5% decrease in word error rate (WER) and 2% decrease in character error rate (CER).",
"pdf_parse": {
"paper_id": "2021",
"_pdf_hash": "",
"abstract": [
{
"text": "It is widely known that a good language model (LM) can dramatically improve the quality of automatic speech recognition (ASR). However, when dealing with a lowresource language, it is often the case that not only aligned audio data is scarce, but there are also not enough texts to train a good LM. This is the case of Beserman, an unwritten dialect of Udmurt (Uralic > Permic). With about 10 hours of aligned audio and about 164K words of texts available for training, the word error rate of a Deepspeech model with the best set of parameters equals 56.4%. However, there are other linguistic resources available for Beserman, namely a bilingual Beserman-Russian dictionary and a rule-based morphological analyzer. The goal of this paper is to explore whether and how these additional resources can be exploited to improve the ASR quality. Specifically, I attempt to use them in order to expand the existing LM by generating a large number of fake sentences that in some way look like genuine Beserman text. It turns out that a sophisticated enough augmented LM generator can indeed improve the ASR quality. Nevertheless, the improvement is far from dramatic, with about 5% decrease in word error rate (WER) and 2% decrease in character error rate (CER).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "\u043a\u0435 \u0440\u0430\u0441\u043f\u043e\u0437\u043d\u0430\u0432\u0430\u0442\u0435\u043b\u0435\u0437 \u0434\u044b\u0448\u0435\u0442\u043e\u043d \u043f\u043e\u043d\u043d\u0430 \u0432\u043e\u043b\u044f\u0442\u044d\u043c \u043a\u0443\u0430\u0440\u0430, \u043e\u0437\u044c\u044b \u0438\u043a \u0443\u043c\u043e\u0439 \u043a\u044b\u043b \u043c\u043e\u0434\u0435\u043b\u0435\u0437 \u0434\u044b\u0448\u0435\u0442\u043e\u043d \u043f\u043e\u043d\u043d\u0430 \u0442\u0435\u043a\u0441\u0442\u044a\u0451\u0441 \u04f5\u0435\u043c\u044b\u0441\u044c \u0442\u0443\u0436 \u04e7\u0436\u044b\u0442 \u043b\u0443\u043e. \u04f4\u0430\u043f\u0430\u043a \u0442\u0430\u04f5\u0435 \u044e\u0433\u0434\u0443\u0440 \u0443\u0434\u043c\u0443\u0440\u0442 \u043a\u044b\u043b\u043b\u044d\u043d \u0433\u043e\u0436\u044a\u044f\u0441\u044c\u043a\u0435\u0442\u0442\u044d\u043c \u0431\u0435\u0441\u0435\u0440\u043c\u0430\u043d \u0432\u0435\u0440\u0430\u0441\u044c\u043a\u0435\u0442\u044d\u043d\u044b\u0437 \u043a\u044b\u043b\u0434\u044d\u043c\u044b\u043d. \u041a\u0438 \u0443\u043b\u0430\u043c\u044b \u0432\u0430\u043d\u044c 10 \u0447\u0430\u0441 \u043f\u0430\u043b\u0430 \u0432\u043e\u043b\u044f\u0442\u044d\u043c \u043d\u043e \u0440\u0430\u0441\u0448\u0438\u0444\u0440\u043e\u0432\u0430\u0442\u044c \u043a\u0430\u0440\u0435\u043c \u043a\u0443\u0430\u0440\u0430 \u043d\u043e 164 \u0441\u044e\u0440\u0441 \u043f\u0430\u043b\u0430 \u0443\u0436\u0435 \u043a\u0443\u0442\u044d\u043c \u043a\u044b\u043b\u044a\u0451\u0441\u044b\u043d \u0442\u0435\u043a\u0441\u0442\u044a\u0451\u0441. \u0422\u0430 \u0442\u043e\u0434\u044d\u0442\u044a\u0451\u0441\u044b\u043d \u0434\u044b\u0448\u0435\u0442\u0441\u043a\u044b\u0441\u0430, Deepspeech \u0441\u0438\u0441\u0442\u0435\u043c\u0430 \u0432\u043e\u0437\u044c\u043c\u0430\u0442\u044d 56,4% WER (\u043c\u044b\u0434\u043b\u0430\u043d\u044c \u0440\u0430\u0441\u043f\u043e\u0437\u043d\u0430\u0442\u044c \u043a\u0430\u0440\u0435\u043c \u043a\u044b\u043b\u044a\u0451\u0441\u043b\u044d\u043d \u043f\u0440\u043e\u0446\u0435\u043d\u0442\u0441\u044b).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "\u041e\u0437\u044c\u044b \u043a\u0435 \u043d\u043e \u0431\u0435\u0441\u0435\u0440\u043c\u0430\u043d \u0432\u0435\u0440\u0430\u0441\u044c\u043a\u0435\u0442\u044a\u044f \u0432\u0430\u043d\u044c \u043d\u0430 \u043c\u0443\u043a\u0435\u0442 \u043a\u044b\u043b\u0442\u043e\u0434\u043e\u043d \u0432\u0430\u043d\u0451\u0441\u044a\u0451\u0441: \u0431\u0435\u0441\u0435\u0440\u043c\u0430\u043d-\u04df\u0443\u0447 \u043a\u044b\u043b\u043b\u044e\u043a\u0430\u043c \u043d\u043e \u0448\u043e\u043d\u0435\u0440\u0440\u0430\u0434\u044a\u044f\u043d \u043c\u043e\u0440\u0444\u043e\u043b\u043e\u0433\u0438 \u0430\u043d\u0430\u043b\u0438\u0437\u0430\u0442\u043e\u0440. \u0422\u0430 \u0443\u0436\u043b\u044d\u043d \u0446\u0435\u043b\u0435\u0437 -\u0432\u0430\u043b\u0430\u043d\u044b, \u043b\u0443\u044d-\u0430 \u0432\u0435\u0440\u0430\u0441\u044c\u043a\u0435\u043c\u0435\u0437 \u0430\u0441\u044d\u0440\u043a\u0430\u0437 \u0442\u043e\u0434\u043c\u0430\u043d\u043b\u044d\u0441\u044c \u04df\u0435\u0447\u043b\u044b\u043a\u0441\u044d \u0431\u0443\u0434\u044d\u0442\u043e\u043d \u043f\u043e\u043d\u043d\u0430 \u0442\u0430 \u0432\u0430\u0442\u0441\u0430\u043c \u0440\u0435\u0441\u0443\u0440\u0441\u044a\u0451\u0441\u0442\u044b \u0443\u0436\u0435 \u043a\u0443\u0442\u044b\u043d\u044b. \u041a\u044b\u043b\u0441\u044f\u0440\u044b\u0441\u044c, \u0441\u043e\u043e\u0441 \u0432\u044b\u043b\u044d \u043f\u044b\u043a\u044a\u044f\u0441\u044c\u043a\u044b\u0441\u0430, \u0442\u0443\u0440\u0442\u0442\u044d\u043c\u044b\u043d \u043a\u044b\u043b \u043c\u043e\u0434\u0435\u043b\u0435\u0437 \u043f\u0430\u0441\u044c\u043a\u044b\u0442\u0430\u0442\u044b\u043d\u044b, \u0441\u043e \u043f\u043e\u043d\u043d\u0430 \u043a\u044b\u043b\u0434\u044b\u0442\u044d\u043c\u044b\u043d \u0432\u0430\u043b \u0442\u0440\u043e\u0441 \u0437\u044d\u043c\u043e\u0441 \u043b\u0443\u0438\u0441\u044c\u0442\u044d\u043c \u0448\u0443\u043e\u0441\u044a\u0451\u0441, \u043a\u0443\u0434\u044a\u0451\u0441\u044b\u0437 \u043a\u0443\u0434-\u043e\u0433 \u043b\u0430\u0441\u044f\u043d\u044c \u0442\u0443\u043f\u0430\u043b\u043e \u0437\u044d\u043c\u043e\u0441 \u0431\u0435\u0441\u0435\u0440\u043c\u0430\u043d \u0442\u0435\u043a\u0441\u0442\u043b\u044b. \u0428\u0443\u043e\u0441\u044a\u0451\u0441\u0442\u044b \u043a\u044b\u043b\u0434\u044b\u0442\u04e5\u0441\u044c \u0433\u0435\u043d\u0435\u0440\u0430\u0442\u043e\u0440 \u0442\u044b\u0440\u043c\u044b\u0442 \u00ab\u0432\u0438\u0437\u044c\u043c\u043e\u00bb \u043a\u0435, \u0440\u0430\u0441\u043f\u043e\u0437\u043d\u0430\u0432\u0430\u043d\u0438\u043b\u044d\u043d \u04df\u0435\u0447\u043b\u044b\u043a\u0435\u0437 \u0437\u044d\u043c\u0437\u044d \u043d\u043e \u0431\u0443\u0434\u044d \u0432\u044b\u043b\u044d\u043c. \u041e\u0437\u044c\u044b \u043a\u0435 \u043d\u043e \u0442\u0430 \u0443\u043c\u043e\u044f\u043d \u0448\u04e7\u0434\u0441\u043a\u044b\u043c\u043e\u043d \u043b\u0443\u044d \u0448\u0443\u044b\u0441\u0430, \u0432\u0435\u0440\u0430\u043d\u044b \u0443\u0433 \u043b\u0443\u044b: WER \u0432\u043e\u0437\u044c\u043c\u0430\u0442\u043e\u043d \u0443\u0441\u0435 5%-\u043b\u044b \u043f\u0430\u043b\u0430, \u043d\u043e\u0448 CER (\u043c\u044b\u0434\u043b\u0430\u043d\u044c \u0440\u0430\u0441\u043f\u043e\u0437\u043d\u0430\u0442\u044c \u043a\u0430\u0440\u0435\u043c \u0431\u0443\u043a\u0432\u0430\u043e\u0441\u043b\u044d\u043d \u043f\u0440\u043e\u0446\u0435\u043d\u0442\u0441\u044b) -2%-\u043b\u044b. \u043e\u0431\u0443\u0447\u0435\u043d\u0438\u044f \u0440\u0430\u0441\u043f\u043e\u0437\u043d\u0430\u0432\u0430\u0442\u0435\u043b\u044f, \u043d\u043e \u0438 \u0441\u043b\u0438\u0448\u043a\u043e\u043c \u043c\u0430\u043b\u043e \u0442\u0435\u043a\u0441\u0442\u043e\u0432 \u0434\u043b\u044f \u043e\u0431\u0443\u0447\u0435\u043d\u0438\u044f \u0445\u043e\u0440\u043e\u0448\u0435\u0439 \u044f\u0437\u044b\u043a\u043e\u0432\u043e\u0439 \u043c\u043e\u0434\u0435\u043b\u0438. \u0418\u043c\u0435\u043d\u043d\u043e \u0442\u0430\u043a\u043e\u0432 \u0441\u043b\u0443\u0447\u0430\u0439 \u0431\u0435\u0441\u0435\u0440\u043c\u044f\u043d\u0441\u043a\u043e\u0433\u043e -\u0431\u0435\u0441\u043f\u0438\u0441\u044c\u043c\u0435\u043d\u043d\u043e\u0433\u043e \u0434\u0438\u0430\u043b\u0435\u043a\u0442\u0430 \u0443\u0434\u043c\u0443\u0440\u0442\u0441\u043a\u043e\u0433\u043e \u044f\u0437\u044b\u043a\u0430. \u0412 \u043d\u0430\u0448\u0435\u043c \u0440\u0430\u0441\u043f\u043e\u0440\u044f\u0436\u0435\u043d\u0438\u0438 \u0438\u043c\u0435\u044e\u0442\u0441\u044f \u043e\u043a\u043e\u043b\u043e 10 \u0447\u0430\u0441\u043e\u0432 \u0437\u0432\u0443\u043a\u0430, \u0432\u044b\u0440\u043e\u0432\u043d\u0435\u043d\u043d\u043e\u0433\u043e \u0441 \u0440\u0430\u0441\u0448\u0438\u0444\u0440\u043e\u0432\u043a\u0430\u043c\u0438, \u0438 \u0442\u0435\u043a\u0441\u0442\u044b \u043e\u0431\u044a\u0451\u043c\u043e\u043c \u043e\u043a\u043e\u043b\u043e 164 \u0442\u044b\u0441. \u0441\u043b\u043e\u0432\u043e\u0443\u043f\u043e\u0442\u0440\u0435\u0431\u043b\u0435\u043d\u0438\u0439. \u041e\u0431\u0443\u0447\u0438\u0432\u0448\u0438\u0441\u044c \u043d\u0430 \u044d\u0442\u0438\u0445 \u0434\u0430\u043d\u043d\u044b\u0445, \u0441\u0438\u0441\u0442\u0435\u043c\u0430 Deepspeech \u0434\u0435\u043c\u043e\u043d\u0441\u0442\u0440\u0438\u0440\u0443\u0435\u0442 WER (\u043f\u0440\u043e\u0446\u0435\u043d\u0442 \u043d\u0435\u043f\u0440\u0430\u0432\u0438\u043b\u044c\u043d\u043e \u0440\u0430\u0441\u043f\u043e\u0437\u043d\u0430\u043d\u043d\u044b\u0445 \u0441\u043b\u043e\u0432), \u0440\u0430\u0432\u043d\u044b\u0439 56,4%. \u041e\u0434\u043d\u0430\u043a\u043e \u0434\u043b\u044f \u0431\u0435\u0441\u0435\u0440\u043c\u044f\u043d\u0441\u043a\u043e\u0433\u043e \u0441\u0443\u0449\u0435\u0441\u0442\u0432\u0443\u044e\u0442 \u0434\u0440\u0443\u0433\u0438\u0435 \u043b\u0438\u043d\u0433\u0432\u0438\u0441\u0442\u0438\u0447\u0435\u0441\u043a\u0438\u0435 \u0440\u0435\u0441\u0443\u0440\u0441\u044b, \u0430 \u0438\u043c\u0435\u043d\u043d\u043e \u0431\u0435\u0441\u0435\u0440\u043c\u044f\u043d\u0441\u043a\u043e-\u0440\u0443\u0441\u0441\u043a\u0438\u0439 \u0441\u043b\u043e\u0432\u0430\u0440\u044c \u0438 \u043f\u0440\u0430\u0432\u0438\u043b\u043e\u0432\u044b\u0439 \u043c\u043e\u0440\u0444\u043e\u043b\u043e\u0433\u0438\u0447\u0435\u0441\u043a\u0438\u0439 \u0430\u043d\u0430\u043b\u0438\u0437\u0430\u0442\u043e\u0440. \u0426\u0435\u043b\u044c \u044d\u0442\u043e\u0439 \u0440\u0430\u0431\u043e\u0442\u044b -\u0432\u044b\u044f\u0441\u043d\u0438\u0442\u044c, \u043c\u043e\u0436\u043d\u043e \u043b\u0438 \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u043e\u0432\u0430\u0442\u044c \u044d\u0442\u0438 \u0434\u043e\u043f\u043e\u043b\u043d\u0438\u0442\u0435\u043b\u044c\u043d\u044b\u0435 \u0440\u0435\u0441\u0443\u0440\u0441\u044b \u0434\u043b\u044f \u0443\u043b\u0443\u0447\u0448\u0435\u043d\u0438\u044f \u0440\u0430\u0441\u043f\u043e\u0437\u043d\u0430\u0432\u0430\u043d\u0438\u044f \u0440\u0435\u0447\u0438. \u0412 \u0447\u0430\u0441\u0442\u043d\u043e\u0441\u0442\u0438, \u043f\u0440\u0435\u0434\u043f\u0440\u0438\u043d\u0438\u043c\u0430\u0435\u0442\u0441\u044f \u043f\u043e\u043f\u044b\u0442\u043a\u0430 \u0440\u0430\u0441\u0448\u0438\u0440\u0438\u0442\u044c \u0441 \u0438\u0445 \u043f\u043e\u043c\u043e\u0449\u044c\u044e \u044f\u0437\u044b\u043a\u043e\u0432\u0443\u044e \u043c\u043e\u0434\u0435\u043b\u044c \u043f\u0443\u0442\u0451\u043c \u043f\u043e\u0440\u043e\u0436\u0434\u0435\u043d\u0438\u044f \u0431\u043e\u043b\u044c\u0448\u043e\u0433\u043e \u043a\u043e\u043b\u0438\u0447\u0435\u0441\u0442\u0432\u0430 \u043d\u0435\u043d\u0430\u0441\u0442\u043e\u044f\u0449\u0438\u0445 \u043f\u0440\u0435\u0434\u043b\u043e\u0436\u0435\u043d\u0438\u0439, \u043a\u043e\u0442\u043e\u0440\u044b\u0435 \u0432 \u043d\u0435\u043a\u043e\u0442\u043e\u0440\u044b\u0445 \u043e\u0442\u043d\u043e\u0448\u0435\u043d\u0438\u044f\u0445 \u043f\u043e\u0445\u043e\u0436\u0438 \u043d\u0430 \u043d\u0430\u0441\u0442\u043e\u044f\u0449\u0438\u0439 \u0431\u0435\u0441\u0435\u0440\u043c\u044f\u043d\u0441\u043a\u0438\u0439 \u0442\u0435\u043a\u0441\u0442. \u041e\u043a\u0430\u0437\u044b\u0432\u0430\u0435\u0442\u0441\u044f, \u0447\u0442\u043e \u0435\u0441\u043b\u0438 \u0433\u0435\u043d\u0435\u0440\u0430\u0442\u043e\u0440 \u043f\u0440\u0435\u0434\u043b\u043e\u0436\u0435\u043d\u0438\u0439 \u0434\u043e\u0441\u0442\u0430\u0442\u043e\u0447\u043d\u043e \"\u0443\u043c\u0451\u043d\", \u043a\u0430\u0447\u0435\u0441\u0442\u0432\u043e \u0440\u0430\u0441\u043f\u043e\u0437\u043d\u0430\u0432\u0430\u043d\u0438\u044f \u043f\u043e\u0441\u043b\u0435 \u044d\u0442\u043e\u0433\u043e \u0434\u0435\u0439\u0441\u0442\u0432\u0438\u0442\u0435\u043b\u044c\u043d\u043e \u0432\u043e\u0437\u0440\u0430\u0441\u0442\u0430\u0435\u0442. \u041e\u0434\u043d\u0430\u043a\u043e \u044d\u0442\u043e \u0443\u043b\u0443\u0447\u0448\u0435\u043d\u0438\u0435 \u0432\u0440\u044f\u0434 \u043b\u0438 \u043c\u043e\u0436\u043d\u043e \u043d\u0430\u0437\u0432\u0430\u0442\u044c \u0441\u0443\u0449\u0435\u0441\u0442\u0432\u0435\u043d\u043d\u044b\u043c: \u043f\u043e\u043a\u0430\u0437\u0430\u0442\u0435\u043b\u044c WER \u043f\u0430\u0434\u0430\u0435\u0442 \u043f\u0440\u0438\u043c\u0435\u0440\u043d\u043e \u043d\u0430 5%, \u0430 CER (\u043f\u0440\u043e\u0446\u0435\u043d\u0442 \u043d\u0435\u043f\u0440\u0430\u0432\u0438\u043b\u044c\u043d\u043e \u0440\u0430\u0441\u043f\u043e\u0437\u043d\u0430\u043d\u043d\u044b\u0445 \u0431\u0443\u043a\u0432) -\u043d\u0430 2%.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "The key to reaching good ASR quality is having lots of data, i.e. thousands or at least hundreds of hours of text-aligned sound recordings. For most languages in the world, however, resources of that size are unavailable. With only a dozen hours of sound at hand, it is currently impossible to reach a WER low enough for the system to be usable in real-world applications. Nevertheless, a system with a WER, which is high, but lower than a certain threshold (e.g. 50%), could still be used in practice. Specifically, the primary motivation behind this research was the need to transcribe large amounts of spoken Beserman for subsequent linguistic research. If an ASR system, despite its high WER, could facilitate and accelerate manual transcription, that would be a useful practical application, even if limited in scope. Other possible applications of such undertrained noisy ASR systems have been proposed by Tyers and Meyer (2021) . This is why it makes sense to experiment with datasets that small.",
