File size: 86,080 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 |
{
"paper_id": "O09-1019",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T08:11:18.925033Z"
},
"title": "\u5f37\u5065\u6027\u8a9e\u97f3\u8fa8\uf9fc\u4e2d\u57fa\u65bc\u5c0f\u6ce2\u8f49\u63db\u4e4b\u5206\u983b\u7d71\u8a08\u88dc\u511f\u6280\u8853\u7684\u7814\u7a76 A Study of Sub-band Feature Statistics Compensation Techniques Based on a Discrete Wavelet Transform for Robust Speech Recognition",
"authors": [
{
"first": "Hao-Teng",
"middle": [],
"last": "\u8303\u9865\u9a30",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "",
"middle": [],
"last": "Fan",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Wen-Hsiang",
"middle": [],
"last": "\u675c\u6587\u7965",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "",
"middle": [],
"last": "Tu",
"suffix": "",
"affiliation": {},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "The environ mental m ismatch caused by additiv e noise and/or channel distortion often degrades th e perform ance of a s peech reco gnition sys tem seriously. V arious ro bustness techniques have been proposed to reduce this mismatch, and one category of them aim s t o normalize the statistics of speech fea tures in bo th training and testing conditions. In general, these statistics norm alization methods deal with the sp eech feature sequ ences in a f ull-band manner, which som ewhat ignores the fact th at dif ferent m odulation frequency com ponents",
"pdf_parse": {
"paper_id": "O09-1019",
"_pdf_hash": "",
"abstract": [
{
"text": "The environ mental m ismatch caused by additiv e noise and/or channel distortion often degrades th e perform ance of a s peech reco gnition sys tem seriously. V arious ro bustness techniques have been proposed to reduce this mismatch, and one category of them aim s t o normalize the statistics of speech fea tures in bo th training and testing conditions. In general, these statistics norm alization methods deal with the sp eech feature sequ ences in a f ull-band manner, which som ewhat ignores the fact th at dif ferent m odulation frequency com ponents",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "have unequal importance for speech recognition.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "With the above observations, in this paper we propose that the speech feature streams be proce ssed in a sub-band ma nner. The processed temporal-domain feature sequence is first decomposed into non-uniform sub-bands us ing discrete wavelet transform (DWT), and then each sub-band stream is individuall y processed by the well-known normalization methods, like m ean and variance norm alization (MVN) and histogram equalization (HEQ) . Finally, we reconstruct the feature stream w ith all th e modif ied sub-band streams using inverse D WT. W ith this process, the com ponents that correspond to m ore important modulation spectral bands in the feature sequ ence can be processed separately . For the Aurora-2 clean-condition training task, the new proposed su b-band MVN and HEQ provide relative error rate reductions of 20.32% a nd 16.39% over the conventional MVN a nd HEQ, respectively. These results re veal that the proposed m ethods significantly enhance the robustness of speech features in noise-corrupted environments. ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "( ) H H z \u8207 ( ) L H z \u8868\u793a\u70ba\u5206\u6790(analysis) \uf984\u6ce2\u5668\u4e4b\u9ad8\u901a\u8207\u4f4e\u901a\u7684\u8f49\u63db\u51fd\uf969(transfer function)\uff0c ( ) H G z \u8207 ( ) L G z \u5247\u70ba\u5408\u6210(synthesis)\uf984\u6ce2\u5668\u4e4b\u9ad8\u901a\u8207\u4f4e\u901a\u7684\u8f49\u63db\u51fd\uf969\uff0c\u4e14\u5b83\u5011 \u9808\u7b26\u5408\u4ee5\u4e0b\u7684\u689d\u4ef6\uff1a ( ) ( ) L H G z H z = , ( ) ( ) H L G z H z = \u2212 \u2212 (3-1) \u800c\u5728\u9ad8\u901a\u8207\u4f4e\u901a\u5206\u6790\uf984\u6ce2\u5668\u4e4b\u9593\u5b58\u6709\u4ee5\u4e0b\u95dc\u4fc2\uff1a ( ) ( ) H L H z H z = \u2212 \u5f0f(3-2) \u610f\u5373\u5176\u983b\uf961\u7279\u6027\u70ba ( ) ( ) ( ) j j H L H e H e \u03c9 \u03c0 \u03c9 \u2212 = \uff0c\u5176\u610f\u7fa9\u5728\u65bc\u9ad8\u901a\u8207\u4f4e\u901a\uf984\u6ce2\u5668\u4e4b\u983b\uf961\u97ff\u61c9\u6703 2 2 2 2 ( ) L H z ( ) H H z ( ) H G z ( ) L G z x[n] y[n] \u4ee5 2 \u03c0 \u03c9 = \u70ba\u4e2d\u5fc3\u5f62\u6210\u5de6\u53f3\u5c0d\u7a31\u7684\u5716\u5f62\uff0c\u5728\u5c0f\u6ce2\u8f49\u63db\u4e2d\uff0c\u5373\uf9dd\u7528\u6b64\u5f62\u5f0f\u7684\uf984\u6ce2\u5668\uf92d\u5c0d\u8a0a\u865f \u4f5c\u5206\u983b\u8655\uf9e4\u3002 \u5716\u56db\u6240\u8868\u793a\uf9ba\u4e00\u8a0a\u865f\u85c9\u7531\u4e0a\u8ff0\u4e4b\uf984\u6ce2\u5668\u8655\uf9e4\u7684\u5206\u89e3\u7a0b\u5e8f(decomposition process) \uff0c\u5373 \uf9ea\u6563\u5c0f\u6ce2\u8f49\u63db\u7684\u5206\u983b\u8655\uf9e4\u3002\u5176\u4e2d\uff0c\u4e00\uf99a\uf905\u7684\uf978\u500d\u983b(octave-band)\u5206\u6790\uf984\u6ce2\u7d44\u8207\u4e4b\u5f8c\u7684\ufa09 \u4f4e\u53d6\u6a23(down-sampling)\u7684\u7d44\u5408\u901a\u5e38\u88ab\u7a31\u4f5c\u4e8c\u5143\u6a39(binary tree)\u7d50\u69cb\uff0c\u55ae\u4e00\u8f38\u5165\u5e8f\uf99c\u7d93\u7531\u5206 \u983b\u8655\uf9e4\u8207\ufa09\u4f4e\u53d6\u6a23\u5668(down-sampler)\u7684\u8f49\u63db\uff0c\u8f38\u51fa\u8b8a\u70ba\u5404\u5b50\u983b\u5e36\u5e8f\uf99c(sub-sequences)\u7684\u96c6 \u5408\u3002\u5728\u5716\u56db\u4e2d\uff0c\u6211\u5011\u770b\u5230\uf9ba\u4e00\u500b\u4e09\u968e(three-level)\u7684\u4e8c\u5143\u6a39\u5206\u6790\uf984\u6ce2\u5668\u7d44\u7d50\u69cb\uff0c\u5176\u4e2d\u9ad8\u901a ( ( ) H H z ) \u8207 \u4f4e \u901a ( ( ) L H z ) \uf984 \u6ce2 \u5668 \u90fd \u5177 \u6709 \u5b8c \u5168 \u91cd \u69cb (perfect reconstru ction) \u7684 \u96d9 \u901a \u9053 (two-channel)\u7279\u6027\uff0c\u5373\u8a0a\u865f\u901a\u904e\u6b64\uf978\uf984\u6ce2\u5668\u4e4b\u5f8c\uff0c\u4e26\u672a\u55aa\u5931\u4efb\u4f55\u8cc7\u8a0a\u6216\u5f15\u9032\u672a\u77e5\u7684\u5e72\u64fe \u8a0a\u865f\uff0c\u800c\u5f97\u4ee5\u5c07\u5206\u983b\u5f8c\u7684\u8a0a\u865f\u5b8c\u7f8e\u91cd\u5efa\u56de\u539f\u59cb\u8a0a\u865f\u3002\u53e6\u5916\uff0c\u5982\u679c\u8f38\u5165\u6b64\uf984\u6ce2\u5668\u7d44\u7684\u8a0a\u865f [ ] x n \u9577\ufa01\u70ba N \uff0c\u5728\u7b2c\u4e00\u968e\u9ad8\u901a\u5206\u6790\uf984\u6ce2\u5668\u4e4b\u8f38\u51fa [ ] 1 x n \u5373\u7d04\u70ba /2 N \uff0c\u800c\u518d\u4e0b\u4e00\u968e\u9ad8\u901a\u5206 \u6790\uf984\u6ce2\u5668\u8f38\u51fa [ ] 2 x n \u7d04\u70ba / 4 N \uff0c\u5982\u6b64\u91cd\u8907\u9019\u7a0b\u5e8f\uff0c\u5c31\u53ef\u4ee5\u5f97\u5230\u6240\u6709\u968e\u5c64\u4e4b\uf984\u6ce2\u5668\u7684\u8f38 \u51fa\u3002\u8868\u4e00\uf99c\u51fa\uf9ba\u5404\u5c64\uf984\u6ce2\u5668\u7684\u5176\u983b\u5e36\u7bc4\u570d\u53ca\u8f38\u51fa\u8a0a\u865f\u7684\u9577\ufa01\u3002\u4ee5\u4e0a\u6240\u8ff0\u4e4b\uf978\u500d\u983b(octave) \u5b8c\u5168\u91cd\u69cb QMF \uf984\u6ce2\u5668\u7d44\u5c0d\u8f38\u5165\u8a0a\u865f\u7684\u8655\uf9e4\u7a0b\u5e8f\uff0c\u5373\u70ba\uf9ea\u6563\u5c0f\u6ce2\u8f49\u63db(discrete wavelet transform, DWT)\uff0c\u7531\u4e0a\u8ff0\u53ef\u77e5\uff0c\u5982\u679c\u6240\u7528\u4e4b\uf984\u6ce2\u5668\u7d44\u7684\u968e\u5c64\uf969\u70ba L (\u76f8\u7576\u65bc L \u5c64\u7684\uf9ea\u6563 \u5c0f\u6ce2\u8f49\u63db) \uff0c\u5247\u8f38\u51fa\u7684\u7e3d\u5206\u983b\u8a0a\u865f\u7684\uf969\u76ee\u5247\u70ba 1 L + \u500b\u3002 \u5716\u56db \uf9ea\u6563\u5c0f\u6ce2\u8f49\u63db\u7684\u5206\u89e3\u7a0b\u5e8f\u5716(\u968e\u5c64\uf969\u70ba 3) \u8868\u4e00\u3001\u4e09\u5c64\uf9ea\u6563\u5c0f\u6ce2\u8f49\u63db(DWT)\u6bcf\u4e00\u968e\u5c64\u7684\u8f38\u51fa\u8a0a\u865f\u9ede\uf969\u53ca\u76f8\u5c0d\u61c9\u7684\u983b\uf961\u7bc4\u570d ( [ ] x n \u53d6\u6a23\u983b\uf961\u70ba s F Hz ) \u8a0a\u865f \u7e3d\u9ede\uf969 \u983b\uf961\u7bc4\u570d [ ] x n N 0, 2 Hz s F \u23a1 \u23a4 \u23a2 \u23a5 \u23a3 \u23a6 [ ] 1 x n 2 N 4 Hz, 2 Hz s s F F \u23a1 \u23a4 \u23a2 \u23a5 \u23a3 \u23a6 [ ] 2 x n 4 N 8 Hz, 4 Hz s s F F \u23a1 \u23a4 \u23a2 \u23a5 \u23a3 \u23a6 [ ] 3 x n 8 N 16 