File size: 66,596 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 |
{
"paper_id": "O13-1018",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T08:04:04.863877Z"
},
"title": "Multilingual Acoustic Model Splitting and Merging by Latent Dirichlet Allocation",
"authors": [
{
"first": "Jui-Feng",
"middle": [],
"last": "\u8449\u745e\u5cf0",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "",
"middle": [],
"last": "Yeh",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Sheng-Feng",
"middle": [],
"last": "\u674e\u52dd\u8c50",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "",
"middle": [],
"last": "Li",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Shi-Sheng",
"middle": [],
"last": "\u8a31\u5e0c\u8056",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "",
"middle": [],
"last": "Shiu",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "David",
"middle": [
"M"
],
"last": "\u500b\u968e\u5c64\u5f0f\u7684\u6578\u5b78\u6a21\u578b\uff0c\u65e9\u671f\u662f\u7531",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "",
"middle": [],
"last": "Blei",
"suffix": "",
"affiliation": {},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "To avoid the confusion of phonetic acoustic models between different languages is one of the most challenges in multilingual speech recognition. We proposed the method based on Latent Dirichlet Allocation to avoid the confusion of phonetic acoustic models between different languages. We split phonetic acoustic models based on tri-phone. And merging the group that selected by Latent Dirichlet Allocation Detector to solve pronunciation variants problems between different languages. This paper has three parts. First part is introduced the Pronunciation Event and Articulatory Features. Second part is about Latent Dirichlet Allocation and the acoustic model selecting method using Latent Dirichlet Allocation. Latent Dirichlet Allocation is a Hierarchical math model that proposed by David M. Blei at 2003. It is often used on documents classification and document generation. The structure of LDA is also suitable for speech recognition and nature language processing. The final is experiment result and verification the method we proposed.",
"pdf_parse": {
"paper_id": "O13-1018",
"_pdf_hash": "",
"abstract": [
{
"text": "To avoid the confusion of phonetic acoustic models between different languages is one of the most challenges in multilingual speech recognition. We proposed the method based on Latent Dirichlet Allocation to avoid the confusion of phonetic acoustic models between different languages. We split phonetic acoustic models based on tri-phone. And merging the group that selected by Latent Dirichlet Allocation Detector to solve pronunciation variants problems between different languages. This paper has three parts. First part is introduced the Pronunciation Event and Articulatory Features. Second part is about Latent Dirichlet Allocation and the acoustic model selecting method using Latent Dirichlet Allocation. Latent Dirichlet Allocation is a Hierarchical math model that proposed by David M. Blei at 2003. It is often used on documents classification and document generation. The structure of LDA is also suitable for speech recognition and nature language processing. The final is experiment result and verification the method we proposed.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "(Internal Phone Alphabet ,IPA)\u4e4b\u5b9a\u7fa9\u4f86\u9032\u884c\u767c\u97f3\u4e8b\u4ef6\u5206\u985e\u3002\u7d93\u904e\u5c0d\u53f0\u8a9e\u97f3\u6a19(ForPA)\u3001\u570b\u8a9e\u97f3\u6a19\u548c\u82f1\u6587 (KK \u97f3\u6a19)\u5c0d\u61c9\u5f8c\uff0c\u4f7f\u7528\u5230\u7684\u767c\u97f3\u65b9\u6cd5(Manner)\u7e3d\u5171\u6709\u516d\u985e\uff1a\u9f3b\u97f3(Nasal)\u3001\u585e\u97f3(Stop)\u3001 \u6469\u64e6\u97f3(Fricative)\u3001\u8fd1\u97f3(Approximant)\u3001\u585e\u64e6\u97f3(Affricate)\u548c\u986b\u97f3(Trill)\u3002\u800c\u4f7f\u7528\u7684\u767c\u97f3\u90e8 \u4f4d(Place)\u7e3d\u5171\u6709\u5341\u4e09\u985e\uff1a\u96d9\u5507\u97f3(Bilabial)\u3001\u5507\u9f52\u97f3(Labio-dental)\u3001\u5507\u8edf\u984e(Labio-velar)\u3001 \u9f52\u97f3(Dental) \u3001\u9f66\u97f3(Alveolar) \u3001\u9f66\u5f8c\u97f3(Post-alv) \u3001\u6372\u820c(Retroflex) \u3001\u9f66\u984e\u97f3(Alveolo-palatal) \u3001 \u9f66\u786c\u984e(Palato-alveolar)\u3001\u786c\u984e\u97f3(Patatal)\u3001\u8edf\u984e(Velar)\u3001\u5c0f\u820c\u97f3(Uvular)\u548c\u8072\u9580\u97f3(Glottal)\u3002 \u56db\u3001\u6f5b\u85cf\u72c4\u5f0f\u5206\u4f48\u5075\u6e2c\u5668 (\u4e00)\u6982\u8ff0 \u6f5b\u85cf\u72c4\u5f0f\u5206\u4f48\u662f\u4e00\u7a2e\u968e\u5c64\u5316\u7684\u751f\u6210\u6a5f\u7387\u6a21\u578b\uff0c\u662f\u7531 David M. Blei[1]\u65bc",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": ", \u03b2) = \u222c p(\u03b8 M |\u03b1 M ) p(\u03b8 P |\u03b1 P ) {\u220f p(Z M |\u03b8 M )p(Z P |\u03b8 P )p(w|Z M , Z P , L, \u03b2) N n=1 } d\u03b8 M d\u03b8 P (\u5f0f 4.1) \u5716\u4e8c\u3001\u6f5b\u85cf\u72c4\u5f0f\u5206\u4f48\u5075\u6e2c\u5668\u4e4b\u5716\u578b\u8868\u793a\u6cd5 \u8868\u4e00\u3001\u6f5b\u85cf\u72c4\u5f0f\u5206\u4f48\u5075\u6e2c\u5668\u4e4b\u53c3\u6578\u4ee3\u8868\u610f\u7fa9 \u7b26\u865f \u63cf\u8ff0 \u03b1M K \u5411\u91cf\uff0c\u767c\u97f3\u65b9\u6cd5\u7684 dirichlet \u5206\u5e03 \u03b1P K \u5411\u91cf\uff0c\u767c\u97f3\u90e8\u4f4d\u7684 dirichlet",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Latent Dirichlet Allocation",
"authors": [
{
"first": "David",
"middle": [
"M"
],
"last": "Blei",
"suffix": ""
},
{
"first": "Andrew",
"middle": [
"Y"
],
"last": "Ng",
"suffix": ""
},
{
"first": "Michael",
"middle": [
"I"
],
"last": "Jordan",
"suffix": ""
}
],
"year": 2003,
"venue": "Journal of Machine Learning Research",
"volume": "3",
"issue": "",
"pages": "993--1022",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "David M. Blei, Andrew Y. Ng, Michael I. Jordan, Latent Dirichlet Allocation, Journal of Machine Learning Research 3, pp.993-1022, 2003.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Automatic Evaluation of English Pronunciation Based on Speech Recognition Techniques",
"authors": [
{
"first": "H",
"middle": [],
"last": "Hamada",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Miki",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Nakatsu",
"suffix": ""
}
],
"year": 1993,
"venue": "IEICE Trans. Inf. and Sys",
"volume": "",
"issue": "3",
"pages": "352--359",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Hamada, H., S. Miki, and R. Nakatsu. Automatic Evaluation of English Pronunciation Based on Speech Recognition Techniques, IEICE Trans. Inf. and Sys. 1993 E76-D(3):352-359.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Pronunciation Scoring of Foreign Language Student Speech In ICSLP' 96",
"authors": [
{
"first": "L",
"middle": [],
"last": "Neumeyer",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Franco",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Weintraub",
"suffix": ""
},
{
"first": "P",
"middle": [],
"last": "Price",
"suffix": ""
}
],
"year": null,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Neumeyer, L., H. Franco, M. Weintraub, and P. Price. Pronunciation Scoring of Foreign Language Student Speech In ICSLP' 96. Philadelphia, USA, Oct.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Automatic Detection of Mispronunciation for Language Instruction",