"cite_spans": [
{
"start": 912,
"end": 934,
"text": "Tyers and Meyer (2021)",
"ref_id": "BIBREF17"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "A number of techniques have been used to achieve better results in low-resource ASR systems. This includes pre-training the model on the data from another (possibly related or phonologically similar) language (Stoian et al., 2020) , augmenting the sound data with label-preserving transformations (T\u00fcske et al., 2014; Park et al., 2019) , and training the LM on a larger set of texts taken e.g. from a written corpus (Leinonen et al., 2018) . That a good language model can play an important role can be seen e.g. from the experiments on ASR for varieties of Komi, a language closely related to Udmurt, as described by (Hjortnaes et al., 2020b) and (Hjortnaes et al., 2020a) . Replacing a LM with a larger and more suitable one (in terms of domain) can decrease WER significantly.",
"cite_spans": [
{
"start": 209,
"end": 230,
"text": "(Stoian et al., 2020)",
"ref_id": "BIBREF15"
},
{
"start": 297,
"end": 317,
"text": "(T\u00fcske et al., 2014;",
"ref_id": "BIBREF16"
},
{
"start": 318,
"end": 336,
"text": "Park et al., 2019)",
"ref_id": "BIBREF14"
},
{
"start": 417,
"end": 440,
"text": "(Leinonen et al., 2018)",
"ref_id": "BIBREF11"
},
{
"start": 619,
"end": 644,
"text": "(Hjortnaes et al., 2020b)",
"ref_id": "BIBREF8"
},
{
"start": 649,
"end": 674,
"text": "(Hjortnaes et al., 2020a)",
"ref_id": "BIBREF7"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Beserman is traditionally classified as a dialect of Udmurt (Uralic > Permic) and is spoken by around 2200 people in NW Udmurtia, Russia. Unlike standard Udmurt, it lacks a codified orthography and is not used in the written form outside of scientific publications. This paper describes experiments with training Deepspeech (Hannun et al., 2014) on transcribed and elicited Beserman data. I am particularly interested in augmenting the LM with the help of linguistic resources that exist for Beserman: a Beserman-Russian dictionary and a morphological analyzer. The former is used, among other things, to transfer information from a model trained on Russian data. Same kinds of data augmentation could be relevant for many other under-resourced languages and dialects, since bilingual dictionaries and rule-based tools often exist for varieties, which are poor in raw data.",
"cite_spans": [
{
"start": 324,
"end": 345,
"text": "(Hannun et al., 2014)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The paper is organized as follows. In Section 2, I describe the dataset and lay out the reasons why improving the LM could be challenging. In Section 3, the training setup is outlined. In Section 4, I describe how the artificially augmented LM was generated. In 5, the original results are compared to that of the augmented LM. This is followed by a conclusion.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The Beserman dataset I have at hand consists of about 15,000 transcribed sound files with recordings from 10 speakers, both male and female, total-ing about 10 hours (with almost no trailing silence). Most of them come from a sound-aligned Beserman corpus, whose recordings were made in 2012-2019 and have varying quality. Another 2,700 files, totaling 2.5 hours, come from a sound dictionary and contain three pronunciations of a headword each. The duration of most files lies between 1 and 5 seconds. In addition to the texts of the sound-aligned corpus, there are transcriptions of older recordings, which are not sound-aligned as of now, and a corpus of usage examples based on the Beserman-Russian dictionary\u00b9 (Arkhangelskiy, 2019). All these sources combined contain about 27,400 written Beserman sentences (some very short, some occurring more than once), with a total of 164K words.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The data",
"sec_num": "2"
},
{
"text": "Such amount of textual data is insufficient for producing a well performing LM. Since Beserman is a morphologically rich language, most forms of most lexemes are absent from the sample and thus cannot be recognized, being out-of-vocabulary words. Unlike in some other studies mentioned above, it is hardly possible to find Beserman texts elsewhere. One way of doing that would be to use texts in literary Udmurt, which are available in larger quantities (tens of millions of words). Although I have not explored that option yet\u00b2, I doubt it could have the desired effect because the available Udmurt texts belong to a completely different domain. While most Beserman texts are narratives about the past or the life in the village, or everyday dialogues, most Udmurt texts available in digital form come from mass media. There is a pronounced difference between the vocabularies and grammatical constructions used in these two domains.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The data",
"sec_num": "2"
},
{
"text": "Instead, I attempt to utilize linguistic resources available for Beserman: a Beserman-Russian dictionary comprising about 6,000 entries and a morphological analyzer. The latter is rule-based and is based on the dictionary itself. Apart from the information necessary for morphological analysis, it contains some grammatical tags, such as animacy for nouns and transitivity for verbs. The analyzer \u00b9Available for search at http://beserman.ru; a large part of it has been published as Usacheva et al. (2017) .",
"cite_spans": [
{
"start": 483,
"end": 505,
"text": "Usacheva et al. (2017)",
"ref_id": "BIBREF18"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "The data",
"sec_num": "2"
},
{
"text": "\u00b2There are certain phonological, morphological and lexical differences between the standard language and the Beserman dialect. Before an Udmurt model can be used in Beserman ASR, the texts should be \"translated\" into Beserman. Although such attempts have been made (Miller, 2017) , making the translations look Beserman enough would require quite a lot of effort. recognizes about 97% of words in the textual part of the Beserman dataset. A small set of Constraint Grammar rules (Karlsson, 1990; Bick and Didriksen, 2015 ) is applied after the analysis, which reduces the average ambiguity to 1.25 analyses per analyzed word.",
"cite_spans": [
{
"start": 265,
"end": 279,
"text": "(Miller, 2017)",
"ref_id": "BIBREF12"
},
{
"start": 479,
"end": 495,
"text": "(Karlsson, 1990;",
"ref_id": "BIBREF9"
},
{
"start": 496,
"end": 520,
"text": "Bick and Didriksen, 2015",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "The data",
"sec_num": "2"
},
{
"text": "The idea is to inflate the text corpus used to produce the LM by generating a large number of fake sentences, using real corpus sentences as the starting point and the source of lemma frequencies, and incorporating data from the linguistic resources in the process.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The data",