Hz, 8 Hz s s F F \u23a1 \u23a4 \u23a2 \u23a5 \u23a3 \u23a6 [ ] 4 x n 8 N 0, 16 Hz s F \u23a1 \u23a4 \u23a2 \u23a5 \u23a3 \u23a6 \u5f9e\u4e0a\u8868\u4e00\u53ef\u77e5\uff0c\u5982\u679c\u5e8f\uf99c [ ] x n \u6db5\u84cb\u7684\u983b\uf961\u7bc4\u570d\u70ba[0, 2 s F Hz] \uff0c\u5176\u4e2d s F \u70ba [ ] x n \u7684\u53d6 \u6a23 \u983b \uf961 \uff0c \u5247 \u7d93 \u7531 \u7b2c \u4e00 \u968e \u6b63 \u4ea4 \u93e1 \u50cf \uf984 \u6ce2 \u5668 \u7d44 \u4e4b \u9ad8 \u983b \u8f38 \u51fa [ ] 1 x n \uff0c \u983b \uf961 \u7bc4 \u570d \u70ba 4 Hz, 2 Hz s s F F \u23a1 \u23a4 \u23a2 \u23a5 \u23a3 \u23a6 \uff0c\u4f9d\u6b64\uf9d0\u63a8\uff0c\u9010\u6b65\u5f80\u4f4e\u983b\uf961\u90e8\u4efd\u505a\uf967\u7b49\u5206\ufa00\u5272\uff0c\u96a8\u8457\u983b\uf961\u8d8a\u9ad8\uff0c\u5176 L H 2 \u2193 2 \u2193 H H L H 2 \u2193 2 \u2193 H H L H 2 \u2193 2 \u2193 [ ] x n [ ] 1 x n [ ] 2 x n [ ] 3 x n [ ] 4 x n H H Level 1 Level 3 Level 2 \u983b\u5bec\u5247\u8d8a\u5927\u3002\u7531\u4e0a\u6240\u8ff0\uff0c\uf9ea\u6563\u5c0f\u6ce2\u8f49\u63db\u7684\u7b2ck \u500b\u8f38\u51fa [ ] k x n \uff0c\u76f8\u7576\u65bc\u662f\u539f\u59cb\u5e8f\uf99c [ ] x n \u8207 \u7b2ck \u500b\u5e36\u901a\uf984\u6ce2\u5668\u4e4b\u8108\u885d\u97ff\u61c9(impulse response)\u76f8\u4e92\u647a\u7a4d(convolution)\u7684\u7d50\u679c\uff0c\u5982\u5f0f(3-3) \u6240\u793a\uff1a [ ] [ ] [ ] 1 ,1 2 , 0 -1 , 2 , . k k m k k k m h n m x m k L x n h n m x m k L \u221e + =\u2212\u221e \u221e =\u2212\u221e \u23a7 \u23aa \u23a1 \u23a4 \u23aa \u2212 \u2264 \u2264 \u23aa \u23a3 \u23a6 \u23aa \u23aa = \u23a8 \u23aa \u23aa \u23a1 \u23a4 \u2212 = \u23aa \u23a3 \u23a6 \u23aa \u23aa \u23a9 \u2211 \u2211 \u5f0f(3-3) \u5176\u4e2d 1 ,1 2 k k h n + \u23a1 \u23a4 \u23a3 \u23a6 \u8207 2 k k h n \u23a1 \u23a4 \u23a3 \u23a6 \u70ba\u539f\u59cb\u8108\u885d\u97ff\u61c9 [ ] ,1 k h n \u8207 [ ] k h n \ufa09\u4f4e\u53d6\u6a23\u800c\u5f97\uff0c\u800c\u9ad8\u901a\uf984\u6ce2\u5668\u4e4b \u8f38\u51fa\uff0c\u7a31\u70ba\u7d30\u7bc0(detail)\u4fc2\uf969\uff1b\u4f4e\u901a\uf984\u6ce2\u5668\u4e4b\u8f38\u51fa\u5247\u7a31\u70ba\u8fd1\u4f3c(approximation)\u4fc2\uf969\u3002 \uf974 \u8981 \u85c9 \u7531 \u6240 \u6709 \u5b50 \u983b \u5e36 \u8a0a \u865f \u7684 \u96c6 \u5408 \u5f97 \u5230 \u539f \u59cb \u5e8f \uf99c [ ] x n \uff0c \u5176 \u904e \u7a0b \u7a31 \u70ba \u91cd \u5efa \u7a0b \u5e8f (reconstruction process)\uff0c\u6b64\u6070\u70ba\u524d\u8ff0\u4e4b\u5206\u89e3\u7a0b\u5e8f\u7684\u53cd\u7a0b\u5e8f(inverse process)\uff0c\u5373\u4f7f\u7528\u6240\u5f97 \u4e4b [ ] { } k x n \u7d93 L \u968e\uf978\u500d\u983b\u5b8c\u5168\u91cd\u69cb QMF \u5408\u6210\uf984\u6ce2\u5668\u7d44\u9010\u5c64\u8655\uf9e4\uff0c\u6b64\u904e\u7a0b\u5373\u70ba\u53cd\uf9ea\u6563\u5c0f\u6ce2 \u8f49\u63db(inverse discrete wavelet transform, IDWT)\uff0c\u5982\u4e0b\u5716\u4e94\u6240\u793a\uff1a \u5716\u4e94 \u53cd\uf9ea\u6563\u5c0f\u6ce2\u8f49\u63db\u7684\u91cd\u5efa\u7a0b\u5e8f\u5716(\u968e\u5c64\uf969\u70ba 3) \u9084\u539f\u7a0b\u5e8f\u5176\uf969\u5b78\u5f0f\u5982\u5f0f(3-4)\uff1a [ ] [ ] [ ] 1 1 ,1 0 2 2 L k L k k k L k m m x n g n m x m g n m x m \u2212 \u221e \u221e + = =\u2212\u221e =\u2212\u221e \u23a1 \u23a4 \u23a1 \u23a4 = \u2212 + \u2212 \u23a3 \u23a6 \u23a3 \u23a6 \u2211 \u2211 \u2211 , (3-4) \u5176\u4e2d 1 ,1 2 k k g n + \u23a1 \u23a4 \u23a3 \u23a6 \u8207 2 k k g n \u23a1 \u23a4 \u23a3 \u23a6 \u5206\u5225\u70ba\u539f\u59cb\u8108\u885d\u97ff\u61c9 [ ] ,1 k g n \u8207 [ ] k g n \u63d0\u5347\u53d6\u6a23\u800c\u5f97\u3002\u5716\u4e94\u4e4b\u9084\u539f \u7a0b\u5e8f\uff0c\u5373\u662f\u5c07\u5404\u5b50\u983b\u5e36\u7684\u8a0a\u865f\u4ee5\u63d0\u5347\u53d6\u6a23(up-sampling)\u7684\u65b9\u5f0f\u589e\u52a0\u5e8f\uf99c\u9ede\uf969\uff0c\u518d\u7d93\u904e\u9ad8 \u901a( ( ) ( ) H H G z H z = )\u8207\u4f4e\u901a( ( ) ( ) L L G z H z = )\u4e4b\u5408\u6210\uf984\u6ce2\u5668\u8655\uf9e4\uff0c\u5982\u679c\u7b2c\u4e09\u968e\u8f38\u5165\u8a0a\u865f\u9ede \uf969\u70ba 8 N \uff0c\u5247\u5728\u7b2c\u4e09\u968e\u8f38\u51fa\u8a0a\u865f\u9ede\uf969\u7d04\u70ba 