"authors": [
{
"first": "O",
"middle": [],
"last": "Ronen",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Neumeyer",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Franco",
"suffix": ""
}
],
"year": null,
"venue": "Proceedings Eurospeech",
"volume": "97",
"issue": "",
"pages": "649--652",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ronen, O., Neumeyer, L. and Franco, H. Automatic Detection of Mispronunciation for Language Instruction, Proceedings Eurospeech 97, Rhodes, Greece, 649-652.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Automatic Detection of Phone-Level Mispronunciation for Language Learning",
"authors": [
{
"first": "H",
"middle": [],
"last": "Franco",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Neumeyer",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Ramos",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Bratt",
"suffix": ""
}
],
"year": null,
"venue": "Proceedings Eurospeech '99",
"volume": "",
"issue": "",
"pages": "851--854",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Franco, H., Neumeyer, L., Ramos, M., and Bratt, H. Automatic Detection of Phone-Level Mispronunciation for Language Learning, Proceedings Eurospeech '99, Budapest, Hungary, 851-854.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "EM Training of Finite-State Transducers and its Application to Pronunciation Modeling",
"authors": [
{
"first": "H",
"middle": [],
"last": "Shu",
"suffix": ""
},
{
"first": "I",
"middle": [
"L"
],
"last": "Hetherington",
"suffix": ""
}
],
"year": 2002,
"venue": "Proc. ICSLP",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "H. Shu and I. L. Hetherington, EM Training of Finite-State Transducers and its Application to Pronunciation Modeling, Proc. ICSLP, Denver, CO, September 2002.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "A Vector Space Modeling Approach to Spoken Language Identification\u2016, Audio, Speech, and Language Processing",
"authors": [
{
"first": "H",
"middle": [],
"last": "Li",
"suffix": ""
},
{
"first": "B",
"middle": [],
"last": "Ma",
"suffix": ""
},
{
"first": "C",
"middle": [
"H"
],
"last": "Lee",
"suffix": ""
}
],
"year": 2007,
"venue": "IEEE Transactions",
"volume": "15",
"issue": "1",
"pages": "271--284",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "H. Li, B. Ma, and C.H. Lee. A Vector Space Modeling Approach to Spoken Language Identification\u2016, Audio, Speech, and Language Processing, IEEE Transactions on vol. 15, NO. 1, JANUARY, pp 271-284, 2007.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Experiments on Cross-Language Attribute Detection and Phone Recognition With Minimal Target-Specific Training Data",
"authors": [
{
"first": "Dau-Cheng",
"middle": [],
"last": "Sabato Marco Siniscalchi",
"suffix": ""
},
{
"first": "Torbj\u00f8rn",
"middle": [],
"last": "Lyu",
"suffix": ""
},
{
"first": "Chin-Hui",
"middle": [],
"last": "Svendsen",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Lee",
"suffix": ""
}
],
"year": 2012,
"venue": "IEEE TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING",
"volume": "20",
"issue": "3",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Sabato Marco Siniscalchi, Dau-Cheng Lyu, Torbj\u00f8rn Svendsen, Chin-Hui Lee, Experiments on Cross-Language Attribute Detection and Phone Recognition With Minimal Target-Specific Training Data, IEEE TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING, VOL. 20, NO. 3, MARCH 2012.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Proceedings of the Twenty-Fifth Conference on Computational Linguistics and Speech Processing",
"authors": [],
"year": null,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Proceedings of the Twenty-Fifth Conference on Computational Linguistics and Speech Processing (ROCLING 2013)",
"links": null
}
},
"ref_entries": {
"TABREF0": {
"html": null,
"num": null,
"content": "<table><tr><td>\u4e00\u3001\u7dd2\u8ad6 (\u4e00)\u7814\u7a76\u52d5\u6a5f \u7531\u65bc\u76ee\u524d\u7db2\u8def\u8207\u4ea4\u901a\u7684\u767c\u9054\uff0c\u4f7f\u5f97\u5168\u7403\u5316\u6210\u70ba\u5fc5\u7136\u7684\u8da8\u52e2\uff0c\u518d\u52a0\u4e0a\u53f0\u7063\u672c\u8eab\u5c31\u662f\u5c6c \u65bc\u4e00\u500b\u591a\u65cf\u7fa4\u793e\u6703\uff0c\u56e0\u6b64\u5728\u65e5\u5e38\u751f\u6d3b\u74b0\u5883\u4e2d\u63a5\u89f8\u5230\u5176\u4ed6\u8a9e\u8a00\u7684\u6a5f\u6703\u4e5f\u65e5\u9032\u589e\u52a0\u3002\u9664\u4e86\u5e73 \u5e38\u6240\u807d\u5230\u548c\u770b\u5230\u7684\u8cc7\u8a0a\u542b\u6709\u5176\u4ed6\u8a9e\u8a00\u5916\uff0c\u73fe\u5728\u9023\u5e73\u6642\u5c0d\u8a71\u4e5f\u6703\u7d93\u5e38\u542b\u6709\u82f1\u6587\u3001\u53f0\u8a9e\u751a\u81f3 \u662f\u65e5\u8a9e\u7684\u60c5\u6cc1\u7522\u751f\u3002\u4e5f\u56e0\u6b64\u8a9e\u97f3\u8fa8\u8b58\u6210\u70ba\u8fd1\u5e74\u4f86\u71b1\u9580\u7684\u79d1\u5b78\u7814\u7a76\u9805\u76ee\u4e4b\u4e00\uff0c\u800c\u4e14\u5e02\u9762\u4e0a \u4e5f\u6709\u8a31\u591a\u76f8\u95dc\u61c9\u7528\u985e\u7522\u54c1\uff0c\u4f8b\u5982\uff1aGoogle Android \u5e73\u53f0\u7684 Google Voice Search\u3001Apple \u7684 Siri \u8a9e\u97f3\u52a9\u7406\u2026\u7b49\uff0c\u4f46\u76ee\u524d\u9019\u4e9b\u5e73\u53f0\u90fd\u53ea\u80fd\u5c0d\u55ae\u4e00\u7a2e\u8a9e\u8a00\u9032\u884c\u8fa8\u8b58\uff0c\u56e0\u6b64\u96e3\u4ee5\u61c9\u4ed8\u65e5\u9032 \u589e\u52a0\u7684\u591a\u8a9e\u74b0\u5883\uff0c\u56e0\u6b64\u9700\u8981\u4e00\u7a2e\u53ef\u4ee5\u8fa8\u8b58\u591a\u7a2e\u8a9e\u8a00\u7684\u8fa8\u8b58\u5668\u3002 \u65e9\u671f\u7684\u591a\u8a9e\u8fa8\u8b58\u5668\u7b2c\u4e00\u6b65\u662f\u5148\u5c07\u8a9e\u8a00\u985e\u578b\u8fa8\u8b58\u51fa\u4f86\u5f8c\uff0c\u518d\u5c07\u8f38\u5165\u7684\u8a9e\u97f3\u8a0a\u865f\u9001\u9032 \u5c0d\u61c9\u7684\u8fa8\u8b58\u5668\u8fa8\u8b58\u3002\u4f46\u662f\u7531\u65bc\u4e0d\u540c\u8a9e\u8a00\u7684\u97f3\u6a19\u96c6\u5408\u4e26\u4e0d\u5b8c\u5168\u76f8\u540c\uff0c\u4e14\u591a\u6578\u7684\u8a9e\u97f3\u8fa8\u8b58\u5668 \u662f\u91dd\u5c0d\u7279\u5b9a\u8a9e\u8a00\u7684\u97f3\u6a19\u96c6\u5408\u9032\u884c\u6a21\u578b\u4e4b\u8a13\u7df4\uff0c\u56e0\u6b64\u82e5\u662f\u5728\u7b2c\u4e00\u968e\u6bb5\u7684\u8a9e\u8a00\u7a2e\u985e\u8fa8\u8b58\u932f\u8aa4\uff0c \u800c\u5c07\u8f38\u5165\u8a9e\u53e5\u9001\u5165\u6240\u5c0d\u61c9\u7684\u932f\u8aa4\u8a9e\u8a00\u8fa8\u8b58\u5668\u4e2d\uff0c\u5247\u6240\u7522\u751f\u7684\u7d50\u679c\u5c07\u6703\u662f\u5e7e\u4e4e\u5b8c\u5168\u4e0d\u7b26\u9810 \u671f\u3002\u800c\u5176\u7d9c\u5408\u8fa8\u8b58\u6b63\u78ba\u7387\u4e5f\u6703\u53d7\u5230\u5169\u500b\u5143\u4ef6\u7684\u932f\u8aa4\u758a\u52a0\u800c\u964d\u4f4e\u3002\u70ba\u4e86\u89e3\u6c7a\u6b64\u554f\u984c\uff0c\u5c07\u524d \u5f8c\u6574\u5408\u5728\u4e00\u8d77\u8a2d\u8a08\u4e00\u500b\u8fa8\u8b58\u5668\u53ef\u4ee5\u76f4\u63a5\u8fa8\u8b58\u591a\u570b\u8a9e\u8a00\u7684\u67b6\u69cb\u5247\u88ab\u63a1\u7528\uff0c\u4ee5\u6e1b\u5c11\u932f\u8aa4\u7387\u7684 \u7d2f\u52a0\u3002\u8981\u5c07\u591a\u500b\u8a9e\u8a00\u6574\u5408\u5728\u540c\u4e00\u8fa8\u8b58\u5668\u5167\u5927\u81f4\u4e0a\u53ef\u5206\u70ba\u4e09\u7a2e\u65b9\u6cd5\uff0c\u7b2c\u4e00\u7a2e\u662f\u5c07\u5404\u500b\u55ae\u4e00 \u8a9e\u8a00\u7684\u97f3\u6a19\u96c6\u5408\u4f75\u6210\u5171\u540c\u97f3\u6a19\u96c6\u5408\uff0c\u4e26\u4ee5\u6b64\u70ba\u4f9d\u64da\u4f86\u5efa\u7acb\u591a\u8a9e\u7684\u8fa8\u8b58\u5668\u3002\u7b2c\u4e8c\u7a2e\u662f\u4f7f\u7528 \u5c08\u5bb6\u77e5\u8b58\u6240\u5efa\u7acb\u7684\u8de8\u8a9e\u8a00\u97f3\u6a19\u96c6\u5408\u4f86\u5408\u4f75\u4e0d\u540c\u8a9e\u8a00\u7684\u97f3\u6a19\uff0c\u76ee\u524d\u570b\u969b\u97f3\u6a19\u96c6\u5408\u6709\u570b\u969b\u97f3 \u4e9b\u4e0d\u540c\u7684\u60c5\u6cc1\u3002\u6216\u8005\u662f\u4e0d\u540c\u8a9e\u8a00\u7684\u4e0d\u540c\u7684\u97f3\u6a19\uff0c\u4f46\u662f\u5176\u767c\u97f3\u6a21\u5f0f\u537b\u6975\u70ba\u76f8\u4f3c\u3002\u70ba\u4e86\u89e3\u6c7a \u4e0a\u8ff0\u4e4b\u554f\u984c\uff0c\u56e0\u6b64\u5c07\u540c\u97f3\u6a19\u4f46\u767c\u97f3\u4e0d\u540c\u7684\u8072\u5b78\u6a21\u578b\u9032\u884c\u62c6\u5206\uff0c\u800c\u97f3\u6a19\u4e0d\u540c\u767c\u97f3\u985e\u4f3c\u7684\u8072 \u5b78\u6a21\u578b\u9032\u884c\u5408\u4f75\uff0c\u4ee5\u6e1b\u5c11\u6df7\u6dc6\u7684\u60c5\u6cc1\u3002 (\u4e8c)\u76f8\u95dc\u7814\u7a76 Haizhou Li, Bin Ma, and Chin-Hui Lee \u65bc\u671f\u520a\u4e0a\u767c\u8868\u7684\u7814\u7a76\u591a\u8a9e\u8fa8\u8b58[7]\uff0c\u63d0\u51fa\u4e86\u65b0 \u7684\u8fa8\u8b58\u55ae\u5143\u69cb\u60f3\uff0c\u4e0d\u518d\u4ee5\u97f3\u6a19\u70ba\u55ae\u5143\u800c\u662f\u7528\u4eba\u985e\u5be6\u969b\u767c\u97f3\u7684\u65b9\u6cd5\u4f86\u505a\u70ba\u55ae\u4f4d\u3002\u6b64\u65b9\u6cd5\u662f \u5f9e\u8a0a\u865f\u9762\u5c0b\u627e\u5404\u7a2e\u7279\u5fb5\uff0c\u4e26\u4ee5\u9019\u4e9b\u7279\u5fb5\u5efa\u7acb\u51fa\u5404\u7a2e\u767c\u97f3\u4e8b\u4ef6(Pronunciation Event)\u7684\u8fa8\u8b58 \u5668\uff0c\u4e26\u5c07\u6240\u6709\u767c\u97f3\u4e8b\u4ef6\u8fa8\u8b58\u5668\u7d50\u679c\u9032\u884c\u4ea4\u53c9\u6bd4\u5c0d\u51fa\u6240\u8981\u8fa8\u8b58\u55ae\u5143\u4e4b\u97f3\u6a19\uff0c\u6b64\u65b9\u6cd5\u53ef\u4ee5\u6709 \u6548\u6e1b\u4f4e\u8072\u5b78\u6a21\u578b\u6578\u91cf\uff0c\u56e0\u6b64\u5177\u6709\u8f03\u4f73\u7684\u6297\u566a\u80fd\u529b[8]\u3002 \u738b\u5c0f\u5ddd[9]\u6559\u6388\u5728\"\u8a9e\u97f3\u4fe1\u865f\u8655\u7406\"\u4e00\u66f8\u4e2d\u5c0d\u65bc\u8a9e\u97f3\u8fa8\u8b58\u7684\u57fa\u790e\u77e5\u8b58\u80cc\u666f\u548c\u6d41\u7a0b\u6709\u76f8 \u7576\u8a73\u7d30\u7684\u6558\u8ff0\uff0c\u5c0d\u65bc\u60f3\u8981\u5b78\u7fd2\u8a9e\u97f3\u8fa8\u8b58\u7684\u4eba\u662f\u4e00\u672c\u5f88\u597d\u7684\u5165\u9580\u66f8\u7c4d\u3002\u9577\u5e9a\u5927\u5b78\u5247\u662f\u9577\u5e74 \u4f86\u9032\u884c\u53f0\u7063\u672c\u571f\u8a9e\u8a00\u7684\u7814\u7a76\uff0c\u6881\u654f\u96c4[10]\u5efa\u7acb\u4ee5 IPA \u70ba\u57fa\u790e\u7684\u53f0\u8a9e\u97f3\u6a19\u96c6\u5408 ForPA\u3001\u6587 \u5b57\u8f49\u8a9e\u97f3(TTS)\u548c\u767c\u97f3\u5075\u932f\u61c9\u7528\u65bc\u8a9e\u8a00\u6559\u5b78\u8207\u53f0\u8a9e\u8a9e\u97f3\u8a9e\uf9be\u5eab\u8a2d\u8a08\u8207\u6536\u96c6\u3002\u9673\u5fd7\u5b87[11]\u5247 \u63a2\u8a0e\u4e86\u540c\u6642\u5c0d\u570b\u53f0\u96d9\u8a9e\u7684\u5927\u8a5e\u5f59\u9023\u7e8c\u8a9e\u97f3\u8fa8\u8b58\u3002\u694a\u6c38\u6cf0[12]\u5c07\u97f3\u6a19\u66ff\u63db\u6539\u8b8a\u6210\u4e2d\u6587\u800c\u5c07 \u96b1\u85cf\u5f0f\u99ac\u53ef\u592b\u6a21\u578b\u904b\u7528\u5728\u4e2d\u6587\u8fa8\u8b58\u4e0a\u3002 \u975e\u672c\u570b\u6bcd\u8a9e\u4eba\u58eb\u6240\u8b1b\u7684\u8a9e\u8a00\u6703\u7522\u751f\u767c\u97f3\u8b8a\u7570\uff0c\u4f8b\u5982\u53f0\u7063\u4eba\u8b1b\u82f1\u6587\uff0c\u6b64\u7a2e\u767c\u97f3\u8b8a\u7570\u6703 \u4f7f\u7cfb\u7d71\u7522\u751f\u6975\u5927\u7684\u8aa4\u5224\u3002\u570b\u7acb\u6210\u529f\u5927\u5b78\u7684\u8521\u4f69\u73ca\u3001\uf972\u6db5\u5e73\u3001\u5433\u5b97\u61b2[13]\u63d0\u51fa\u4ee5\u66f4\u5c0f\u7684\u55ae \u4f4d \u2500 senone \u4f5c\u70ba\u57fa\u672c\u7684\u8fa8\u8b58\u55ae\u5143\uff0c\u4ee5\u66f4\u52a0\u8a73\u7d30\u7684\u6a21\u64ec\u767c\u97f3\u8b8a\u7570\uff0c\u4e26\u5efa\u7acb\u5305\u542b\u767c\u97f3\u8b8a\u7570 \u4e4b\u82f1\u6587\u8072\u5b78\u6a21\u578b\u3002 \u4e8c\u3001\u7cfb\u7d71\u67b6\u69cb (\u4e00)\u7cfb\u7d71\u6846\u67b6 \u5716\u4e00\u70ba\u672c\u8ad6\u6587\u4e4b\u7cfb\u7d71\u67b6\u69cb\u3002\u6839\u64da\u5176\u8655\u7406\u6d41\u7a0b\u53ef\u4ee5\u5206\u70ba\u8a13\u7df4\u968e\u6bb5\u8207\u6e2c\u8a66\u968e\u6bb5\u3002\u8a13\u7df4\u968e \u6bb5\u70ba\u4f7f\u7528\u72c4\u5f0f\u5206\u4f48\u9032\u884c\u5c0d\u97f3\u7d20\u8072\u5b78\u6a21\u578b\u7684\u6df7\u6dc6\u5075\u6e2c\uff0c\u6e2c\u8a66\u968e\u6bb5\u5247\u662f\u5c07\u6df7\u6dc6\u767c\u751f\u7684\u6a21\u578b\u9032 \u884c\u5206\u5272\u6216\u5408\u4f75\u5f8c\u9032\u884c\u6548\u80fd\u6e2c\u8a66\u3002 \u5716\u4e00\u3001\u7cfb\u7d71\u67b6\u69cb (\u4e8c)\u8a13\u7df4\u90e8\u5206 \u5728\u8a13\u7df4\u6642\u7684\u521d\u59cb\u968e\u6bb5\u70ba\u4f7f\u7528 39 \u7dad MFCC \u53c3\u6578\u548c HTK \u4f86\u5efa\u7acb\u8d77\u4ee5\u96b1\u85cf\u5f0f\u99ac\u53ef\u592b\u6a21 \u578b(HMM, Hidden Markov Model)\u70ba\u57fa\u790e\u7684\u591a\u8a9e\u8a00 Tri-phone \u8072\u5b78\u6a21\u578b\u3002\u6b64\u968e\u6bb5\u5c07\u4e0d\u540c\u8a9e \u8a00\u76f8\u540c\u97f3\u6a19\u8996\u70ba\u4e0d\u540c\u97f3\u6a19\uff0c\u4e26\u4e14\u6839\u64da Tri-phone \u5b9a\u7fa9\u5206\u88c2\u6a21\u578b\uff0c\u4f46\u662f\u4e0d\u9032\u884c State-Tying \u548c\u6a21\u578b\u5408\u4f75\u3002\u7b2c\u4e8c\u968e\u6bb5\u70ba\u4f7f\u7528\u8a9e\u6599\u6a19\u8a18\u5efa\u7acb\u4ee5\u72c4\u5f0f\u5206\u4f48(LDA, Latent Dirichlet Allocation) \u70ba\u57fa\u790e\u7684\u5075\u6e2c\u5668\uff1a\u7fa4\u5167\u9a57\u8b49\u5075\u6e2c\u5668(Intra-clustering verification Detector)\u3002\u904e\u4f86\u6839\u64da\u4ee5\u72c4 \u5f0f\u5206\u4f48\u70ba\u57fa\u790e\u7684\u7fa4\u5167\u9a57\u8b49\u5075\u6e2c\u5668\u5206\u5225\u6839\u64da\u767c\u97f3\u90e8\u4f4d\u548c\u767c\u97f3\u65b9\u6cd5\u9032\u884c\u5206\u7fa4\uff0c\u7531\u65bc\u6bcf\u4e00\u500b\u97f3 \u7d20\u90fd\u53ef\u4ee5\u5c0d\u61c9\u7684\u55ae\u4e00\u7684\u767c\u97f3\u90e8\u4f4d\u548c\u767c\u97f3\u65b9\u6cd5\uff0c\u56e0\u6b64\u5c07\u767c\u97f3\u65b9\u6cd5\u8207\u767c\u97f3\u90e8\u4f4d\u7686\u5206\u985e\u5728\u540c\u4e00 (\u4e09)\u6e2c\u8a66\u90e8\u5206 \u6e2c\u8a66\u90e8\u5206\u6700\u4e3b\u8981\u7684\u662f\u5c0d\u8a13\u7df4\u968e\u6bb5\u6240\u4fee\u6b63\u7684\u8072\u5b78\u6a21\u578b\u9032\u884c\u6548\u80fd\u8207\u6b63\u78ba\u6027\u6e2c\u8a66\u3002\u5f9e\u4f7f\u7528 \u8005\u8f38\u5165\u7684\u8a9e\u53e5\u62bd\u53d6 39 \u7dad MFCC \u53c3\u6578\u548c\u8a9e\u97f3\u5c6c\u6027\u4e26\u4f7f\u7528\u4fee\u6b63\u904e\u7684\u8072\u5b78\u6a21\u578b\u9032\u884c\u8fa8\u8b58\uff0c\u4e26 \u6839\u64da\u8fa8\u8b58\u7d50\u679c\u4f86\u5224\u65b7\u662f\u5426\u6709\u6e1b\u5c11\u8072\u5b78\u6a21\u578b\u6df7\u6dc6\u7a0b\u5ea6\u3002 \u4e09\u3001\u767c\u97f3\u4e8b\u4ef6 \u6a19(\u865f\u96c6\uff0c\u4f46\u5728\u4e0d\u540c\u8a9e\u8a00\u7684\u60c5\u6cc1\u4e0b\uff0c\u6703\u51fa\u73fe\u96d6\u7136\u5c6c\u65bc\u540c\u4e00\u500b\u97f3\u6a19\uff0c\u4f46\u662f\u5176\u767c\u97f3\u6a21\u5f0f\u4ecd\u7136\u6703\u6709 \u70ba\u4e86\u89e3\u6c7a\u767c\u97f3\u8b8a\u7570\u800c\u9020\u6210\u8a9e\u97f3\u6a21\u578b\u6df7\u6dc6\u7684\u60c5\u5f62\uff0c\u570b\u5916\u5b78\u8005\u6839\u64da\u4e0d\u540c\u7684\u767c\u97f3\u8b8a\u7570\u6cd5\u9032 \u70ba\u4e86\u89e3\u6c7a\u8a9e\u97f3\u8fa8\u8b58\u7684\u74f6\u9838\uff0c\u7f8e\u570b\u55ac\u6cbb\u4e9e\u7406\u5de5\u5b78\u9662\u7684\u674e\u9326\u8f1d(C.