"sec_num": "2"
},
{
"text": "All Beserman texts were encoded in a version of the Uralic Phonetic Alphabet so that each Unicode character represents one phoneme. Although there are a couple of regular phonetic processes not reflected in the transcription, such as optional final devoicing or regressive voicing of certain consonants, the characters almost always correspond to actual sounds. Therefore, CER values reported below must closely resemble PER (phone error rates)\u00b3. All sound files were transformed into 16 KHz, singlechannel format.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Deepspeech training",
"sec_num": "3"
},
{
"text": "Deepspeech architecture (Hannun et al., 2014 ) (Mozilla implementation\u2074) was used for training. This involves training a 5-layer neural network with one unidirectional LSTM layer. After each epoch, the quality is checked against a development dataset not used in training. After the training is complete, the evaluation is performed on the test dataset. The train/development/test split was randomly created once and did not change during the experiments. The development dataset contains 1737 sentences; the test dataset, 267 sentences. No sound dictionary examples were included in either development or test datasets, otherwise their unnaturally high quality would lead to overly optimistic WER and CER values. It has to be pointed out though that the training dataset contains data from all speakers of the test dataset. This is in line with the primary usage scenario I had in mind, i.e. pretranscription of field data, because most untranscribed recordings in my collection are generated by the same speakers. However, for a real-world scenario where the set of potential speakers is unlimited, this setting would \u00b3This property is the reason why UPA rather than Udmurt Cyrillic script was used for encoding. Otherwise, the choice of encoding is hardly important because UPA can be converted to Cyrillics and vice versa.",
"cite_spans": [
{
"start": 24,
"end": 44,
"text": "(Hannun et al., 2014",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Deepspeech training",
"sec_num": "3"
},
{
"text": "\u2074https://github.com/mozilla/DeepSpeech produce an overly optimistic estimate. No transfer learning was applied.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Deepspeech training",
"sec_num": "3"
},
{
"text": "A number of hyperparameter values were tested: learning rate between 0.00005 and 0.001, dropout rate 0.3 or 0.4, training batch size between 16 and 36. These value ranges have been demonstrated to yield optimal results on a dataset of similar size by (Agarwal and Zesch, 2019) . The results did not depend in any significant way on these values, except for almost immediate overfitting when the learning rate was close to 0.001. Under all these settings, the training ran for 8 or 9 epochs, after which the loss on the development dataset started rising due to overfitting. The model used for the evaluation was trained with the following parameters: learning rate 0.0002, droupout rate 0.4, training batch size 24.",
"cite_spans": [
{
"start": 251,
"end": 276,
"text": "(Agarwal and Zesch, 2019)",
"ref_id": "BIBREF0"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Deepspeech training",
"sec_num": "3"
},
{
"text": "The output of the trained Deepspeech model is filtered using kenlm, an n-gram language model (Heafield, 2011) . 3-gram and 4-gram models were tried, with no substantial difference; the figures below refer to the 4-gram models. When using the model with Deepspeech, there are two adjustable parameters, \u03b1 and \u03b2. \u03b1 (between 0 and 1) is the LM weight: higher values make the filter pay more attention to the n-gram probabilities provided by the model. \u03b2 defines the penalty for having too many words: higher values increase the average number of words in transcribed sentences and decrease the average length of a word. A number of \u03b1 and \u03b2 combinations were tested (see below).",
"cite_spans": [
{
"start": 93,
"end": 109,
"text": "(Heafield, 2011)",
"ref_id": "BIBREF6"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Deepspeech training",
"sec_num": "3"
},
{
"text": "As could be immediately seen from the test results, at least one of the reasons why the automatic transcription was wrong in many cases is that the corpus used to train the LM simply lacked the forms. Since Beserman is morphologically rich, a corpus of 164K words will inevitably lack most forms of most lexemes. Thankfully, this gap can be filled relatively easily, since Beserman morphological analyzer and dictionary can be turned into a morphological generator. (Another option, not explored here, would be to use subwords instead of words (Leinonen et al., 2018; Egorova and Burget, 2018) .) However, if one just generated all possible word forms and added them to the corpus packed into random sentences, that would completely skew the occurrence and co-occurrence probabilities of forms, which would lead to even worse performance. The real trick would be to add the lacking forms without losing too much information from the original model, i.e. without significantly distorting the probabilities. Specifically, one would need to make the following values as close to the original ones as possible:",
"cite_spans": [
{
"start": 544,
"end": 567,
"text": "(Leinonen et al., 2018;",
"ref_id": "BIBREF11"
},
{
"start": 568,
"end": 593,
"text": "Egorova and Burget, 2018)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Augmented language model",
"sec_num": "4"
},
{
"text": "\u2022 relative frequencies of lemmata;",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Augmented language model",
"sec_num": "4"
},
{
"text": "\u2022 relative frequencies of affix combinations, such as \"genitive plural\";",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Augmented language model",
"sec_num": "4"
},
{
"text": "\u2022 constraints on co-occurrence of certain grammatical forms (e.g. \"verb in the first person is not expected after a second-person pronoun\");",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Augmented language model",
"sec_num": "4"
},
{
"text": "\u2022 lexical constraints on contexts (e.g. \"mother eats apples\" should be fine, while \"apple eat mother\" should not).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Augmented language model",
"sec_num": "4"
},
{
"text": "Of course, traditional word-based text generation models strive to achieve exactly that. However, they could hardly be applied here because the objective of correctly generating a lot of previously unseen forms would be missed. Instead, I developed a sentence generator that utilizes not only the texts, but also the linguistic resources available for Beserman.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Augmented language model",