4 N \uff0c\u800c\u7b2c\u4e8c\u968e\u8f38\u51fa\u8a0a\u865f\u9ede\uf969\u7d04\u70ba 2 N \uff0c\u5982\u6b64 \u91cd\u8986\u6b64\u7a0b\u5e8f\uff0c\u5247\u6700\u5f8c\u6240\u5f97\u4e4b\u8a0a\u865f\u70ba\u539f\u59cb N \u9ede\u4e4b\u8a0a\u865f [ ] x n : \u4ee5\u4e0a\u6240\u8ff0\u70ba\u5c0f\u6ce2\u8f49\u63db\u4e4b\u5206\u6790(analysis)\u8207\u5408\u6210(synthesis)\u7a0b\u5e8f\uff0c\u7d93\u7531\u6b64\u8f49\u63db\u5f8c\uff0c\u8a0a\u865f \u88ab\u5206\u89e3\u6210\u5404\u500b\u5b50\u983b\u5e36\u4e4b\u8a0a\u865f\uff0c\u5982\u8868\u4e00\u6240\u793a\uff0c\u4f4e\u983b\u90e8\u5206\u7684\u5b50\u983b\u5e36\u983b\u5bec\u8f03\u5c0f\uff0c\u800c\u9ad8\u983b\u90e8\u5206\u7684 \u5b50\u983b\u5e36\u983b\u5bec\u8f03\u5927\u3002\u85c9\u7531\u4ee5\u4e0a\u6240\u8ff0\u7684\uf9ea\u6563\u5c0f\u6ce2\u8f49\u63db\u7a0b\u5e8f\uff0c\u6211\u5011\u53ef\u4ee5\u5c07\u8a9e\u97f3\u7279\u5fb5\u6642\u9593\u5e8f\uf99c\u4f5c \u5206\u983b\u7684\u8655\uf9e4\uff0c\u9032\u800c\u91dd\u5c0d\uf967\u540c\u8abf\u8b8a\u983b\u5e36\u6210\u5206\u7684\u8a9e\u97f3\u7279\u5fb5\u5e8f\uf99c\u5206\u5225\u4f5c\u8655\uf9e4\uff0c\u5728\u4e0b\u4e00\u7ae0\uf9e8\uff0c\u6211 \u5011\u5c07\u4ecb\u7d39\u5176\u5c0d\u61c9\u7684\u7684\u5206\u983b\u5f0f\u7279\u5fb5\u7d71\u8a08\u88dc\u511f\u6cd5\u3002 \u56db\u3001\u5206\u983b\u5e36\u7279\u5fb5\u7d71\u8a08\u6b63\u898f\u5316\u6cd5 \u5728\u9019\u4e00\u7ae0\u4e2d\uff0c\u6211\u5011\u9996\u5148\u5728\u7b2c\u4e00\u7bc0\u4ecb\u7d39\u6240\u65b0\u63d0\u51fa\u4e4b\u5206\u983b\u5e36\u7279\u5fb5\u7d71\u8a08\u88dc\u511f\u6cd5\u7684\u6b65\u9a5f\u53ca\u7279 \u6027\uff0c\u63a5\u8457\u5728\u7b2c\u4e8c\u7bc0\u4e2d\uff0c\u6211\u5011\u5c07\u4ee5\u4e00\u8a9e\uf906\u70ba\uf9b5\uff0c\u9a57\u8b49\u6240\u63d0\u4e4b\u65b0\u65b9\u6cd5\u8db3\u4ee5\u6709\u6548\ufa09\u4f4e\u96dc\u8a0a\u5c0d\u8a9e \u97f3\u8abf\u8b8a\u983b\u8b5c\u4e4b\u5e72\u64fe\u3002 Level 3 [ ] 4 x n [ ] 3 x n 2 \u2191 2 \u2191 L G H G \u2295 2 \u2191 L G 2 \u2191 H G [ ] 2 x n \u2295 2 \u2191 L G 2 \u2191 H G [ ] 1 x n \u2295 [ ] x n Level 2 Level 1 (\u4e00) \u5206\u983b\u5e36\u7279\u5fb5\u7d71\u8a08\u6b63\u898f\u5316\u6cd5\u7684\u6b65\u9a5f\uf96f\u660e \u5047\u8a2d\u4e00\u6bb5\u8a9e\uf906(utterance)\u7684\u67d0\u4e00\u7dad\u6885\u723e\u5012\u983b\u8b5c\u8a9e\u97f3\u7279\u5fb5\u4ee5\u4e0b\u5f0f(4-1)\u8868\u793a: ( ) [ ] { } ;1 , 0 1 , m x n n N m M < \u2264 \u2264 \u2264 \u2212 (4-1) \u5176\u4e2d N \u70ba\u6b64\u7279\u5fb5\u5e8f\uf99c\u7684\u7e3d\u97f3\u6846\uf969\uff0c M \u8868\u793a\u6bcf\u4e00\u97f3\u6846\u4e2d\u7684\u7279\u5fb5\u7e3d\uf969\u3002\u6b64\u7279\u5fb5\u5e8f\uf99c\u76f8\u7576\u65bc \u6db5\u84cb\uf9ba\u5168\u8abf\u8b8a\u983b\u5e36(full-band)\u7684\u8a9e\u97f3\u8cc7\u8a0a\uff0c\u7136\u800c\uff0c\u7531\u524d\u9762\u7ae0\u7bc0\u6240\u8ff0\uff0c\uf967\u540c\u7684\u983b\u5e36\u6210\u4efd\uff0c \u5c0d\u65bc\u8a9e\u97f3\u8fa8\u8a8d\u7684\u91cd\u8981\u6027\u6709\u6240\uf967\u540c\uff0c\u57fa\u65bc\u6b64\u9805\uf9e4\u7531\uff0c\u9019\uf9e8\u6211\u5011\u4f7f\u7528\u5206\u983b\u7684\u6280\u8853\uff0c\u5c07\u6b64\u7279\u5fb5 \u5e8f\uf99c\u5206\u89e3\u6210\u5404\uf967\u540c\u983b\uf961\u7684\u6210\u5206\uff0c\u5982\u4ee5\u4e0b\u6b65\u9a5f (\u70ba\uf9ba\u7c21\uf9e0\uf96f\u660e\u8d77\ufa0a\uff0c\u6211\u5011\u5728\u4e4b\u5f8c\u7684\u8a0e\uf941\u4e2d\uff0c \u5c07\uf96d\uf976\u5f0f(4-1)\u4e2d\u4ee3\u8868\uf967\u540c\u7dad\u7279\u5fb5\u7684\u4e0a\u6a19\" ( ) m \"\uff0c\u56e0\u70ba\u6211\u5011\u662f\u5c0d\u6bcf\u4e00\u500b\uf967\u540c\u7dad\u7684\u7279\u5fb5\u5e8f\uf99c \u7686\u4f5c\u540c\u6a23\u8655\uf9e4)\uff1a \u9996\u5148\uff0c\u6211\u5011\u5c07\u539f\u59cb\u7279\u5fb5\u5e8f\uf99c [ ] { } x n \ufa00\u5272\u6210 L \u500b\u5206\u983b\u5e36\u4e14\u5047\u8a2d\u6bcf\u4e00\u5206\u983b\u5e36\u90fd\u70ba\u5404\u81ea\u7368 \uf9f7\uff0c\u800c\u6bcf\u4e00\u983b\u5e36\u4e2d\u7684\u5e8f\uf99c\u8868\u793a\u70ba [ ] { },1 x n L \u2264 \u2264 \uff0c\u6b64\ufa00\u5272\u983b\u5e36\u7684\u65b9\u6cd5\u662f\u5c07\u539f\u59cb\u7279\u5fb5 \u901a\u904e\u4e00\u500d\u983b(octave-band)\u5e36\u901a\uf984\u6ce2\u5668\u7d44\uff0c\u6bcf\u4e00\u5b50\u983b\u5e36\u8a0a\u865f\u518d\u4f5c\ufa09\u4f4e\u53d6\u6a23(down-sampling) \u8655\uf9e4\uff0c\u6b64\u6b65\u9a5f\u7b49\u6548\u65bc\u57f7\ufa08( ) 1 L \u2212 \u968e\u7684\uf9ea\u6563\u5c0f\u6ce2\u8f49\u63db(discrete wavelet transform, DWT) \u65bc \u7279\u5fb5\u5e8f\uf99c [ ] x n \u4e0a\u3002\u53e6\u5916\uff0c\u5047\u8a2d\u7279\u5fb5\u5e8f\uf99c [ ] { } x n \u97f3\u6846\u53d6\u6a23\uf961\u70ba s F (Hz) \uff0c\u5247\u5176\u8abf\u8b8a\u983b\u8b5c\u983b \uf961\u7bc4\u570d\u70ba[ ] 0, / 2 s F \uff0c\u56e0\u6b64\uff0c\u7b2c \u500b\u5206\u983b\u5e36\u5e8f\uf99c\u7684\u983b\uf961\u7bc4\u570d\uff0c\u53ef\u88ab\u8fd1\u4f3c\u8868\u793a\u6210\u5f0f(4-2)\uff1a ( ) ( ) ( ) 1 2 1 1 1 1 0, / 2 if =1 2 2 2 / 2 , / 2 if 2, 3, , 2 2 s L s s L L F F F L \u2212 \u2212 \u2212 \u2212 \u2212 \u23a7\u23a1 \u23a4 \u23aa \u23aa \u23a2 \u23a5 \u23aa \u23a2 \u23a5 \u23aa\u23a3 \u23a6 \u23aa \u23a8 \u23a1 \u23a4 \u23aa \u23aa \u23a2 \u23a5 = \u23aa \u23aa\u23a2 \u23a5 \u23aa\u23a3 \u23a6 \u23a9 (4-2) \u5728 DWT \u7a0b\u5e8f\u4e2d\uff0c\u5176\u65b9\u5f0f\u662f\u5c07\u4e00\u4e3b\u983b\u5e36\u4f9d\u983b\u5bec\u5148\u7b49\ufa00\u70ba\uf978\u500b\u526f\u983b\u5e36\uff0c\u7136\u5f8c\u4fdd\u6301\u9ad8\u983b \u5e36\uf967\u52d5\uff0c\u5c07\u4f4e\u983b\u5e36\u518d\u7b49\ufa00\u6210\uf978\u500b\u526f\u983b\u5e36\uff0c\u5982\u6b64\u53cd\u8986\u9032\ufa08\uff0c\u56e0\u6b64\u76f8\u7576\u65bc\u4f4e\u983b\u90e8\u4efd\u6703\u4f7f\u7528\u8f03 \u591a\u500b\u983b\u5bec\u8f03\u5c0f\u7684\uf984\u6ce2\u5668\uff0c\u800c\u9ad8\u983b\u90e8\u4efd\u5247\u7528\u8f03\u5c11\u500b\u983b\u5bec\u8f03\u5927\u7684\uf984\u6ce2\u5668\uff0c\u800c\u56e0\u70ba DWT \u7a0b\u5e8f \u4e2d\u7684\ufa09\u4f4e\u53d6\u6a23(down-sampling)\u7684\u904b\u7b97\uff0c\u6240\u4ee5\u6bcf\u4e00\u5206\u983b\u5e36\u7684\u5e8f\uf99c [ ] { } x n \u9577\ufa01\u7d04\u6b63\u6bd4\u65bc\u983b \u5bec\u7684\u5927\u5c0f\u3002 \u63a5\u8457\uff0c\u5c07\u4e0a\u6b65\u9a5f\u6240\u5f97\u7684\u5206\u983b\u5e36\u5e8f\uf99c [ ] { } x n \u505a\u7279\u5fb5\u7d71\u8a08\u6b63\u898f\u5316\uff0c\u5f97\u5230\u65b0\u7684\u5206\u983b\u5e36\u5e8f \uf99c\uff0c\u8868\u793a\u70ba [ ] { } x n \uff0c\u5176\u7279\u5fb5\u7d71\u8a08\u6b63\u898f\u5316\u7684\u65b9\u5f0f\u662f\u5c07\u6bcf\u4e00\u8a9e\uf906\u4e4b\u5b50\u983b\u5e36\u7279\u5fb5 [ ] { } x n \u7684\u7d71 \u8a08\uf97e\uff0c\u8b6c\u5982\u5e73\u5747\u503c(mean)\u3001\u8b8a\uf962\uf969(variance)\u6216\u662f\uf901\u9ad8\u968e\u7684\u52d5\u5dee(moments)\u4f5c\u8655\uf9e4\uff0c\u4f7f\u65b0 \u7684\u7279\u5fb5\uf96b\uf969 [ ] { } x n \u7684\u7d71\u8a08\uf97e\u7b49\u540c\u6216\u903c\u8fd1\u4e00\u76ee\u6a19(target)\u7d71\u8a08\uf97e\uff0c\u800c\u6b64\u76ee\u6a19\u7d71\u8a08\uf97e\u662f\u7531\u4e7e \u6de8\u8a13\uf996\u8a9e\uf9be\u5eab\u4e2d\uff0c\u6240\u6709\u8a9e\uf906\u4e4b\u5b50\u983b\u5e36\u7279\u5fb5 [ ] { } x n \u4f30\u6e2c\u8a08\u7b97\u800c\u5f97\u3002\u5728\u9019\uf9e8\u6211\u5011\u4f7f\u7528\u7684\u7279 \u5fb5\u7d71\u8a08\u6b63\u898f\u5316\u6cd5\u6709\uf978\u7a2e\uff0c\u5206\u5225\u70ba\u5012\u983b\u8b5c\u5e73\u5747\u503c\u8207\u8b8a\uf962\uf969\u6b63\u898f\u5316\u6cd5(MVN)\u8207\u7d71\u8a08\u5716\u7b49\u5316\u6cd5 (HEQ)\uff0c\u4ee5 MVN \u6cd5\u800c\u8a00\uff0c\u6240\u5f97\u65b0\u7684\u5206\u983b\u5e36\u5e8f\uf99c [ ] { } c n \u53ef\u8868\u793a\u70ba\u4e0b\u5f0f(4-3)\uff1a [ ] [ ] , , , , s t t s x n x n \u03bc \u03c3 \u03bc \u03c3 \u239b \u239e \u2212 \u239f \u239c \u239f \u239c = \u00d7 + \u239f \u239c \u239f \u239f \u239c \u239d \u23a0 (4-3) \u5176\u4e2d ,s \u03bc \u8207 2 ,s \u03c3 \u5206\u5225\u70ba\u76ee\u524d\u8655\uf9e4\u7684\u55ae\u4e00(single)\u5206\u983b\u5e36\u5e8f\uf99c [ ] { } x n \u7684\u5e73\u5747\u503c\u8207\u8b8a\uf962\uf969\uff0c\u800c ,t \u03bc \u8207 2 ,t \u03c3 \u70ba\u76ee\u6a19(target)\u5e73\u5747\u503c\u8207\u8b8a\uf962\uf969\uff0c\u6b64\u76ee\u6a19\u5e73\u5747\u503c\u8207\u8b8a\uf962\uf969\u662f\u7531\u539f\u59cb\u4e7e\u6de8\u8a13\uf996\u8a9e\uf9be \u5eab\u4e2d\u6240\u6709\u5206\u983b\u5e36\u7279\u5fb5\u5e8f\uf99c [ ] { } x n \u4f30\u6e2c\u800c\u5f97\u3002\u540c\u6a23\u5730\uff0c\u5982\u4ee5 HEQ \u4f5c\u70ba\u7d71\u8a08\u88dc\u511f\u6cd5\uff0c\u5247 [ ] { } x n \u8207 [ ] { } x n \u5f7c\u6b64\u95dc\u4fc2\u70ba\u4e0b\u5f0f(4-4)\uff1a [ ] [ ] ( ) ( ) 1 , , X t X s x n F F x n \u2212 = (4-4) \u5176\u4e2d ( ) , . X s F \u70ba\u76ee\u524d\u8655\uf9e4\u7684\u55ae\u4e00\u5206\u983b\u5e36\u5e8f\uf99c [ ] { } x n \u6240\u4f30\u6e2c\u7684\u6a5f\uf961\u5206\u4f48\u51fd\uf969(probability distribution function)\uff0c\u800c ( ) , . X t F \u662f\u7531\u539f\u59cb\u4e7e\u6de8\u8a13\uf996\u8a9e\u5eab\u4e2d\u6240\u6709\u5206\u983b\u5e36\u7279\u5fb5\u5e8f\uf99c [ ] { } x n \u6240 \u4f30\u6e2c\u800c\u5f97\u7684\u6a5f\uf961\u5206\u4f48\u51fd\uf969\u3002 \u6700\u5f8c\uff0c\u5c07\u6240\u6709\u7684\u5206\u983b\u5e36\u5e8f\uf99c [ ] { } x n (\u5305\u542b\uf9ba\uf901\u65b0\u904e\u5f8c\u8207\u672a\uf901\u65b0\u7684\u5206\u983b\u5e36\u5e8f\uf99c)\u900f\u904e ( ) 1 L \u2212 \u968e\u53cd\uf9ea\u6563\u5c0f\u6ce2\u8f49\u63db(inverse discrete wavelet transform, IDWT)\uff0c\u91cd\u5efa\u70ba\u65b0\u7684\u7279\u5fb5\u6642 \u9593\u5e8f\uf99c\uff0c\u6b64\u5373\u70ba\u6211\u5011\u6700\u5f8c\u4f7f\u7528\u4e4b\u8a9e\u97f3\u7279\u5fb5\u5e8f\uf99c [ ] { } x n \u3002 \u4e0a\u8ff0\u5206\u983b\u5e36\u7d71\u8a08\u6b63\u898f\u5316\u6cd5\u7684\uf9ca\u7a0b\u5716\u7e6a\u65bc\u4e0b\u5716\uf9d1\uff1a x[n] H L (z) H H (z) \u21932 \u21932 H H (z) \u21932 \u21932 H L (z) H H (z) \u21932 S S S S \u21932 \u21912 \u21912 H L (z) G L (z) G H (z) \u21912 \u21912 G L (z) G H (z) \u21912 \u21912 G L (z) G H (z) [ ] x n [ ] 1 x n [ ] 2 x n [ ] 3 x n [ ] 4 x n [ ] 1 x n [ ] 2 x n [ ] 3 x n [ ]",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "\u5716\uf9d1 \u5206\u983b\u5e36\u7279\u5fb5\u7d71\u8a08\u6b63\u898f\u5316\u6cd5\u7684\u904b\u4f5c\u7a0b\u5e8f\u5716 \u70ba\uf9ba\u5728\u4e4b\u5f8c\u7684\u8a0e\uf941\u4e2d\uff0c\u6709\u6548\u5340\u9694\u50b3\u7d71\u65b9\u6cd5\u8207\u6240\u63d0\u51fa\u7684\u65b0\u65b9\u6cd5\uff0c\u5c0d\u50b3\u7d71\u5168\u983b\u5e36 (full-band)\u7684\u7279\u5fb5\u7d71\u8a08\u6b63\u898f\u5316\u6cd5 MVN \u8207 HEQ\uff0c\u6211\u5011\u5206\u5225\u7a31\u4e4b\u70ba FB-MVN \u8207 FB-HEQ\uff0c \u800c \u5982 \u5f0f (4-3) \u8207 (4-4) \u4e2d \u5206 \u983b (sub-band) \u8655 \uf9e4 \u7684 \u7279 \u5fb5 \u7d71 \u8a08 \u6b63 \u898f \u5316 \u6cd5 \uff0c \u6211 \u5011 \u5247 \u5206 \u5225 \u7a31 \u70ba SB-MVN \u548c SB-",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "De-noising by soft-thresholding",
"authors": [
{
"first": "D",
"middle": [
"L"
],
"last": "Donoho",
"suffix": ""
}
],
"year": 1995,
"venue": "IEEE Trans. on Information Theory",
"volume": "41",
"issue": "3",
"pages": "613--627",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "D.L. Donoho, \"De-noising by soft-thresholding\", IEEE Trans. on Information Theory, vol. 41, no. 3, pp. 613-627, May 1995.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Suppression of aco ustic noise in speech u sing spectral subtraction",
"authors": [
{
"first": "S",
"middle": [
"F"
],
"last": "Boll",
"suffix": ""
}
],
"year": 1979,
"venue": "IEEE Trans. on Acoustics, Speech and Signal Processing",
"volume": "27",
"issue": "2",
"pages": "113--120",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "S. F . Boll, \"Suppression of aco ustic noise in speech u sing spectral subtraction \", IEEE Trans. on Acoustics, Speech and Signal Processing, vol. 27, no. 2, pp. 113-120, 1979",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Noise spectrum estim ation with entropy-based V AD i n non-stationary environm ents",
"authors": [
{
"first": "K.-C",
"middle": [],
"last": "B.-F . W U",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Ang",
"suffix": ""
}
],
"year": 2006,
"venue": "IEICE Trans. on Fundamentals of Electronics",
"volume": "",