-H. Lee)\u6559\u6388\u63d0\u51fa\u4e86\u5075 \u884c\u4e86\u8a31\u591a\u7684\u7814\u7a76\u3002\u95dc\u65bc\u500b\u4eba\u5316\u767c\u97f3\u8b8a\u7570\u7684\u7814\u7a76\uff0c1993 \u5e74 Hamada \u548c Miki \u7b49\u4eba\u63d0\u51fa\uff0c\u904b \u6e2c\u5f0f(Detection-based)\u7684\u65b9\u6cd5\uff0c\u5176\u4e3b\u8981\u6982\u5ff5\u70ba\u4eba\u985e\u518d\u767c\u97f3\u7684\u6642\u5019\u6703\u6709\u767c\u97f3\u90e8\u4f4d(Place)\u8207\u767c \u7528\u52d5\u614b\u898f\u5283(dynamic programming)\u548c\u5411\uf97e\uf97e\u5316(vector quantization)\u7684\u65b9\u5f0f\u6bd4\u8f03 Native \u97f3\u65b9\u6cd5(Manner)\uff0c\u85c9\u7531\u767c\u97f3\u8a9e\u8a00\u5b78\u4f86\u63cf\u8ff0\u8a9e\u97f3\u3002\u800c\u767c\u97f3\u90e8\u4f4d\u8207\u767c\u97f3\u65b9\u6cd5\u5247\u7d71\u7a31\u70ba\u767c\u97f3\u4e8b Speaker \u8207 Non-Native Speaker \u5728\u540c\u4e00\u500b\u5b57\u767c\u97f3\u4e0a\u7684\u5dee\u7570[2] \u30021996 \u5e74 Neumeyer \u548c Franc \u4ef6\u3002\u800c\u70ba\u4e86\u8981\u5075\u6e2c\u767c\u97f3\u4e8b\u4ef6\uff0c\u5247\u85c9\u7531\u76f4\u63a5\u89c0\u5bdf\u8a9e\u97f3\u8a0a\u865f\u4f86\u5c0b\u627e\u51fa\u7279\u5b9a\u767c\u97f3\u4e8b\u4ef6\u4e0b\u7684\u6709\u6548 \u7b49\u4eba\u5247\u5b9a\u7fa9\u4e86 HMM Log-Likelihood\u3001Segment duration \u548c Timing \u7b49\u7279\u5fb5\u53c3\u6578\uff0c\u91dd\u5c0d\u6574 \u7279\u5fb5\u4f86\u5075\u6e2c\uff0c\u800c\u9019\u4e9b\u7279\u5fb5\u5247\u7a31\u70ba\u8a9e\u97f3\u5c6c\u6027(Articulatory Features)\u3002\u5176\u7279\u6027\u662f\u85c9\u7531\u591a\u5c64\u5316\u67b6 \u500b\u53e5\u5b50\u505a\u767c\u97f3\u8a55\u4f30\uff0c\u800c\u4e14\u5be6\u9a57\u7684\u7d50\u679c\u767c\u73fe normalized segment duration scores \u8207\u5c08\u5bb6\u7d66 \u69cb\u800c\u7e2e\u6e1b\u4e86\u6a21\u578b\u6578\u91cf\uff0c\u4e5f\u56e0\u70ba\u8207\u4eba\u985e\u767c\u97f3\u7684\u65b9\u6cd5\u6709\u6240\u95dc\u9023\uff0c\u56e0\u6b64\u5176\u5177\u6709\u8f03\u9ad8\u7684\u5f37\u5065\u6027\u3002 \u56e0\u6b64\u5c0d\u65bc\u4e0d\u540c\u8a9e\u8a00\u76f8\u540c\u97f3\u6a19\u4e5f\u8f03\u4e0d\u5bb9\u6613\u7522\u751f\u6df7\u6dc6\u3002\u70ba\u4e86\u8655\u7406\u5728\u591a\u8a9e\u8fa8\u8b58\u7684\u74b0\u5883\u4e0b\u7684\u97f3\u6a19 \u6df7\u6dc6\uff0c\u672c\u6587\u4f7f\u7528\u4e86\u767c\u97f3\u4e8b\u4ef6\u4f86\u9032\u884c\u5075\u6e2c\u3002 \u767c\u97f3\u4e8b\u4ef6\u6709\u8a31\u591a\u4e0d\u540c\u7684\u5206\u985e\u6cd5\uff0c\u4f46\u5176\u5171\u901a\u7279\u6027\u5247\u70ba\u53ef\u4ee5\u8de8\u8a9e\u8a00\u7684\u5c0d\u61c9\u5230\u7279\u5b9a\u7684\u97f3\u6a19\uff0c \u4e88\u7684\u5206\u6578\u4e2d\u6709\u6700\u9ad8\u7684\u76f8\u95dc\u6027[3]\u30021997 \uf98e Final State Transducer (FST) [6]\u3002 \u56e0\u6b64\u70ba\u4e86\u5efa\u7acb\u8d77\u8de8\u8a9e\u8a00\u7684\u97f3\u6a19\u96c6\u5408\uff0c\u672c\u6587\u4e3b\u8981\u4f7f\u7528\u4e86\u570b\u969b\u97f3\u6a19\u96c6\u5408</td></tr><tr><td>\u7fa4\u4e4b\u97f3\u7d20\u5b9a\u70ba\u5408\u4f75\u76ee\u6a19\uff0c\u800c\u5c07\u6c92\u6709\u5169\u8005\u7686\u5206\u985e\u5728\u540c\u4e00\u7fa4\u7684\u97f3\u6a19\u793a\u70ba\u4e0d\u540c\u4e4b\u5206\u7fa4\u3002\u6700\u5f8c\u518d</td></tr><tr><td>\u6839\u64da\u6b64\u5206\u7fa4\u5c0d\u7b2c\u4e00\u968e\u6bb5\u6240\u5efa\u7acb\u7684\u8072\u5b78\u6a21\u578b\u9032\u884c\u5408\u4f75\uff0c\u6700\u5f8c\u5373\u53ef\u5f97\u5230\u4fee\u6b63\u904e\u4e4b\u8072\u5b78\u6a21\u578b\u3002</td></tr></table>",
"text": "The International Phonetic Alphabet, IPA)\u3001\u5b57\u6bcd\u97f3\u6a19\u8a55\u4f30\u6cd5(Speech Assessment Methods Phonetic Alphabet, SAMPA)\u3001Worldbet \u7b49\u3002\u7b2c\u4e09\u7a2e\u662f\u8a08\u7b97\u4e0d\u540c\u97f3\u6a19\u4e4b\u9593\u7684\u76f8\u4f3c\u6027\uff0c\u4e26\u7531 \u4f30\u7b97\u51fa\u4f86\u7684\u76f8\u4f3c\u6027\u4f86\u5efa\u7acb\u5171\u540c\u7684\u97f3\u6a19\u96c6\u3002 \u672c\u6587\u4e3b\u8981\u6ce8\u91cd\u65bc\u4f7f\u7528\u5c08\u5bb6\u77e5\u8b58\u6240\u5efa\u7acb\u7684\u8de8\u8a9e\u8a00\u97f3\u6a19\u96c6\u5408\u4e4b\u65b9\u6cd5\uff0c\u5728\u6b64\u4e3b\u8981\u662f\u4f7f\u7528 \u570b\u969b\u97f3\u6a19(The International Phonetic Alphabet, IPA)\uff0c\u96d6\u7136 IPA \u63d0\u4f9b\u4e86\u4e00\u500b\u901a\u7528\u7684\u97f3\u6a19\u7b26 Ronen \u7b49\u4eba\u63d0\u51fa MisPronunciation network \u7684\u6982 \uf9a3\u3002\u8003\u616e\u6bcf\u500b\u97f3\u7d20\u5728 native speaker \u8207 non-native speaker \u7684\u767c\u97f3\u60c5\u5f62\uff0c\u5efa\uf9f7\u5c0d\u61c9\u7684 HMM Model\uff0c\u8fa8\u8b58\u7684\u6642\u5019\u767c\u97f3\u7db2\uf937\u540c\u6642\u8003\u616e native speaker \u8207 non-native speaker \u7684\u767c\u97f3\u60c5\u5f62 [4]\u30021999 \uf98e Franco \u7b49\u5247\u662f\u4f7f\u7528\u5169\u500b\u5206\u5225\u7531 native speaker \u8207 non-native speaker \u6240\u8a13\uf996 \u7684\u8072\u5b78\u6a21\u578b\uff0c\u5229\u7528 log-likelihood ratio \u4f86\u8a55\u4f30\u767c\u97f3\u932f\u8aa4\uff0c\u540c\u6642\u4e5f\u8b49\u660e\uf9ba\u9019\u6a23\u7684\u65b9\u6cd5\u6bd4\u5229 \u7528 a posterior score \u7684\u65b9\u5f0f\u8207\u5c08\u5bb6\u6240\u7d66\u4e88\u7684\u5206\u6578\u6709\uf901\u9ad8\u7684\u76f8\u95dc\u6027[5]\u3002\u5728\u6a21\u578b\u65b9\u9762\uff0c\u5247\u662f \u9ebb\u7701\u7406\u5de5\u5b78\u9662\u7684",
"type_str": "table"
},
"TABREF1": {
"html": null,
"num": null,
"content": "<table><tr><td>(\u4e8c)\u6f5b\u85cf\u72c4\u5f0f\u5206\u4f48\u6a21\u578b(Latent Dirichlet Allocation Model)</td></tr><tr><td>1. \u6f5b\u85cf\u72c4\u5f0f\u5206\u4f48(Latent Dirichlet Allocation)</td></tr><tr><td>\u6f5b \u85cf \u72c4\u5f0f\u5206\u4f48(Latent Dirichlet Allocation, LDA) \u662f \u7531 \u6a5f \u7387 \u5f0f \u6f5b \u85cf \u8a9e \u610f \u5206 \u6790</td></tr><tr><td>(Probabilistic Latent Semantic Analysis, PLSA)\u5ef6\u4f38\u767c\u5c55\u800c\u4f86\uff0c\u800c\u6a5f\u7387\u5f0f\u6f5b\u85cf\u8a9e\u610f\u5206\u6790\u53c8</td></tr><tr><td>\u662f\u7531\u6f5b\u85cf\u8a9e\u610f\u5206\u6790(Latent Semantic Analysis, LSA)\u767c\u5c55\u800c\u4f86\u3002\u4e0a\u8ff0\u4e09\u500b\u6a21\u578b\u90fd\u662f\u5c6c\u65bc\u6f5b</td></tr><tr><td>\u85cf\u4e3b\u984c\u6a21\u578b(Latent Topic Model)\uff0c\u7531\u65bc N \u9023\u6a21\u578b\u6240\u9762\u81e8\u7684\u7f3a\u4e4f\u9577\u8ddd\u96e2\u8cc7\u8a0a\u548c\u8cc7\u6599\u7a00\u758f\u554f</td></tr><tr><td>\u984c\uff0c\u800c\u6f5b\u85cf\u4e3b\u984c\u6a21\u578b\u5247\u662f\u89e3\u6c7a\u554f\u984c\u591a\u7a2e\u65b9\u6cd5\u7684\u5176\u4e2d\u4e4b\u4e00\u3002\u5176\u6982\u5ff5\u70ba\u4f7f\u7528\u975e\u76e3\u7763\u5f0f\u5b78\u7fd2\u65b9</td></tr><tr><td>\u6cd5(Unsupervised Training)\u4f86\u627e\u51fa\u96b1\u542b\u65bc\u6587\u4ef6\u6216\u6587\u7ae0\u4e2d\u7684\u6700\u4e3b\u8981\u7684\u4e3b\u984c\u8a9e\u610f\u8cc7\u8a0a\u3002</td></tr><tr><td>\u6f5b\u85cf\u72c4\u5f0f\u5206\u4f48\u4e0d\u540c\u65bc\u6a5f\u7387\u5f0f\u6f5b\u85cf\u8a9e\u610f\u5206\u6790\u7684\u5730\u65b9\u5728\u65bc\u5f9e\u6587\u4ef6\u5230\u4e3b\u984c\u4e4b\u9593\u591a\u4e86\u4e00\u5c64</td></tr><tr><td>\u7684\u72c4\u5f0f\u5206\u5e03(Dirichlet Distribution)\uff0c\u4f7f\u5f97\u6a21\u578b\u53c3\u6578\u6578\u91cf\u4e0d\u6703\u56e0\u70ba\u8a9e\u6599\u589e\u52a0\u800c\u5927\u5e45\u5ea6\u7684\u589e</td></tr><tr><td>\u52a0\uff0c\u540c\u6642\u5c0d\u65bc\u51fa\u73fe\u5728\u8a9e\u6599\u5eab\u5916\u7684\u6587\u4ef6\u4e5f\u53ef\u4ee5\u5f9e\u72c4\u5f0f\u5206\u4f48\u4e2d\u53d6\u51fa\u4e00\u500b\u6700\u9069\u5408\u6b64\u6587\u4ef6\u7684\u6f5b\u85cf</td></tr><tr><td>\u4e3b\u984c\u6a5f\u7387\u5206\u4f48\u3002\u7531\u65bc\u6f5b\u85cf\u72c4\u5f0f\u5206\u4f48\u662f\u85c9\u7531\u9006\u5411\u901a\u904e\u6587\u4ef6\u5efa\u7acb\u751f\u6210\u6a21\u578b\uff0c\u56e0\u6b64\u8981\u5148\u7406\u89e3\u6f5b</td></tr><tr><td>\u85cf\u72c4\u5f0f\u5206\u4f48\u662f\u5982\u4f55\u7522\u751f\u4e00\u7bc7\u6587\u7ae0\u7684\u3002</td></tr><tr><td>\u5047\u8a2d\u8a9e\u6599\u5eab M 2.