"sec_num": "4"
},
{
"text": "After a series of sequential improvements, the resulting sentence generator works as follows.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Augmented language model",
"sec_num": "4"
},
{
"text": "First, the sentences from the Beserman corpora are morphologically analyzed and turned into sentence templates. In a template, content words (nouns, verbs, adjectives, adverbs and numerals) are replaced with \"slots\", while the rest (pronouns, postpositions etc.) are left untouched. The idea is that the lemma in a slot can be replaced by another lemma with similar characteristics, while the remaining words should not be replaced with anything else. Certain high-frequency or irregular verbs or adverbs are also not turned into slots, e.g. negative verbs or discourse clitics. Templates where less than one-third of the elements were turned into slots, or that contain fewer than three words, are discarded.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Augmented language model",
"sec_num": "4"
},
{
"text": "A slot contains the inflectional affixes the word used to have, its tags (e.g. \"N,anim\" for \"animate noun\"), as well as the original lemma.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Augmented language model",
"sec_num": "4"
},
{
"text": "Second, the data from the grammatical dictionary of the analyzer is processed. For each item, its lemma, stem(s) and tags (part of speech among them) is loaded. A global frequency dictionary is created. If a lemma is present in the corpora, its total number of occurrences is stored as its frequency; for the remainder of the lemmata, the frequency is set to 1.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Augmented language model",
"sec_num": "4"
},
{
"text": "Third, semantic similarity matrices are created for nouns, verbs and adjectives separately. The semantic similarities are induced from the Russian translations the lemmata have in the Beserman-Russian dictionary.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Augmented language model",
"sec_num": "4"
},
{
"text": "Each translation is stripped of usage notes in brackets and parentheses and of one-letter words.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Augmented language model",
"sec_num": "4"
},
{
"text": "After that, the first word of the remaining string is taken as the Russian equivalent of the Beserman word. The similarities between Russian translations are then calculated with an embedding model ruwikiruscorpora_upos_skipgram_300_2_2019\u2075 trained on the data of Russian National Corpus and Russian Wikipedia (Kutuzov and Kuzmenko, 2017) . The resulting pairwise similarities are then condensed into a JSON file where each Beserman lemma contains its closest semantic neighbors together with the corresponding similarity value. The similarity threshold of 0.39 was set to only keep lemmata which are sufficiently similar to the lemma in question in terms of their distribution. After that, an average lemma contains about 66 semantic neighbors.",
"cite_spans": [
{
"start": 310,
"end": 338,
"text": "(Kutuzov and Kuzmenko, 2017)",
"ref_id": "BIBREF10"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Augmented language model",
"sec_num": "4"
},
{
"text": "After these preparatory steps, the sentence generation starts. A template is chosen at random, after which each slot is filled with a word. If a slot contains multiple ambiguous analyses, one of them is chosen at random with equal probability, apart from several manually defined cases where one of the analyses is much more probable than the others. The original lemma of the slot is looked up in the list of semantic neighbors. If found, its semantic neighbors are used as its possible substitutes. Neighbors whose tags differ from the slot tags (e.g. inanimate nouns instead of animate) are filtered out. A random similarity threshold is chosen, which can further narrow down the list of substitutes. This way, more similar lemmata have a higher chance of ending up on the list of potential substitutes. When the list is ready, a lemma is chosen at random, with probability of each lemma proportional to its frequency in the global frequency list. Its stem is combined with the inflectional affixes in the slot, taking certain morphophonological alternations into account. The resulting word is added to the sentence. Template elements that are not slots are generally used as is, but words from a certain manually defined list can be omitted with a probability of 0.2 (this mostly includes discourse particles).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Augmented language model",
"sec_num": "4"
},
{
"text": "\u2075https://rusvectores.org/en/models/ The sentences generated this way do not always make sense, but many of them at least are not completely ungrammatical, and some actually sound quite acceptable.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Augmented language model",
"sec_num": "4"
},
{
"text": "I did not check how the size of the training dataset affects the quality of the model. However, it is interesting to note that the addition of 2.5 hours of triple headword pronunciations from the sound dictionary apparently did not add to the quality. The results were almost the same when they were omitted from the training set.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results and comparison",
"sec_num": "5"
},
{
"text": "As already mentioned in Section 3, the output of a trained Deepspeech model is filtered with an n-gram model trained on a text corpus, with parameters \u03b1 and \u03b2. I evaluated the model on the test dataset with three kenlm models: based only on the real Beserman sentences (base), and two augmented models trained on real and generated sentences (gen). The first augmented model was trained on 2M additional sentences (about 170K word types), the second, on 10M additional sentences (about 300K word types). The difference between the two augmented models was almost nonexistent. The larger model performed slightly better than the smaller for most parameter values, except in the case of \u03b1 = 1.0; the difference in WER in most cases did not exceed 0.5%. The figures below are given for the larger model.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results and comparison",
"sec_num": "5"
},
{
"text": "The following \u03b1 values were tested: 1.0, 0.9, 0.75, 0.6, 0.4. The values of \u03b2 between 1.0 and 7.0 with the step of 1 were tested. The WER values for \u03b2 \u2265 5.0 were always worse than with lower \u03b2 values and are not represented below.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results and comparison",
"sec_num": "5"
},
{