"issue": "2",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "B.-F . W u, K.-C. W ang, \"Noise spectrum estim ation with entropy-based V AD i n non-stationary environm ents\", IEICE Trans. on Fundamentals of Electronics, Communications and Computer Sciences, vol. E89-A, no. 2, Feb 2006",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Wavelet speech e nhancement based on voiced /unvoiced decision",
"authors": [
{
"first": "J",
"middle": [
"K"
],
"last": "Lee",
"suffix": ""
},
{
"first": "C",
"middle": [
"D"
],
"last": "Yoo",
"suffix": ""
}
],
"year": 2003,
"venue": "",
"volume": "",
"issue": "",
"pages": "4149--4156",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "J.K. Lee , C.D. Yoo, \"Wavelet speech e nhancement based on voiced /unvoiced decision\", 32 nd Inter-Noise, pp. 4149-4156, Aug 2003",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "A speech endpoint d etection method based on wavelet coefficient varian ce an d sub-band am plitude varian ce",
"authors": [
{
"first": "X",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "Z",
"middle": [],
"last": "Zhao",
"suffix": ""
},
{
"first": "G",
"middle": [],
"last": "Zhao",
"suffix": ""
}
],
"year": 2006,
"venue": "International Conference on Innovative Computing",
"volume": "3",
"issue": "",
"pages": "83--86",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "X. Zhang, Z. Zhao and G. Zhao, \"A speech endpoint d etection method based on wavelet coefficient varian ce an d sub-band am plitude varian ce\", International Conference on Innovative Computing, Information and Control, vol. 3, pp. 83-86, 2006",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Mel-scale d discre te wavelet co efficients for speech recognition",
"authors": [
{
"first": "J",
"middle": [
"N"
],
"last": "Gowdy",
"suffix": ""
},
{
"first": "Z",
"middle": [],
"last": "Tufekci",
"suffix": ""
}
],
"year": 2000,
"venue": "Proc. IEEE International Conference on Acoustics, Speech and Signal Processing",
"volume": "3",
"issue": "",
"pages": "1351--1354",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "J.N. Gowdy and Z. Tufekci, \"Mel-scale d discre te wavelet co efficients for speech recognition\", in Proc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), vol. 3, pp. 1351-1354, 2000",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Objective wavelet packets features for speaker verification",
"authors": [
{
"first": "M",
"middle": [],
"last": "Siafarikas",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Anchev",
"suffix": ""
},
{
"first": "N",
"middle": [],
"last": "Fakotak",
"suffix": ""
}
],
"year": 2002,
"venue": "Proc. International Conference on Spoken Language Processing (ICSLP)",
"volume": "",
"issue": "",
"pages": "2365--2368",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "M. Siafarikas, T. G anchev and N. Fakotak is, \"Objective wavelet packets features for speaker verification\" , in Proc. International Conference on Spoken Language Processing (ICSLP), pp. 2365-2368, 2002",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Cepstral analysis technique for automatic speaker verification",
"authors": [
{
"first": "S",
"middle": [],
"last": "Furui",
"suffix": ""
}
],
"year": 1981,
"venue": "IEEE Trans. on Acoustics, Speech and Signal Processing",
"volume": "29",
"issue": "2",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "S. Furui, \"Cepstral analysis technique for automatic speaker verification\", IEEE Trans. on Acoustics, Speech and Signal Processing, vol. 29, no. 2, Apr 1981",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Bilm es, '' Frontend post-processing and backend model enhancement on The Aurora 2.0/3.0 databases' '",
"authors": [
{
"first": "C.-P",
"middle": [],
"last": "Chen",
"suffix": ""
},
{
"first": "K",
"middle": [],
"last": "Filaliy",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "",
"suffix": ""
}
],
"year": 2002,