\u6f5b\u85cf\u72c4\u5f0f\u5206\u4f48\u5075\u6e2c\u5668</td></tr><tr><td>\u96d6\u7136\u5075\u6e2c\u5f0f\u65b9\u6cd5(Detection Based)\u5728\u591a\u8a9e\u74b0\u5883\u4e0b\u5c0d\u65bc\u4e0d\u540c\u8a9e\u8a00\u540c\u97f3\u6a19\u4e4b\u767c\u97f3\u4ecd\u7136\u5177</td></tr><tr><td>\u6709\u9ad8\u5f37\u5065\u6027\u4e0d\u81f3\u65bc\u6df7\u6dc6\uff0c\u4f46\u662f\u5f80\u5f80\u4e0d\u540c\u8a9e\u8a00\u7684\u8a9e\u8005\u5c0d\u65bc\u76f8\u540c\u97f3\u6a19\u4e4b\u767c\u97f3\u8b8a\u7570\u4ecd\u7136\u6703\u4f7f\u5f97</td></tr><tr><td>\u767c\u97f3\u4e8b\u4ef6\u7522\u751f\u6df7\u6dc6\uff0c\u6709\u9452\u65bc\u6b64\u73fe\u8c61\u96e3\u4ee5\u4f7f\u7528\u77ed\u8ddd\u96e2\u7684\u8a0a\u865f\u4f86\u89e3\u6c7a\uff0c\u56e0\u6b64\u6211\u5011\u4f7f\u7528\u9577\u8ddd\u96e2</td></tr><tr><td>\u8a5e\u5f59\u8a9e\u610f\u8cc7\u8a0a\u4f86\u5354\u52a9\u5075\u6e2c\u51fa\u6df7\u6dc6\u7684\u554f\u984c\u3002\u4e26\u5c07\u6703\u767c\u8072\u6df7\u6dc6\u7684\u8072\u5b78\u6a21\u578b\u4ee5\u53ca\u6027\u8cea\u76f8\u540c\u7684\u8072</td></tr><tr><td>\u5b78\u6a21\u578b\u9032\u884c\u5408\u4f75\u3002\u4e0d\u540c\u65bc\u4e09\u9023\u97f3\u7d20\u6a21\u578b(Tri-Phone)\u7684\u72c0\u614b\u805a\u985e(State-Tying)\u6839\u64da\u77ed\u8ddd\u96e2\u7684 2003 \u5e74\u6240 \u8cc7\u6599\u76f8\u4f3c\u5ea6\u5408\u4f75\uff0c\u6211\u5011\u85c9\u7531\u6f5b\u85cf\u72c4\u5f0f\u5206\u4f48\u64f7\u53d6\u9577\u8ddd\u96e2\u8a5e\u5f59\u8a9e\u610f\u8cc7\u8a0a\uff0c\u4f8b\u5982\u89f8\u767c\u8a5e\u5c0d\uff0c\u4ee5</td></tr><tr><td>\u63d0\u51fa\u7684\uff0c\u6700\u521d\u662f\u7528\u65bc\u6587\u7ae0\u548c\u6587\u672c\u7684\u4e3b\u984c\u5075\u6e2c\u3002\u72c4\u5f0f\u5206\u4f48\u6a21\u578b\u6709\u4e00\u500b\u5148\u6c7a\u689d\u4ef6\uff1a\u8a5e\u888b\u5047\u8a2d \u53ca\u4f7f\u7528\u4e86\u767c\u97f3\u8a9e\u8a00\u5b78\u7684\u5b9a\u7fa9\u5c0d\u5206\u88c2\u7684\u4e09\u9023\u97f3\u7d20\u6a21\u578b\u9032\u884c\u8072\u5b78\u6a21\u578b\u7684\u5408\u4f75\u3002</td></tr><tr><td>(Bag of Words Assumption)\uff0c\u4e5f\u5c31\u662f\u4e0d\u8003\u616e\u8a5e\u5f59(Word)\u5728\u6587\u7ae0\u4e2d\u7684\u51fa\u73fe\u9806\u5e8f\u548c\u6587\u6cd5\u95dc\u4fc2\uff0c</td></tr><tr><td>\u53ea\u8003\u616e\u55ae\u4e00\u8a5e\u5f59\u5728\u7279\u5b9a\u6587\u7ae0\u4e2d\u4e4b\u51fa\u73fe\u6b21\u6578\u3002\u5176\u6240\u4f7f\u7528\u4e4b\u539f\u7406\u70ba\u67d0\u4e9b\u8a5e\u5f59\u5728\u7279\u5b9a\u4e3b\u984c\u4e0b\u51fa \u7531\u65bc\u6f5b\u85cf\u72c4\u5f0f\u5206\u4f48\u5177\u6709\u8a5e\u888b\u5047\u8a2d\u7684\u524d\u63d0\u4e0b\uff0c\u5176\u6240\u89c0\u5bdf\u5230\u7684\u8cc7\u6599\u9577\u5ea6\u8db3\u5920\u5305\u542b\u9577\u8ddd\u96e2\u7684\u8a5e</td></tr><tr><td>\u73fe\u4e4b\u6a5f\u7387\u548c\u6b21\u6578\u8f03\u9ad8\uff0c\u56e0\u6b64\u72c4\u5f0f\u5206\u4f48\u7684\u6a21\u578b\u5efa\u7acb\u65b9\u5f0f\u70ba\u7d71\u8a08\u5404\u500b\u8a5e\u5f59(Word)\u5728\u6587\u672c\u4e2d\u51fa \u5f59\u8a9e\u610f\u8cc7\u8a0a\u3002\u56e0\u70ba\u8a9e\u97f3\u4e8b\u4ef6\u548c\u8a9e\u97f3\u5c6c\u6027\u7684\u76f8\u4e92\u95dc\u4fc2\u4e5f\u662f\u5c6c\u65bc\u968e\u5c64\u5316\u67b6\u69cb\uff0c\u56e0\u6b64\u6839\u64da\u5176\u8072</td></tr><tr><td>\u73fe\u7684\u6b21\u6578\u4f86\u8a08\u7b97\u3002\u7531\u65bc\u6b64\u65b9\u6cd5\u5728\u6587\u672c\u548c\u6587\u7ae0\u5206\u985e\u53ef\u4ee5\u53d6\u5f97\u76f8\u7576\u597d\u7684\u6210\u6548\uff0c\u56e0\u6b64\u88ab\u5ee3\u6cdb\u7684 \u5b78\u6a21\u578b\u7684\u76f8\u4e92\u95dc\u4fc2\u5c0d\u6f5b\u85cf\u72c4\u5f0f\u5206\u4f48\u4e4b\u67b6\u69cb\u91cd\u65b0\u5efa\u69cb\u5f8c\u4e4b\u5716\u578b\u8868\u793a\u6cd5\u5982\u5716\u4e8c \uff0c\u5716\u5f62\u4e2d\u53c3</td></tr><tr><td>\u4f7f\u7528\u65bc\u4e3b\u984c\u5075\u6e2c\u548c\u4e3b\u984c\u5206\u985e\u7684\u61c9\u7528\u4e0a\u3002 \u6578\u6240\u4ee3\u8868\u4e4b\u610f\u7fa9\u5982\u8868\u4e00\u3002</td></tr><tr><td>\u5728\u591a\u8a9e\u74b0\u5883\u4e0b\uff0c\u56e0\u70ba\u4e0d\u540c\u8a9e\u8a00\u7684\u8a9e\u8005\u9593\u4e4b\u767c\u97f3\u8b8a\u5316\u5e45\u5ea6\u6703\u76f8\u7576\u7684\u5927\uff0c\u4e5f\u56e0\u6b64\u82e5\u662f\u6839 \u5728\u9019\u88e1\u5c07\u6bcf\u4e00\u6bb5\u8a9e\u53e5(Utterance)\u8996\u70ba\u4e00\u500b\u6587\u4ef6(Document)\uff0c\u56e0\u6b64\u6839\u64da\u6f5b\u85cf\u5f0f\u72c4\u5f0f\u5206</td></tr><tr><td>\u64da\u4ee5\u5f80\u7684\u8072\u5b78\u6a21\u578b\u53ea\u4f7f\u7528\u77ed\u671f\u5167(Short-term)\u7684\u8cc7\u6599\u4f86\u8fa8\u8b58\u6240\u80fd\u63d0\u5347\u7684\u6548\u80fd\u5e45\u5ea6\u6709\u9650\uff0c\u800c \u4f48\u5b9a\u7fa9\u5247\u53ef\u4ee5\u5f97\u5230\u7279\u5b9a\u767c\u97f3\u4e8b\u4ef6\u8207\u8a9e\u610f\u4e0a\u7684\u76f8\u5c0d\u95dc\u4fc2\u3002\u800c\u806f\u5408\u6a5f\u7387\u5206\u4f48\u70ba</td></tr><tr><td>\u4e14\u5728\u767c\u97f3\u8b8a\u7570\u7522\u751f\u7684\u72c0\u6cc1\u4e0b\u4e5f\u96e3\u4ee5\u53ea\u4f7f\u7528\u8a0a\u865f\u4e0a\u77ed\u671f\u7684\u8cc7\u6599\u4f86\u52a0\u4ee5\u8b58\u5225\u3002\u4f46\u8a9e\u97f3\u8a0a\u865f\u7684 p(M|\u03b1 M , \u03b1 P</td></tr><tr><td>\u6642\u8b8a\u6027\u76f8\u7576\u7684\u5927\uff0c\u56e0\u6b64\u5728\u8a0a\u865f\u9762\u4e00\u6b21\u4f7f\u7528\u8f03\u9577\u8cc7\u6599\u7684\u65b9\u6cd5\u96e3\u5ea6\u4e5f\u9ad8\u3002\u6240\u4ee5\u672c\u6587\u6839\u64da\u8a9e\u6599</td></tr><tr><td>\u6a19\u8a18\u4f7f\u7528\u72c4\u5f0f\u5206\u4f48\u5c07\u6574\u53e5\u7684\u8cc7\u8a0a\u4e5f\u4e00\u4f75\u8003\u616e\u4ee5\u63d0\u5347\u5c0d\u767c\u97f3\u8b8a\u7570\u4e4b\u5f37\u5065\u6027\u3002</td></tr></table>",
"text": "\u4e2d\u6709 K \u500b\u4e3b\u984c\uff0cT1,T2,T3,\u2026.Tk\uff0c\u4e26\u4e14\u6709 V \u500b\u5b57\u5f59\uff0c\u7576\u96a8\u6a5f\u9078\u53d6\u4e00\u500b \u4e3b\u984c Ti \u7684\u6642\u5019\uff0c\u4ee5 Ti \u70ba\u4e3b\u984c\u7684\u6587\u7ae0\u5247\u6709\u4e00\u7684\u5e8f\u5217\u6587\u5b57\uff0c\u800c\u9019\u4e9b\u6587\u5b57\u8207 Ti \u6709\u95dc\u9023\uff0c\u4e26\u4e14 \u6709\u4e00\u500b\u6a5f\u7387\u503c\u4ee3\u8868\u5728\u4e3b\u984c Ti \u4e0b\u6642\u6bcf\u500b\u6587\u5b57\u6240\u51fa\u73fe\u4e4b\u6a5f\u7387\u3002\u800c\u9078\u64c7\u5176\u4ed6\u4e3b\u984c\u6642\u5f8c\u4e5f\u6703\u6709 \u76f8\u540c\u7684\u53c3\u6578\u4f86\u63cf\u8ff0\u8a72\u4e3b\u984c\u3002\u6b64\u6642\u9650\u5b9a\u6587\u6a94\u9577\u5ea6\u70ba N\uff0c\u4e0d\u505c\u7684\u6311\u9078\u6587\u5b57\u76f4\u5230\u6578\u91cf\u5230 N \u5f8c\uff0c \u4fbf\u53ef\u4ee5\u85c9\u7531\u5c0d\u61c9\u7684\u53c3\u6578\u5f97\u5230\u8a72\u5efa\u7acb\u6587\u4ef6\u5c0d\u65bc\u5404\u4e3b\u984c\u7684\u76f8\u95dc\u6027\u3002",
"type_str": "table"
},
"TABREF2": {
"html": null,
"num": null,
"content": "<table><tr><td>\u5206\u5e03 \u67d0\u6bb5\u8a9e\u53e5\u7684\u767c\u97f3\u65b9\u6cd5\u767c\u751f\u4e4b\u6a5f\u7387 \u6240\u6709\u8a9e\u8a00\u4e0b\u4e4b\u8072\u5b78\u6a21\u578b\u6a5f\u7387 \u67d0\u6bb5\u8a9e\u53e5\u7684\u767c\u97f3\u90e8\u4f4d\u767c\u751f\u4e4b\u6a5f\u7387 \u8072\u5b78\u6a21\u578b(phone model) \u8a9e\u6599\u5eab \u67d0\u6bb5\u8a9e\u53e5\u7684\u8072\u5b78\u6a21\u578b\u96c6\u5408 \u8a9e\u8a00 \u767c\u97f3\u90e8\u4f4d(Place) \u767c\u97f3\u65b9\u6cd5(Manner) \u6839\u64da\u5f0f 4.1 \u8981\u8a08\u7b97\u7684\u53c3\u6578\u70ba \u03b1M, \u03b1P ,\u03b2\uff0c\u4f46\u7531\u65bc\u7528\u4f86\u4f30\u7b97\u6f5b\u85cf\u72c4\u5f0f\u5206\u4f48\u53c3\u6578\u7684\u6700\u5927\u671f\u671b\u6f14 \u03b2 \u03b8M \u03b8P W M N L ZP ZM \u7b97\u6cd5(Expectation-maximization algorithm, EM algorithm)\u4e00\u6b21\u53ea\u80fd\u4f30\u6e2c\u5169\u500b\u53c3\u6578\uff0c\u800c\u6839\u64da \u767c\u97f3\u8a9e\u8a00\u5b78\u4e4b\u5b9a\u7fa9\u767c\u97f3\u65b9\u6cd5\u548c\u767c\u97f3\u90e8\u4f4d\u5169\u4e8b\u4ef6\u70ba\u4e92\u76f8\u7368\u7acb(independent)\uff0c\u56e0\u6b64\u6211\u5011\u53ef\u4ee5 \u5c07\u5176\u62c6\u6210 \u03b1M\u3001\u03b2 \u548c \u03b1P \u3001\u03b2 \u5169\u500b\u90e8\u5206\u5206\u958b\u4f30\u6e2c\u3002 \u800c\u62c6\u958b\u5f8c\u5206\u5225\u7684\u806f\u5408\u6a5f\u7387\u5247\u5206\u5225\u70ba\u70ba\u5f0f 4.2 \u8207\u5f0f 4.3 \u6240\u793a p(M|\u03b1 M , \u03b2) = \u222b p(\u03b8 M |\u03b1 M ) {\u220f p(Z M |\u03b8 M )p(w|Z M , L, \u03b2) N n=1 } d\u03b8 M (\u5f0f 4.2) p(M|\u03b1 P , \u03b2) = \u222b p(\u03b8 P |\u03b1 P ) {\u220f p(Z P |\u03b8 P )p(w|Z P , L, \u03b2) N n=1 } d\u03b8 P (\u5f0f 4.3) \u4e4b\u5f8c\u6211\u5011\u53ef\u4ee5\u7531 \u03b2 \u5f97\u5230\u6bcf\u500b\u4e3b\u984c\u4e0b\u6240\u6709\u97f3\u7d20\u7684\u51fa\u73fe\u6a5f\u7387\uff0c\u800c\u9032\u800c\u5f97\u5230\u5206\u985e\u904e\u5f8c\u7684\u97f3\u7d20\u96c6 \u5408\u3002\u5728\u539f\u59cb\u7684\u72c4\u5f0f\u5206\u4f48\u4e2d\u6703\u767c\u751f\u82e5\u55ae\u4e00\u6587\u4ef6\u904e\u77ed\u6703\u56e0\u70ba\u6709\u6548\u8cc7\u8a0a\u904e\u5c11\u800c\u4f7f\u5f97\u6a21\u578b\u8a13\u7df4\u904e \u7a0b\u7121\u6cd5\u6536\u6582\u6216\u8005\u662f\u5f71\u97ff\u5230\u7d50\u679c\u3002\u4f46\u5728\u6211\u5011\u7684\u7814\u7a76\u4e2d\uff0c\u8a9e\u53e5\u4e2d\u7684\u6bcf\u4e00\u500b\u97f3\u7d20\u5c0d\u61c9\u5230\u6f5b\u85cf\u72c4 \u5f0f\u5206\u4f48\u90fd\u53ef\u4ee5\u8996\u70ba\u5177\u6709\u8cc7\u8a0a\u7684\u55ae\u8a5e\uff0c\u56e0\u6b64\u5373\u4f7f\u53ea\u6709 2-4 \u500b\u5b57\u8a5e\u7684\u77ed\u53e5\u7684\u8a9e\u53e5\u4e5f\u5177\u6709\u8db3\u5920 \u7684\u8cc7\u8a0a\u4f86\u52a0\u4ee5\u5224\u5b9a\u5176\u7d44\u6210\u3002 (\u4e09)\u8a9e\u97f3\u4e8b\u4ef6\u964d\u7dad\u8207\u5408\u4f75\u8072\u5b78\u6a21\u578b\u4e4b\u9078\u64c7 1.\u8a9e\u97f3\u4e8b\u4ef6\u964d\u7dad \u5728\u672c\u6587\u4e2d\u5c07\u4e00\u6bb5\u8a9e\u53e5\u8996\u70ba\u4e00\u500b\u6587\u4ef6\uff0c\u56e0\u6b64\u6839\u64da\u6f5b\u85cf\u72c4\u5f0f\u5206\u4f48\u4e4b\u7269\u7406\u610f\u7fa9\uff1a\u82e5\u67d0\u500b \u5b57\u8a5e(Word)\u5f9e\u5c6c\u65bc\u67d0\u500b\u4e3b\u984c\uff0c\u5247\u7576\u67d0\u6587\u4ef6\u5c6c\u65bc\u67d0\u500b\u7279\u5b9a\u4e3b\u984c\u7684\u6642\u5019\uff0c\u5247\u5f9e\u5c6c\u65bc\u67d0\u500b\u4e3b\u984c \u7684\u5b57\u8a5e\u51fa\u73fe\u6b21\u6578\u6703\u8f03\u9ad8\u3002\u5c07\u4e3b\u984c\u5c0d\u61c9\u5230\u7684\u5247\u662f\u5b57\u8a5e(Word)\u3002\u56e0\u70ba\u5728\u7279\u5b9a\u7684\u8a9e\u8a00\u4e0b\uff0c\u82e5\u4ee5 \u5b57\u8a5e\u70ba\u55ae\u4f4d\uff0c\u6709\u4e9b\u97f3\u7d20\u7d93\u5e38\u6027\u7684\u6703\u51fa\u73fe\u5728\u4e00\u8d77\uff0c\u56e0\u6b64\u672c\u6587\u5c07\u539f\u72c4\u5f0f\u5206\u4f48\u4e4b\u4e3b\u984c\u6578\u8a2d\u5b9a\u70ba \u5b57\u8a5e\u6578\u91cf\u3002 \u4f46\u7531\u65bc\u5b57\u8a5e\u6578\u91cf\u904e\u591a\uff0c\u800c\u4e14\u6f5b\u85cf\u72c4\u5f0f\u5206\u4f48\u4e4b\u904b\u7b97\u6642\u9593\u8207\u4e3b\u984c\u6578 N \u6210\u6b63\u6bd4\uff0c\u82e5\u5c07\u6240 \u6709\u53ef\u80fd\u51fa\u73fe\u7684\u5b57\u8a5e\u6578\u91cf\u8a2d\u5b9a\u70ba\u4e3b\u984c\u6578\uff0c\u5176\u4e3b\u984c\u6578\u904e\u591a\u4e0d\u50c5\u5728\u904b\u7b97\u6642\u9593\u4e0a\u4e0d\u5141\u8a31\uff0c\u4e5f\u56e0\u70ba \u76ee\u6a19\u4e4b\u96c6\u5408\u904e\u5927\uff0c\u8a9e\u6599\u6703\u9762\u81e8\u56b4\u91cd\u4e0d\u8db3\u4e4b\u60c5\u6cc1\uff0c\u56e0\u6b64\u5be6\u969b\u4e0a\u4e26\u4e0d\u53ef\u884c\u3002\u56e0\u6b64\u70ba\u4e86\u89e3\u6c7a\u6b64 \u554f\u984c\uff0c\u6211\u5011\u5c07\u5b57\u7684\u7d44\u6210\u97f3\u7d20\u6839\u64da\u570b\u969b\u97f3\u6a19\u96c6\u5408\u4e4b\u5b9a\u7fa9\uff0c\u5c07\u6bcf\u500b\u97f3\u7d20\u62c6\u89e3\u6210\u767c\u97f3\u65b9\u6cd5 (Manner)\u8207\u767c\u97f3\u4f4d\u7f6e(Place)\uff0c\u800c\u767c\u97f3\u65b9\u6cd5\u548c\u767c\u97f3\u4f4d\u7f6e\u5728\u9019\u88e1\u5373\u662f\u8a9e\u97f3\u4e8b\u4ef6\u3002 2.\u5408\u4f75\u8072\u5b78\u6a21\u578b\u4e4b\u9078\u64c7 \u672c\u6587\u6839\u64da\u570b\u969b\u97f3\u6a19\u96c6\u5408\u5b9a\u7fa9\u5c07\u767c\u97f3\u90e8\u4f4d(Place)\u5206\u70ba 13 \u985e\uff0c\u800c\u767c\u97f3\u65b9\u6cd5(Manner)\u5b9a \u70ba 6 \u985e\uff0c\u800c\u6bcd\u97f3\u5247\u7531\u820c\u9762\u524d\u5f8c\u5171 5 \u985e\u3001\u820c\u9762\u9ad8\u4f4e\u5171 7 \u985e\u3001\u5507\u5f62\u5171 2 \u985e\u3002\u820c\u9762\u524d\u5f8c\u70ba\u524d (Front)\u3001\u6b21\u524d(Near-front)\u3001\u592e (Central)\u3001\u6b21\u5f8c(Near-back)\u3001\u5f8c (Back)\u3002\u820c\u9762\u9ad8\u4f4e\u70ba\u9589(Close)\u3001 \u6b21\u9589(Near-close) \u3001\u534a\u9589(Close-mid) \u3001\u4e2d (Mid) \u3001\u534a\u958b(Open-mid) \u3001\u6b21\u958b(Near-open) \u3001\u958b (Open) \u3002 \u8123\u5f62\u70ba\u5713\u5507(Rounded)\u8207\u975e\u5713\u5507(Unrounded)\u3002 \u6211\u5011\u6703\u5148\u5c07\u767c\u97f3\u65b9\u6cd5\u548c\u767c\u97f3\u90e8\u4f4d\u4f7f\u7528\u4e0d\u540c\u7684\u6f5b\u85cf\u72c4\u5f0f\u5206\u4f48\u5075\u6e2c\u5668\u5206\u5225\u5075\u6e2c\u4e26\u5206\u7fa4\uff0c \u6700\u5f8c\u518d\u5c07\u767c\u97f3\u90e8\u4f4d\u76f8\u540c\u4f46\u767c\u97f3\u65b9\u6cd5\u4e0d\u540c\u4ee5\u53ca\u767c\u97f3\u90e8\u4f4d\u4e0d\u540c\u4f46\u767c\u97f3\u65b9\u6cd5\u76f8\u540c\u7684\u97f3\u7d20\u8996\u70ba \u4e0d\u540c\u5206\u7fa4\uff0c\u63db\u53e5\u8a71\u8aaa\u5247\u662f\u53ea\u7559\u4e0b\u767c\u97f3\u90e8\u4f4d\u8207\u767c\u97f3\u65b9\u6cd5\u7686\u5206\u985e\u518d\u4e00\u8d77\u7684\u97f3\u7d20\u9032\u884c\u5408\u4f75\u3002 \u4e94\u3001\u5be6\u9a57\u8a2d\u8a08\u8207\u5206\u6790 (\u4e00) \u5be6\u9a57\u8a9e\u6599\u8207\u5de5\u5177 \u53f0\u8a9e\u5be6\u9a57\u8a9e\u6599\u4f7f\u7528\u826f\u654f\u96c4\u535a\u58eb\u6240\u9304\u88fd\u4e4b\u8a9e\u6599\uff0c\u8a9e\u6599\u70ba 16kHz, 16bit \u4e4b\u9ea5\u514b\u98a8\u8a9e\u6599\u5171 \u6bcf\u6b21\u4f4d\u79fb\u55ae\u4f4d(Shift)\u70ba 10ms\u3002\u96b1\u85cf\u5f0f\u99ac\u53ef\u592b\u6a21\u578b\u6bcf\u500b\u72c0\u614b(State)\u7559\u4e0b 16 \u500b\u8def\u5f91\uff0c\u6700\u5f8c \u518d\u53d6\u524d 10 \u540d\u3002\u6240\u6709\u8a9e\u6599\u7686\u7121\u6642\u9593\u6a19\u8a18\uff0c\u6642\u9593\u65b7\u9ede\u4f7f\u7528 Baum-Welch algorithms \u8a08\u7b97\u3002 (\u4e8c)\u5be6\u9a57\u9805\u76ee 1.\u8a55\u4f30\u65b9\u5f0f \u8981\u5206\u6790\u8a9e\u97f3\u8fa8\u8b58\u7684\u6b63\u78ba\u7387\u5fc5\u9808\u6e96\u78ba\u7684\u8fa8\u8b58\u51fa\u6b63\u78ba\u7684\u5b57\u8a5e\u3002\u8fa8\u8b58\u7d50\u679c\u8207\u8a9e\u6599\u6a19\u8a18\u4e92 \u76f8\u6bd4\u8f03\u5f8c\uff0c\u7d50\u679c\u8207\u6a19\u8a18\u5b8c\u5168\u76f8\u7b26\u70ba\u8fa8\u8b58\u6b63\u78ba(Correct)\uff0c\u8207\u6a19\u8a18\u4e0d\u540c\u7684\u932f\u8aa4\u5247\u6839\u64da\u5b9a\u7fa9\u5206 \u985e\u6210\u4e0b\u5217\u5e7e\u7a2e\u932f\u8aa4\uff1a \u5176\u4e2d N \u70ba\u8a9e\u6599\u6a19\u8a18\u5167\u6240\u6709\u7684\u97f3\u7d20\u7e3d\u6578\uff0cS \u70ba\u53d6\u4ee3\u932f\u8aa4\u6578\u91cf\uff0cD \u70ba\u522a\u9664\u932f\u8aa4\u6578\u91cf\uff0cI \u70ba\u63d2 \u5165\u932f\u8aa4\u6578\u91cf\uff0cC \u70ba\u8fa8\u8b58\u6b63\u78ba\u5b57\u8a5e\u6578\u91cf\uff0cC=N-D-S\u3002 2.