"text": "One can see that the values obtained with the augmented model are better than the baseline across the board, so the sentence generation has had a positive effect on ASR quality. Also, the augmented model tolerates larger \u03b2 values, whereas the baseline model starts producing too much short words in place of longer words absent from its vocabulary in that case. Nevertheless, the difference is not that large: the best gen value, 51.4, is lower than the best base value, 56.4, only by 5%. The difference in CER is even less pronounced:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results and comparison",
"sec_num": "5"
},
{
"text": "A more in-depth analysis of the data reveals that the effect of LM augmentation is most visible on longer sound files. If only tested on sentences whose ground-truth transcription contained at least 6 words, the best WER value for gen equals 52.1, as \u03b2 = 1 \u03b2 = 2 \u03b2 = 3 \u03b2 = 4 \u03b1 = 1.0 57.3 / 55.6 56.5 / 54.1 57.7 / 52.7 58.8 / 52.4 \u03b1 = 0.9 57.0 / 53.3 56.8 / 52.7 58.4 / 51.6 59.6 / 53.5 \u03b1 = 0.75 56.4 / 52.9 57.2 / 51.9 58.7 / 51.4 60.9 / 53.4 \u03b1 = 0.6 57.1 / 52.9 58.9 / 51.9 60.4 / 53.8 64.9 / 56.3 \u03b1 = 0.4 59.6 / 55.5 62.3 / 56.4 67.0 / 59.8 75.4 / 66.1 Table 1 : WER for base (before slash) and gen (after slash) models with different \u03b1 and \u03b2 values. \u03b2 = 1 \u03b2 = 2 \u03b2 = 3 \u03b2 = 4 \u03b1 = 1.0 35.0 / 34.4 33.9 / 33.3 33.3 / 32.0 32.8 / 30.9 \u03b1 = 0.9 34.0 / 32.5 33.4 / 32.1 33.0 / 30.7 32.7 / 30.2 \u03b1 = 0.75 32.9 / 31.5 32.4 / 30.4 32.0 / 29.7 31.9 / 29.2 \u03b1 = 0.6 32.2 / 30.1 31.5 / 29.9 31.5 / 29.6 31.7 / 29.3 \u03b1 = 0.4 31.6 / 30.1 31.3 / 29.6 31.3 / 30.3 31.8 / 30.8 Table 2 : CER for base (before slash) and gen (after slash) models with different \u03b1 and \u03b2 values. opposed to only 59.5 for base. On short files, however, the added benefit of having plausibly looking n-grams in the corpus stops playing any role. For sentences (or, rather, sentence fragments) that contained at most 3 words, the best WER value for gen equals 60.5, compared to 61.5 for base.",
"cite_spans": [],
"ref_spans": [
{
"start": 556,
"end": 563,
"text": "Table 1",
"ref_id": null
},
{
"start": 959,
"end": 966,
"text": "Table 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "Results and comparison",
"sec_num": "5"
},
{
"text": "As we can see, the LM augmentation did improve the ASR quality, even if marginally. The most important takeaway from this experiment, however, was that using a bilingual dictionary and a Russian model for approximating semantic similarity was a crucial part of the LM augmentation. Without that step, the generated LM did not visibly differ from base, even when lemma frequencies and tags were taken into account.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results and comparison",
"sec_num": "5"
},
{
"text": "Since, to the best of my knowledge, no Deepspeech (or any other) ASR models existed for standard Udmurt when the experiments were conducted, it was impossible to compare ASR quality for Beserman and standard Udmurt.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results and comparison",
"sec_num": "5"
},
{
"text": "There is a famous statement by Frederick Jelinek, made exactly in the context of ASR development, \"Whenever I fire a linguist our system performance improves\". Indeed, contemporary ASR is largely an engineering enterprise and relies on algorithms and large amounts of data rather than on any linguistic insights. Still, if there is not enough data, can linguistic resources -resources created by linguists and for linguists -be of any help at all? The results of the experiments with the Beserman data are not conclusive. On the one hand, linguistic interven-tion did improve the ASR results, lowering WER by 5% and even more so in the case of longer sentences. Linguistic resources, such as the rule-based analyzer turned into a generator, and the Beserman-Russian dictionary, as well as the corpus of usage examples, seemed indispensable in the process. On the other hand, the result is yet another experimental model for a low-resource language with suboptimal performance, which might be not good enough even for auxiliary uses. In order to make it usable, one would still have to either add more data or change the algorithm (e.g. (Baevski et al., 2021) report results for comparable amounts of Tatar and Kyrgyz data that almost look like magic). It would be interesting to see if the \"linguistic\" LM augmentation adds anything in that case.",
"cite_spans": [
{
"start": 1136,
"end": 1158,
"text": "(Baevski et al., 2021)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "6"
}
],
"back_matter": [
{
"text": "This research was funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) -Project ID 428175960.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgments",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "German endto-end speech recognition based on DeepSpeech",
"authors": [
{
"first": "Aashish",
"middle": [],
"last": "Agarwal",
"suffix": ""
},
{
"first": "Torsten",
"middle": [],
"last": "Zesch",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Aashish Agarwal and Torsten Zesch. 2019. German end- to-end speech recognition based on DeepSpeech. In KONVENS.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Corpus of usage examples: What is it good for?",
"authors": [
{
"first": "Timofey",
"middle": [],
"last": "Arkhangelskiy",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 3rd Workshop on the Use of Computational Methods in the Study of Endangered Languages",
"volume": "1",
"issue": "",
"pages": "56--63",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Timofey Arkhangelskiy. 2019. Corpus of usage exam- ples: What is it good for? In Proceedings of the 3rd Workshop on the Use of Computational Methods in the Study of Endangered Languages Volume 1 (Pa- pers), pages 56-63, Honolulu. Association for Com- putational Linguistics.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Unsupervised speech recognition",
"authors": [
{
"first": "Alexei",
"middle": [],
"last": "Baevski",
"suffix": ""
},
{
"first": "Wei-Ning",
"middle": [],
"last": "Hsu",
"suffix": ""
},
{
"first": "Alexis",
"middle": [],
"last": "Conneau",
"suffix": ""
},
{
"first": "Michael",
"middle": [],
"last": "Auli",
"suffix": ""
}
],
"year": 2021,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Alexei Baevski, Wei-Ning Hsu, Alexis Conneau, and Michael Auli. 2021. Unsupervised speech recogni- tion.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Cg-3 -beyond classical constraint grammar",
"authors": [
{
"first": "Eckhard",
"middle": [],
"last": "Bick",
"suffix": ""
},
{
"first": "Tino",
"middle": [],
"last": "Didriksen",
"suffix": ""
}
],
"year": 2015,
"venue": "Proceedings of the 20th Nordic Conference of Computational Linguistics",
"volume": "",
"issue": "",
"pages": "31--39",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Eckhard Bick and Tino Didriksen. 2015. Cg-3 -be- yond classical constraint grammar. In Proceedings of the 20th Nordic Conference of Computational Linguis- tics, NODALIDA, pages 31-39. Link\u00f6ping University Electronic Press, Link\u00f6pings universitet.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Out-ofvocabulary word recovery using fst-based subword unit clustering in a hybrid asr system",
"authors": [
{
"first": "Ekaterina",
"middle": [],
"last": "Egorova",
"suffix": ""
},
{
"first": "Luk\u00e1\u0161",
"middle": [],
"last": "Burget",
"suffix": ""
}
],
"year": 2018,
"venue": "2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)",
"volume": "",
"issue": "",
"pages": "5919--5923",
"other_ids": {
"DOI": [
"10.1109/ICASSP.2018.8462221"
]
},
"num": null,
"urls": [],
"raw_text": "Ekaterina Egorova and Luk\u00e1\u0161 Burget. 2018. Out-of- vocabulary word recovery using fst-based subword unit clustering in a hybrid asr system. In 2018 IEEE International Conference on Acoustics, Speech and Sig- nal Processing (ICASSP), pages 5919-5923.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Deep Speech: Scaling up end-to-end speech recognition. arXiv e-prints",
"authors": [
{
"first": "Awni",
"middle": [],
"last": "Hannun",
"suffix": ""
},
{
"first": "Carl",
"middle": [],
"last": "Case",
"suffix": ""
},
{
"first": "Jared",
"middle": [],
"last": "Casper",
"suffix": ""
},
{
"first": "Bryan",
"middle": [],
"last": "Catanzaro",
"suffix": ""
},
{
"first": "Greg",
"middle": [],
"last": "Diamos",
"suffix": ""
},
{
"first": "Erich",
"middle": [],
"last": "Elsen",
"suffix": ""
},
{
"first": "Ryan",
"middle": [],
"last": "Prenger",
"suffix": ""
},
{
"first": "Sanjeev",
"middle": [],
"last": "Satheesh",
"suffix": ""
},
{
"first": "Shubho",
"middle": [],
"last": "Sengupta",
"suffix": ""
},
{
"first": "Adam",
"middle": [],
"last": "Coates",
"suffix": ""
},
{
"first": "Andrew",
"middle": [
"Y"
],
"last": "Ng",
"suffix": ""
}
],
"year": 2014,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1412.5567"
]
},
"num": null,
"urls": [],
"raw_text": "Awni Hannun, Carl Case, Jared Casper, Bryan Catan- zaro, Greg Diamos, Erich Elsen, Ryan Prenger, San- jeev Satheesh, Shubho Sengupta, Adam Coates, and Andrew Y. Ng. 2014. Deep Speech: Scaling up end-to-end speech recognition. arXiv e-prints, page arXiv:1412.5567.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "KenLM: Faster and smaller language model queries",
"authors": [
{
"first": "Kenneth",
"middle": [],
"last": "Heafield",
"suffix": ""
}
],
"year": 2011,
"venue": "Proceedings of the Sixth Workshop on Statistical Machine Translation",
"volume": "",
"issue": "",
"pages": "187--197",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kenneth Heafield. 2011. KenLM: Faster and smaller language model queries. In Proceedings of the Sixth Workshop on Statistical Machine Translation, pages 187-197, Edinburgh, Scotland. Association for Com- putational Linguistics.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Improving the language model for low-resource ASR with online text corpora",
"authors": [
{
"first": "Nils",
"middle": [],
"last": "Hjortnaes",
"suffix": ""
},
{
"first": "Timofey",
"middle": [],
"last": "Arkhangelskiy",
"suffix": ""
},
{
"first": "Niko",
"middle": [],
"last": "Partanen",
"suffix": ""
},
{
"first": "Michael",
"middle": [],
"last": "Rie\u00dfler",
"suffix": ""
},
{
"first": "Francis",
"middle": [],
"last": "Tyers",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 1st Joint Workshop on Spoken Language Technologies for Under-resourced languages (SLTU) and Collaboration and Computing for Under-Resourced Languages (CCURL)",
"volume": "",
"issue": "",
"pages": "336--341",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Nils Hjortnaes, Timofey Arkhangelskiy, Niko Partanen, Michael Rie\u00dfler, and Francis Tyers. 2020a. Im- proving the language model for low-resource ASR with online text corpora. In Proceedings of the 1st Joint Workshop on Spoken Language Technologies for Under-resourced languages (SLTU) and Collabora- tion and Computing for Under-Resourced Languages (CCURL), pages 336-341, Marseille, France. Euro- pean Language Resources association.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Towards a speech recognizer for Komi, an endangered and low-resource uralic language",
"authors": [
{
"first": "Nils",
"middle": [],
"last": "Hjortnaes",
"suffix": ""
},
{
"first": "Niko",
"middle": [],
"last": "Partanen",
"suffix": ""
},
{
"first": "Michael",
"middle": [],
"last": "Rie\u00dfler",
"suffix": ""
},
{
"first": "Francis",
"middle": [
"M"
],
"last": "Tyers",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the Sixth International Workshop on Computational Linguistics of Uralic Languages",
"volume": "",
"issue": "",
"pages": "31--37",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Nils Hjortnaes, Niko Partanen, Michael Rie\u00dfler, and Francis M. Tyers. 2020b. Towards a speech rec- ognizer for Komi, an endangered and low-resource uralic language. In Proceedings of the Sixth Inter- national Workshop on Computational Linguistics of Uralic Languages, pages 31-37, Wien, Austria. As- sociation for Computational Linguistics.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Constraint grammar as a framework for parsing unrestricted text",
"authors": [
{
"first": "Fred",
"middle": [],
"last": "Karlsson",
"suffix": ""
}
],
"year": 1990,
"venue": "Proceedings of the 13th International Conference of Computational Linguistics",
"volume": "3",
"issue": "",
"pages": "168--173",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Fred Karlsson. 1990. Constraint grammar as a frame- work for parsing unrestricted text. In Proceedings of the 13th International Conference of Computational Linguistics, volume 3, pages 168-173, Helsinki.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "We-bVectors: A Toolkit for Building Web Interfaces for Vector Semantic Models",
"authors": [
{
"first": "Andrey",
"middle": [],
"last": "Kutuzov",
"suffix": ""
},
{
"first": "Elizaveta",
"middle": [],
"last": "Kuzmenko",
"suffix": ""
}
],
"year": 2017,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"DOI": [
"10.1007/978-3-319-52920-2{_}15"
]
},
"num": null,
"urls": [],
"raw_text": "Andrey Kutuzov and Elizaveta Kuzmenko. 2017. We- bVectors: A Toolkit for Building Web Interfaces for Vector Semantic Models. Springer International Pub- lishing, Cham.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "New baseline in automatic speech recognition for Northern S\u00e1mi",
"authors": [
{
"first": "Juho",
"middle": [],
"last": "Leinonen",
"suffix": ""
},
{
"first": "Peter",
"middle": [],
"last": "Smit",
"suffix": ""
},
{
"first": "Sami",
"middle": [],
"last": "Virpioja",
"suffix": ""
},
{
"first": "Mikko",
"middle": [],
"last": "Kurimo",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of the Fourth International Workshop on Computational Linguistics of Uralic Languages",