\u591a\u8a9e\u74b0\u5883\u4e0b\u6f5b\u85cf\u72c4\u5f0f\u5206\u4f48\u5408\u4f75\u8072\u5b78\u6a21\u578b\u9078\u64c7\u4e4b\u9a57\u8b49 \u70ba\u4e86\u9a57\u8b49\u6f5b\u85cf\u72c4\u5f0f\u5206\u4f48\u5408\u4f75\u8072\u5b78\u6a21\u578b\u5728\u591a\u8a9e\u74b0\u5883\u4e0b\u4e4b\u6548\u80fd\uff0c\u56e0\u6b64\u6211\u5011\u540c\u6a23\u5206\u5225\u5c0d\u672a \u805a\u985e\u4e09\u9023\u97f3\u7d20\u6a21\u578b\u3001\u5df2\u805a\u985e\u4e09\u9023\u97f3\u7d20\u6a21\u578b\u548c\u7531\u6f5b\u85cf\u72c4\u5f0f\u5206\u4f48\u5408\u4f75\u8072\u5b78\u6a21\u578b\u5728\u570b\u53f0\u82f1\u4e09\u8a9e \u74b0\u5883\u4e0b\u9032\u884c\u9a57\u8b49\u3002\u6211\u5011\u5c07\u6703\u5206\u5225\u4f7f\u7528\u5df2\u591a\u8a9e\u8072\u5b78\u6a21\u578b\u8fa8\u8b58\u4e09\u8a9e\u6df7\u5408\u3001\u570b\u8a9e\u3001\u53f0\u8a9e\u548c\u82f1\u8a9e \u5be6\u9a57\u8cc7\u6599\u3002\u4ee5\u9a57\u8b49\u5728\u6c92\u6709\u8a9e\u8a00\u8fa8\u8b58\u7cfb\u7d71\u4e0b\u591a\u8a9e\u8072\u5b78\u6a21\u578b\u5c0d\u65bc\u6bcf\u4e00\u7a2e\u8a9e\u8a00\u4e4b\u6548\u80fd\u3002 \u5be6\u9a57\u74b0\u5883\u548c\u5de5\u5177\u8207\u55ae\u4e00\u8a9e\u8a00\u9032\u884c\u4e4b\u5be6\u9a57\u76f8\u540c\uff0c\u8a13\u7df4\u8a9e\u6599\u548c\u4e09\u8a9e\u6df7\u5408\u4e4b\u6e2c\u8a66\u8cc7\u6599\u70ba\u5c07 TCC300\u3001TIMIT \u548c\u6881\u654f\u96c4\u535a\u58eb\u6240\u9304\u88fd\u4e4b\u53f0\u8a9e\u8a9e\u6599\u6df7\u5408\u4f7f\u7528\uff0c\u5176\u5be6\u55ae\u8a9e\u6e2c\u8a66\u8cc7\u6599\u5247\u548c\u4e0a \u8ff0\u8a55\u4f30\u65b9\u5f0f\u6240\u9032\u884c\u7684\u5be6\u9a57\u4f7f\u7528\u76f8\u540c\u7684\u6e2c\u8a66\u8a9e\u6599\u3002\u8a9e\u6599\u5167\u55ae\u4e00\u8a9e\u53e5\u53ea\u6709\u4e00\u7a2e\u8a9e\u8a00\uff0c\u4e5f\u5c31\u662f \u4e0d\u505a\u55ae\u4e00\u8a9e\u53e5\u591a\u7a2e\u8a9e\u8a00\u6df7\u5408\u4e4b\u5be6\u9a57\u3002\u6240\u6709\u8a9e\u8a00\u7684\u97f3\u7d20\u7686\u4ee5 IPA \u4f86\u8868\u793a\u3002 (1)\u570b\u53f0\u82f1\u4e09\u8a9e\u6df7\u5408\u5be6\u9a57 \u7d50\u679c\u5982\u5716\u4e09\u6240\u793a\uff0c\u5728\u591a\u8a9e\u74b0\u5883\u4e0b\u6f5b\u85cf\u72c4\u5f0f\u5206\u4f48\u5408\u4f75\u7684\u8072\u5b78\u6a21\u578b\u5728 Top5 \u6642\u5019\u6709 47.86%\u7684\u6700\u9ad8\u6e96\u78ba\u7387\uff0c\u4e14\u6574\u9ad4\u6e96\u78ba\u7387\u7686\u9ad8\u65bc\u5176\u4ed6\u5169\u7a2e\u8072\u5b78\u6a21\u578b\u3002\u5728 Top5 \u6642\u7684\u63d2\u5165\u932f\u8aa4\uff0c \u672c\u6587\u63d0\u51fa\u4e4b\u65b9\u6cd5\u8f03\u5df2\u805a\u5408\u4e09\u9023\u97f3\u7d20\u6a21\u578b\u591a 0.24%\uff0c\u4f46\u53d6\u4ee3\u932f\u8aa4\u548c\u522a\u9664\u932f\u8aa4\u5247\u8f03\u5df2\u805a\u5408\u4e09 \u9023\u97f3\u7d20\u6a21\u578b\u5206\u5225\u5c11\u4e86 1.02%\u548c 0.74%\uff0c\u56e0\u6b64\u6574\u9ad4\u7684\u6e96\u78ba\u7387\u8f03\u5df2\u4e09\u9023\u97f3\u7d20\u6a21\u578b\u591a\u4e86 1.52%\u3002 (2)\u570b\u8a9e\u5be6\u9a57 \u5982\u5716\u56db\u6240\u793a\uff0c\u672a\u805a\u985e\u4e09\u9023\u97f3\u7d20\u6a21\u578b\u5728\u591a\u8a9e\u74b0\u5883\u4e0b\u8fa8\u8b58\u570b\u8a9e\u6709\u6700\u9ad8\u7684\u6e96\u78ba\u7387\u3002\u800c\u6709\u7d93 \u904e\u6a21\u578b\u5408\u4f75\u7684\u805a\u985e\u4e09\u9023\u97f3\u7d20\u6a21\u578b\u548c\u6f5b\u85cf\u72c4\u5f0f\u5206\u4f48\u9078\u64c7\u8072\u5b78\u6a21\u578b\u64c7\u662f\u6b63\u78ba\u7387\u5dee\u8ddd\u4e0d\u5927\uff0c\u4f46 \u5169\u8005\u7686\u4f4e\u65bc\u672a\u805a\u985e\u4e09\u9023\u97f3\u7d20\u6a21\u578b\u3002\u800c\u672a\u805a\u985e\u4e09\u9023\u97f3\u7d20\u6a21\u578b\u932f\u8aa4\u7387\u6539\u5584\u4e3b\u8981\u96c6\u4e2d\u5728\u53d6\u4ee3\u932f \u8aa4\u3002\u6f5b\u5728\u72c4\u5f0f\u5206\u4f48\u548c\u805a\u985e\u4e09\u9023\u97f3\u7d20\u6a21\u578b\u7684\u6bd4\u8f03\u5728\u63d2\u5165\u932f\u8aa4\u7565\u9ad8\u800c\u522a\u9664\u932f\u8aa4\u4f4e\u5247\u662f\u8207\u672c\u6587 \u524d\u5217\u55ae\u8a9e\u74b0\u5883\u4e0b\u4e4b\u5be6\u9a57\u76f8\u540c\u3002\u56e0\u6b64\u63a8\u6e2c\u6709\u53ef\u80fd\u662f\u5728\u9032\u884c\u591a\u8a9e\u6a21\u578b\u5408\u4f75\u6642\u5f8c\u570b\u8a9e\u7684\u97f3\u6a19\u5408 \u4f75\u4e0a\u51fa\u73fe\u554f\u984c\u6216\u8005\u662f\u56e0\u70ba\u8a13\u7df4\u8a9e\u6599\u904e\u5c11\u800c\u5c0e\u81f4\u6a21\u578b\u63cf\u8ff0\u4e0d\u5920\u5168\u9762\uff0c\u56e0\u6b64\u5728\u591a\u8a9e\u74b0\u5883\u4e0b\u7522 \u751f\u4e86\u56b4\u91cd\u7684\u6a21\u578b\u6df7\u6dc6\u3002 42.88% 45.06% 46.94% 48.08% 48.99% 38.37% 40.75% 42.42% 43.47% 44.04% 38.95% 40.64% 42.07% 43.23% 44.44% 0.00% 10.00% 20.00% 30.00% 40.00% 50.00% 60.00% Top1 Top2 Top3 Top4 Top5 Accuracy Triphone Tied Triphone LDA Tied Triphone \u5716\u4e94\u3001\u591a\u8a9e\u8072\u5b78\u6a21\u578b\u8fa8\u8b58\u53f0\u8a9e\u4e4b\u8fa8\u8b58\u7d50\u679c (4)\u82f1\u8a9e\u5be6\u9a57 \u5716\u516d\u3001\u591a\u8a9e\u8072\u5b78\u6a21\u578b\u8fa8\u8b58\u82f1\u8a9e\u4e4b\u8fa8\u8b58\u7d50\u679c 31.67% 37.21% 41.17% 43.40% 45.21% 48.00% 50.73% 52.81% 53.02% \u516d\u3001\u7d50\u8ad6\u8207\u672a\u4f86\u7814\u7a76\u767c\u5c55\u65b9\u5411 60.00% 51.00% 48.58% 44.56% 50.00% \u7531\u65bc\u76ee\u524d\u7db2\u8def\u8207\u4ea4\u901a\u7684\u767c\u9054\uff0c\u4f7f\u5f97\u5168\u7403\u5316\u6210\u70ba\u5fc5\u7136\u7684\u8da8\u52e2\uff0c\u800c\u591a\u8a9e\u8fa8\u8b58\u5728\u9019\u500b\u793e\u6703 44.23% 38.67% \u4e5f\u6108\u4f86\u6108\u986f\u5f97\u91cd\u8981\u3002\u4f46\u6709\u8a9e\u8a00\u8fa8\u8b58\u7684\u591a\u8a9e\u8fa8\u8b58\u5b58\u5728\u8457\u932f\u8aa4\u758a\u52a0\u7684\u554f\u984c\u3002\u4e0d\u4f7f\u7528\u8a9e\u8a00\u8fa8\u8b58 38.44% 0.00% 10.00% 20.00% 30.00% 40.00% Top1 Top2 Top3 Top4 Top5 \u7684\u591a\u8a9e\u8fa8\u8b58\u65b9\u6cd5\u88ab\u63d0\u51fa\u7528\u4f86\u89e3\u6c7a\u6b64\u554f\u984c\uff0c\u4f46\u4e0d\u4f7f\u7528\u8a9e\u8a00\u8fa8\u8b58\u7684\u65b9\u6cd5\u6709\u4e0d\u540c\u8a9e\u8a00\u9593\u7684\u8072\u5b78 \u6a21\u578b\u9593\u5bb9\u6613\u6df7\u6dc6\u7684\u554f\u984c\u5b58\u5728\uff0c\u800c\u672c\u6587\u63d0\u51fa\u4f7f\u7528\u6f5b\u5728\u72c4\u5f0f\u5206\u4f48\u4f86\u9032\u884c\u8072\u5b78\u6a21\u578b\u5408\u4f75\u4e4b\u9078\u64c7\u3002 Accuracy Triphone Tied Triphone LDA Tied Triphone 0.00% Top1 Top2 Top3 Top4 Top5 Triphone Tied Triphone LDA Tied Triphone \u6240\u6709\u8a9e\u8a00\u97f3\u7d20\u7686\u4ee5 IPA \u8868\u793a\uff0c\u4ee5\u9054\u5230\u53c3\u6578\u5171\u901a\u4e4b\u76ee\u7684\u3002\u4e26\u4e14\u5c07\u97f3\u7d20\u5c0d\u61c9\u5230\u767c\u97f3\u90e8\u4f4d\u548c\u767c \u97f3\u65b9\u6cd5\u4ee5\u6e1b\u5c11\u6578\u91cf\uff0c\u4e26\u4e14\u9054\u5230\u6e1b\u5c11\u904b\u7b97\u91cf\u4e4b\u76ee\u7684\uff0c\u4f7f\u5f97\u6f5b\u85cf\u72c4\u5f0f\u5206\u4f48\u5075\u6e2c\u5668\u5f97\u4ee5\u904b\u7528\u9577 \u8ddd\u96e2\u8a5e\u5f59\u8a9e\u610f\u8cc7\u8a0a\uff0c\u5c07\u5e38\u5e38\u5148\u5f8c\u6210\u5c0d\u51fa\u73fe\u7684\u97f3\u7d20\u9032\u884c\u5408\u4f75\uff0c\u4ee5\u6e1b\u5c11\u8072\u5b78\u6a21\u578b\u7684\u6df7\u6dc6\u3002 \u5be6\u9a57\u7d50\u679c\u986f\u793a\uff0c\u7531\u6f5b\u85cf\u72c4\u5f0f\u5206\u4f48\u6240\u9078\u64c7\u5408\u4f75\u7684\u8072\u5b78\u6a21\u578b\u548c\u805a\u985e\u4e09\u9023\u97f3\u7d20\u6a21\u578b\u4ee5\u53ca\u672a \u805a\u985e\u4e09\u9023\u97f3\u7d20\u6a21\u578b\u76f8\u6bd4\u5728\u55ae\u8a9e\u74b0\u5883\u4e0b\u8fa8\u8b58\u570b\u8a9e\u6700\u9ad8\u6709 10.16%\u7684\u6e96\u78ba\u7387\u6539\u5584\uff0c\u800c\u8fa8\u8b58\u82f1 \u8a9e\u5247\u6709 4.23%\u7684\u6e96\u78ba\u7387\u6539\u5584\u3002\u5728\u591a\u8a9e\u74b0\u5883\u4e0b\u6df7\u5408\u8fa8\u8b58\u570b\u53f0\u82f1\u4e09\u8a9e\u6df7\u5408\u7684\u60c5\u6cc1\u4e0b\u6709 0.24% \u7684\u6e96\u78ba\u5ea6\u6539\u5584\uff0c\u8fa8\u8b58\u82f1\u8a9e\u6709 3.16%\u7684\u6e96\u78ba\u5ea6\u6539\u5584\u3002 \u672c\u6587\u6240\u63d0\u51fa\u4e4b\u65b9\u6cd5\u96d6\u7136\u5728\u591a\u6578\u7684\u60c5\u6cc1\u4e0b\u90fd\u53ef\u4ee5\u7372\u5f97\u6e96\u78ba\u7387\u7684\u6539\u5584\uff0c\u4f46\u662f\u6574\u9ad4\u8fa8\u8b58\u7387 \u4ecd\u7136\u504f\u4f4e\uff0c\u591a\u8a9e\u8fa8\u8b58\u76f8\u8f03\u65bc\u55ae\u8a9e\u8fa8\u8b58\u7684\u8072\u5b78\u6a21\u578b\u6578\u91cf\u6703\u96a8\u8457\u8a9e\u8a00\u6578\u91cf\u6210\u500d\u6578\u6210\u9577\uff0c\u800c\u8072 \u5b78\u6a21\u578b\u6578\u91cf\u6108\u591a\u5247\u9700\u8981\u66f4\u591a\u7684\u8a13\u7df4\u8a9e\u6599\u4f86\u8a13\u7df4\u3002\u56e0\u6b64\u591a\u8a9e\u8fa8\u8b58\u7d93\u5e38\u9762\u81e8\u8a9e\u6599\u4e0d\u8db3\u7684\u554f\u984c\uff0c \u96d6\u7136\u7d93\u904e\u6a21\u578b\u5408\u4f75\u5f8c\u9a0e\u8072\u5b78\u6a21\u578b\u6578\u91cf\u4ecd\u7136\u76f8\u7576\u9f90\u5927\u3002\u672c\u6587\u6240\u63d0\u51fa\u4e4b\u65b9\u6cd5\u6700\u7d42\u5408\u4f75\u5f8c\u7684\u8072 126000 \u53e5\uff0c\u6a19\u8a18\u70ba ForPA\uff0c\u5176\u4e2d\u5171\u6709\u97fb\u6bcd 9 \u96b1\u85cf\u99ac\u53ef\u592b\u8072\u5b78\u6a21\u578b\u8a13\u7df4\u4f7f\u7528 39 \u7dad\u7684\u6885\u723e\u5012\u983b\u8b5c\u7cfb\u6578\uff0c\u97f3\u6846(frame)\u5927\u5c0f\u70ba 20ms\uff0c (1)\u53d6\u4ee3\u932f\u8aa4(Substitution Errors)\uff1a\u5c07\u6b63\u78ba\u7684\u97f3\u7d20\u66ff\u63db\u6210\u5176\u4ed6\u97f3\u7d20\u3002 (2)\u522a\u9664\u932f\u8aa4(Deletion Errors)\uff1a\u6c92\u6709\u5c07\u8a72\u8fa8\u8b58\u7684\u97f3\u7d20\u8fa8\u8b58\u51fa\u4f86\u3002 (3)\u63d2\u5165\u932f\u8aa4(Insertion Errors)\uff1a\u672c\u4f86\u6c92\u6709\u7684\u97f3\u7d20\u537b\u984d\u5916\u8fa8\u8b58\u51fa\u4f86\u3002 \u70ba\u4e86\u8a55\u4f30\u672c\u7cfb\u7d71\u4e4b\u6548\u80fd\uff0c\u6211\u5011\u5c07\u5206\u958b\u5206\u6790\u53d6\u4ee3\u932f\u8aa4\u7387\u3001\u522a\u9664\u932f\u8aa4\u7387\u3001\u63d2\u5165\u932f\u8aa4\u7387\u4ee5\u53ca\u5b57 \u932f\u8aa4\u7387(Word Error Rate, WER)\u548c\u6e96\u78ba\u5ea6(Accuracy)\u3002\u8a08\u7b97\u516c\u5f0f\u5982\u4e0b\u5f0f\uff1a Substitution Errors Rate = (\u5f0f 5.1) Deletion Errors Rate = (\u5f0f 5.2) Insertion Errors Rate = \u56e0\u6b64\u53ef\u4ee5\u9a57\u8b49\u672c\u6587\u6240\u63d0\u51fa\u4e4b\u65b9\u6cd5\u5728\u591a\u8a9e\u74b0\u5883\u4e0b\u4ecd\u7136\u6709\u8f03\u4f73\u7684\u6548\u80fd\u3002 34.88% 38.96% 41.80% 43.43% 44.73% 36.23% 40.33% 43.05% 44.83% 46.34% 37.89% 42.09% 44.89% 60.00% \u5716\u56db\u3001\u591a\u8a9e\u8072\u5b78\u6a21\u578b\u8fa8\u8b58\u570b\u8a9e\u4e4b\u8fa8\u8b58\u7d50\u679c \u5b78\u6a21\u578b\u6578\u91cf\u9060\u5927\u65bc\u805a\u985e\u4e09\u9023\u97f3\u7d20\u6a21\u578b\uff0c\u56e0\u6b64\u9700\u8981\u66f4\u591a\u7684\u8a13\u7df4\u8a9e\u6599\u3002\u672a\u4f86\u6211\u5011\u6703\u6301\u7e8c\u7684\u6536 \u672c\u6587\u6240\u63d0\u51fa\u4e4b\u65b9\u6cd5\u5728\u591a\u8a9e\u74b0\u5883\u4e0b\u4ee5\u8fa8\u8b58\u82f1\u8a9e\u6e96\u78ba\u5ea6\u63d0\u5347\u7387\u9060\u9ad8\u65bc\u570b\u8a9e\u548c\u53f0\u8a9e\uff0c\u5982\u5716 \u516d\u6240\u793a\uff0c\u5728 Top5 \u6642\u6e96\u78ba\u5ea6\u8f03\u805a\u985e\u4e09\u9023\u97f3\u7d20\u6a21\u578b\u9ad8 3.16%\uff0c\u8207\u672a\u805a\u985e\u4e09\u9023\u97f3\u7d20\u6a21\u578b\u76f8\u6bd4 \u96c6\u5927\u91cf\u7684\u8a9e\u6599\uff0c\u8b93\u8a13\u7df4\u8cc7\u6599\u66f4\u70ba\u5b8c\u5584\uff0c\u4ee5\u6539\u5584\u8a9e\u6599\u4e0d\u8db3\u7684\u554f\u984c\u3002 (3)\u53f0\u8a9e\u5be6\u9a57 47.86% \u5247\u662f\u9ad8 5.98%\u3002\u5982\u8868 5.9 \u6240\u793a\uff0c\u5728 Top5 \u60c5\u6cc1\u4e0b\uff0c\u53d6\u4ee3\u932f\u8aa4\u76f8\u8f03\u65bc\u805a\u985e\u4e09\u9023\u97f3\u7d20\u6a21\u578b\u6709 \u76ee\u524d\u6240\u63d0\u51fa\u4e4b\u65b9\u6cd5\u53ea\u4f7f\u7528\u9577\u8ddd\u96e2\u8a5e\u5f59\u8a9e\u610f\u8cc7\u8a0a\u4f86\u9032\u884c\u5408\u4f75\uff0c\u56e0\u6b64\u7576\u53e5\u5b50\u9577\u5ea6\u8f03\u77ed\u7684 46.45% 10.00% 20.00% 30.00% 40.00% 50.00% 3.2%\u7684\u6539\u5584\uff0c\u522a\u9664\u932f\u8aa4\u5247\u662f\u6709 0.63%\u7684\u6539\u5584\uff0c\u4f46\u63d2\u5165\u932f\u8aa4\u5247\u662f\u8f03\u805a\u985e\u4e09\u9023\u97f3\u7d20\u6a21\u578b\u9ad8 \u6642\u5019\u6e96\u78ba\u5ea6\u6703\u6709\u76f8\u7576\u7a0b\u5ea6\u7684\u4e0b\u964d\uff0c\u56e0\u6b64\u672a\u4f86\u82e5\u53ef\u4ee5\u540c\u6642\u8003\u616e\u77ed\u8ddd\u96e2\u7684\u8cc7\u6599\u7279\u6027\uff0c\u4f8b\u5982\u4e09 \u591a\u8a9e\u8072\u5b78\u6a21\u578b\u8fa8\u8b58\u53f0\u8a9e\u7684\u74b0\u5883\u4e0b\uff0c\u6f5b\u85cf\u72c4\u5f0f\u5206\u4f48\u9078\u64c7\u4e4b\u8072\u5b78\u6a21\u578b\u6e96\u78ba\u5ea6\u5728 Top5 \u6642 0.61%\u3002\u56e0\u6b64\u53ef\u4ee5\u770b\u51fa\u672c\u6587\u63d0\u51fa\u4e4b\u65b9\u6cd5\u5728\u591a\u8a9e\u74b0\u5883\u4e0b\u8fa8\u8b58\u82f1\u6587\u53ef\u4ee5\u6e1b\u5c11\u53d6\u4ee3\u932f\u8aa4\u548c\u522a\u9664 \u9023\u97f3\u7d20\u6a21\u578b\u805a\u5408\u6642\u6240\u4f7f\u7528\u7684\u99ac\u5f0f\u8ddd\u96e2\uff0c\u6216\u8a31\u53ef\u4ee5\u540c\u6642\u5728\u77ed\u53e5\u548c\u9577\u53e5\u90fd\u53d6\u5f97\u8f03\u4f73\u7684\u7d50\u679c\u3002 \u7565\u9ad8\u805a\u985e\u4e09\u9023\u97f3\u7d20\u6a21\u578b 0.21%\uff0c\u800c\u9ad8\u672a\u805a\u985e\u4e09\u9023\u97f3\u7d20\u6a21\u578b 7.81%\uff0c\u5982\u5716\u4e94\u6240\u793a\u3002\u5728 Top5 \u932f\u8aa4\uff0c\u800c\u6574\u9ad4\u7684\u6e96\u78ba\u7387\u4e5f\u6709\u6240\u63d0\u5347\u3002 \u6642\u53d6\u4ee3\u932f\u8aa4\u8f03\u805a\u985e\u4e09\u9023\u97f3\u7d20\u6a21\u578b\u4f4e 0.23%\u3001\u522a\u9664\u932f\u8aa4\u591a 0.02%\u800c\u63d2\u5165\u932f\u8aa4\u5247\u662f\u76f8\u540c\uff0c\u6574 Accuracy \u9ad4\u770b\u4f86\u4ecd\u7565\u512a\u65bc\u805a\u985e\u4e09\u9023\u97f3\u7d20\u6a21\u578b\u3002\u5728 5.2.2 \u7684\u55ae\u8a9e\u74b0\u5883\u4e0b\u53f0\u8a9e\u8072\u5b78\u6a21\u578b\u5be6\u9a57\u4e4b\u7d50\u679c\u986f \u793a\u6f5b\u85cf\u72c4\u5f0f\u5206\u4f48\u9078\u64c7\u7684\u8072\u5b78\u6a21\u578b\u6e96\u78ba\u7387\u7565\u4f4e\u65bc\u805a\u985e\u4e09\u9023\u97f3\u7d20\u6a21\u578b\uff0c\u800c\u5728\u591a\u8a9e\u74b0\u5883\u4e0b\u4e4b\u6e96 \u78ba\u7387\u96d6\u7136\u7565\u6709\u6539\u5584\uff0c\u4f46\u4f9d\u7136\u8207\u805a\u985e\u4e09\u9023\u97f3\u7d20\u6a21\u578b\u5dee\u8ddd\u4e0d\u5927\u3002\u56e0\u6b64\u6211\u5011\u63a8\u6e2c\u672c\u6587\u6240\u63d0\u51fa\u7684 \u65b9\u6cd5\u53ef\u80fd\u5728\u6211\u5011\u6240\u4f7f\u7528\u7684\u53f0\u8a9e\u8a9e\u6599\u74b0\u5883\u4e0b\u6548\u80fd\u6539\u5584\u6709\u9650\uff0c\u4f46\u5728\u591a\u8a9e\u7684\u60c5\u6cc1\u4e0b\u4ecd\u7136\u53ef\u4ee5\u7565 \u9ad8\u65bc\u805a\u985e\u4e09\u9023\u97f3\u7d20\u6a21\u578b\u3002 31.14% 34.46% 36.19% 37.01% 38.05% 37.15% 39.19% 40.80% 40.15% 43.19% 44.03% 50.00% \u81f4\u8b1d 42.35% 39.79% 36.49% 40.00% \u672c\u7814\u7a76\u627f\u8499\u4e2d\u83ef\u6c11\u570b\u570b\u5bb6\u79d1\u5b78\u59d4\u54e1\u6703\u7d93\u8cbb(99-2221-E-415-006)\u652f\u6301\u65b9\u5f97\u4ee5\u5b8c\u6210\uff0c\u7279 33.59% \u5225\u611f\u8b1d\u3002 (\u5f0f 5.3) Word Error Rate = + + (\u5f0f 5.4) Accuracy = \u2212 (\u5f0f 5.5) \u5716\u4e09\u3001\u4e09\u7a2e\u8072\u5b78\u6a21\u578b\u5728\u570b\u53f0\u82f1\u6df7\u5408\u4e0b\u4e4b\u8fa8\u8b58\u7d50\u679c 0.00% Top1 Top2 Top3 Top4 Top5 Triphone Tied Triphone LDA Tied Triphone 10.00% 20.00% 30.00% Accuracy \u53c3\u8003\u6587\u737b</td></tr></table>",
"text": "\u7a2e\u3001\u8072\u6bcd 18 \u7a2e\u548c\u9f3b\u97f3\u5c3e\u8072\u6bcd 5 \u7a2e\u3002\u82f1\u6587\u4f7f \u7528\u8a9e\u6599\u70ba\u6210\u5927\u9304\u88fd\u7684\u9ea5\u514b\u98a8\u8a9e\u6599\u5171 808 \u53e5\u3002\u570b\u8a9e\u5be6\u9a57\u8a9e\u6599\u5f9e TCC300 \u4e2d\u9078\u51fa 16kHz, 16bit \u4e4b\u9ea5\u514b\u98a8\u8a9e\u6599\u5171 2676 \u53e5\u3002\u82f1\u6587\u8a9e\u6599\u4f7f\u7528 TIMIT\uff0c\u8a9e\u6599\u70ba 16kHz, 16bit \u652f\u9ea5\u514b\u98a8\u8a9e\u6599\u5171 4620 \u53e5\u3002\u7279\u5fb5\u4f7f\u7528\u6885\u723e\u5012\u983b\u8b5c\u7cfb\u6578(Mel-scale Frequency Cepstral Coefficients, MFCC)\u3001 \u96b1\u85cf\u99ac\u53ef\u592b(HMM)\u8072\u5b78\u6a21\u578b\u8a13\u7df4\u548c\u8072\u5b78\u6a21\u578b\u5408\u4f75\u90e8\u5206\u5247\u662f\u4f7f\u7528\u82f1\u570b\u528d\u6a4b\u5927\u5b78 HTK toolkit \u4f86\u5efa\u7acb\u3002\u6f5b\u85cf\u5f0f\u72c4\u5f0f\u5206\u4f48\u6a21\u578b\u8a13\u7df4\u5247\u662f\u4ee5\u539f\u4f5c\u8005 Blei et al.\u7684 ToolKit \u70ba\u57fa\u790e\u4f86\u9032 \u884c\u4fee\u6539\u3002",
"type_str": "table"
}
}
}
} |