"volume": "",
"issue": "",
"pages": "87--97",
"other_ids": {
"DOI": [
"10.18653/v1/W18-0208"
]
},
"num": null,
"urls": [],
"raw_text": "Juho Leinonen, Peter Smit, Sami Virpioja, and Mikko Kurimo. 2018. New baseline in automatic speech recognition for Northern S\u00e1mi. In Proceedings of the Fourth International Workshop on Computational Lin- guistics of Uralic Languages, pages 87-97, Helsinki, Finland. Association for Computational Linguistics.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Avtomaticheskoe vyravnivanie slovarej literaturnogo udmurtskogo i jazyka i besermjanskogo dialekta",
"authors": [
{
"first": "Eugenia",
"middle": [],
"last": "Miller",
"suffix": ""
}
],
"year": 2017,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Eugenia Miller. 2017. Avtomaticheskoe vyravnivanie slovarej literaturnogo udmurtskogo i jazyka i beser- mjanskogo dialekta [Automatic alignment of literary Udmurt and Beserman dictionaries].",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Elektronnaja pismennost narodov Rossijskoj Federacii: opyt, problemy i perspektivy",
"authors": [],
"year": null,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "In Elektronnaja pismennost narodov Rossijskoj Federacii: opyt, prob- lemy i perspektivy [Electronic writing of the peoples of the Russian Federation: Experience, challenges and perspectives], Syktyvkar, Russia. KRAGSiU.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "SpecAugment: A Simple Data Augmentation Method for Automatic Speech Recognition",
"authors": [
{
"first": "Daniel",
"middle": [
"S"
],
"last": "Park",
"suffix": ""
},
{
"first": "William",
"middle": [],
"last": "Chan",
"suffix": ""
},
{
"first": "Yu",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "Chung-Cheng",
"middle": [],
"last": "Chiu",
"suffix": ""
},
{
"first": "Barret",
"middle": [],
"last": "Zoph",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Ekin",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Cubuk",
"suffix": ""
},
{
"first": "V",
"middle": [],
"last": "Quoc",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Le",
"suffix": ""
}
],
"year": 2019,
"venue": "Proc. Interspeech",
"volume": "",
"issue": "",
"pages": "2613--2617",
"other_ids": {
"DOI": [
"10.21437/Interspeech.2019-2680"
]
},
"num": null,
"urls": [],
"raw_text": "Daniel S. Park, William Chan, Yu Zhang, Chung-Cheng Chiu, Barret Zoph, Ekin D. Cubuk, and Quoc V. Le. 2019. SpecAugment: A Simple Data Augmentation Method for Automatic Speech Recognition. In Proc. Interspeech 2019, pages 2613-2617.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Analyzing ASR pretraining for lowresource speech-to-text translation",
"authors": [
{
"first": "C",
"middle": [],
"last": "Mihaela",
"suffix": ""
},
{
"first": "Sameer",
"middle": [],
"last": "Stoian",
"suffix": ""
},
{
"first": "Sharon",
"middle": [],
"last": "Bansal",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Goldwater",
"suffix": ""
}
],
"year": 2020,
"venue": "ICASSP 2020 -2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)",
"volume": "",
"issue": "",
"pages": "7909--7913",
"other_ids": {
"DOI": [
"10.1109/ICASSP40776.2020.9053847"
]
},
"num": null,
"urls": [],
"raw_text": "Mihaela C. Stoian, Sameer Bansal, and Sharon Gold- water. 2020. Analyzing ASR pretraining for low- resource speech-to-text translation. In ICASSP 2020 -2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7909- 7913.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Data augmentation, feature combination, and multilingual neural networks to improve asr and kws performance for low-resource languages",
"authors": [
{
"first": "Zolt\u00e1n",
"middle": [],
"last": "T\u00fcske",
"suffix": ""
},
{
"first": "Pavel",
"middle": [],
"last": "Golik",
"suffix": ""
},
{
"first": "David",
"middle": [],
"last": "Nolden",
"suffix": ""
},
{
"first": "Ralf",
"middle": [],
"last": "Schl\u00fcter",
"suffix": ""
},
{
"first": "Hermann",
"middle": [],
"last": "Ney",
"suffix": ""
}
],
"year": 2014,
"venue": "INTERSPEECH-2014",
"volume": "",
"issue": "",
"pages": "1420--1424",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Zolt\u00e1n T\u00fcske, Pavel Golik, David Nolden, Ralf Schl\u00fcter, and Hermann Ney. 2014. Data augmentation, feature combination, and multilingual neural networks to im- prove asr and kws performance for low-resource lan- guages. In INTERSPEECH-2014, pages 1420-1424.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "What shall we do with an hour of data? Speech recognition for the un-and under-served languages of Common Voice",
"authors": [
{
"first": "M",
"middle": [],
"last": "Francis",
"suffix": ""
},
{
"first": "Josh",
"middle": [],
"last": "Tyers",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Meyer",
"suffix": ""
}
],
"year": 2021,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Francis M. Tyers and Josh Meyer. 2021. What shall we do with an hour of data? Speech recognition for the un-and under-served languages of Common Voice.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "\u0422\u0435\u0437\u0430\u0443\u0440\u0443\u0441 \u0431\u0435\u0441\u0435\u0440\u043c\u044f\u043d\u0441\u043a\u043e\u0433\u043e \u043d\u0430\u0440\u0435\u0447\u0438\u044f: \u0418\u043c\u0435\u043d\u0430 \u0438 \u0441\u043b\u0443\u0436\u0435\u0431\u043d\u044b\u0435 \u0447\u0430\u0441\u0442\u0438 \u0440\u0435\u0447\u0438 (\u0433\u043e\u0432\u043e\u0440 \u0434\u0435\u0440\u0435\u0432\u043d\u0438 \u0428\u0430\u043c\u0430\u0440\u0434\u0430\u043d) [Thesaurus of the Beserman dialect: Nouns and auxiliary parts of speech",
"authors": [
{
"first": "Maria",
"middle": [],
"last": "Usacheva",
"suffix": ""
},
{
"first": "Timofey",
"middle": [],
"last": "Arkhangelskiy",
"suffix": ""
},
{
"first": "Olga",
"middle": [],
"last": "Biryuk",
"suffix": ""
},
{
"first": "Vladimir",
"middle": [],
"last": "Ivanov",
"suffix": ""
},
{
"first": "Ruslan",
"middle": [],
"last": "Idrisov",
"suffix": ""
}
],
"year": 2017,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Maria Usacheva, Timofey Arkhangelskiy, Olga Biryuk, Vladimir Ivanov, and Ruslan Idrisov, editors. 2017. \u0422\u0435\u0437\u0430\u0443\u0440\u0443\u0441 \u0431\u0435\u0441\u0435\u0440\u043c\u044f\u043d\u0441\u043a\u043e\u0433\u043e \u043d\u0430\u0440\u0435\u0447\u0438\u044f: \u0418\u043c\u0435\u043d\u0430 \u0438 \u0441\u043b\u0443\u0436\u0435\u0431\u043d\u044b\u0435 \u0447\u0430\u0441\u0442\u0438 \u0440\u0435\u0447\u0438 (\u0433\u043e\u0432\u043e\u0440 \u0434\u0435\u0440\u0435\u0432\u043d\u0438 \u0428\u0430\u043c\u0430\u0440\u0434\u0430\u043d) [Thesaurus of the Beserman dialect: Nouns and aux- iliary parts of speech (Shamardan village variety)].",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "Izdatelskie resheniya",
"authors": [],
"year": null,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Izdatelskie resheniya, Moscow.",
"links": null
}
},
"ref_entries": {}
}
} |