"venue": "Proc. International Conference on Spoken Language Processing (ICSLP)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "C.-P. Chen, K. Filaliy and J. A. Bilm es, '' Frontend post-processing and backend model enhancement on The Aurora 2.0/3.0 databases' ', in Proc. International Conference on Spoken Language Processing (ICSLP), 2002",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "M VA processing of speech featu res",
"authors": [
{
"first": "C.-P",
"middle": [],
"last": "Chen",
"suffix": ""
},
{
"first": "J. -A",
"middle": [],
"last": "Bilm",
"suffix": ""
}
],
"year": 2006,
"venue": "IEEE Trans. on Audio, Speech, and Language Processing",
"volume": "15",
"issue": "1",
"pages": "257--270",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "C.-P. Chen and J. -A. Bilm es, \"M VA processing of speech featu res\", IEEE Trans. on Audio, Speech, and Language Processing, vol.15, no. 1, pp.257-270, Jan 2006.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Quantile base d histogram equalization for noise robust speech recognition",
"authors": [
{
"first": "F",
"middle": [],
"last": "Hilger",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Ney",
"suffix": ""
}
],
"year": 2001,
"venue": "Speech Communication and Technology",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "F. Hilger and H. Ney , \"Quantile base d histogram equalization for noise robust speech recognition\", in European Conference on Speech Communication and Technology (Eurospeech), 2001",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "On the importance of various modulation frequencies for speech recognition",
"authors": [
{
"first": "N",
"middle": [],
"last": "Ka Nedera",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Arai",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Herm",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Pavel",
"suffix": ""
}
],
"year": 1997,
"venue": "Speech Communication and Technology",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "N. Ka nedera, T . Arai, H. Herm ansky, and M. Pavel, \"On the importance of various modulation frequencies for speech recognition \", in European Conference on Speech Communication and Technology (Eurospeech), 1997",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "RASTA Processing of Speech",
"authors": [
{
"first": "H",
"middle": [],
"last": "Hermansky",
"suffix": ""
},
{
"first": "N",
"middle": [],
"last": "Mor Gan",
"suffix": ""
}
],
"year": 1994,
"venue": "IEEE Trans. on Speech and Audio Processing",
"volume": "2",
"issue": "4",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "H. Hermansky and N. Mor gan, \"RASTA Processing of Speech\", IEEE Trans. on Speech and Audio Processing, vol.2, no. 4, Oct. 1994.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Optim ization of tem poral filters for constructing robust features in speech recognition",
"authors": [
{
"first": "J-W",
"middle": [],
"last": "Hung",
"suffix": ""
},
{
"first": "L-S",
"middle": [],
"last": "Lee",
"suffix": ""
}
],
"year": 2006,
"venue": "IEEE Trans. on Audio, Speech and Language Processing",
"volume": "4",
"issue": "3",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "J-W. Hung and L-S. Lee, \"Optim ization of tem poral filters for constructing robust features in speech recognition\", IEEE Trans. on Audio, Speech and Language Processing, vol.4, no. 3, May 2006",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Application of quadrature mirror filters to spilt-band voice coding schem es",
"authors": [
{
"first": "D",
"middle": [],
"last": "Esteban",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Galand",
"suffix": ""
}
],
"year": 1977,
"venue": "Proc. IEEE International Conference on Acoustics, Speech, and Signal Processing",
"volume": "2",
"issue": "",
"pages": "191--195",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "D. Esteban and C. Galand. \"Application of quadrature mirror filters to spilt-band voice coding schem es\", in Proc. IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), vol. 2, pp. 191-195, May 1977.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "The A URORA Experim ental Fram ework for the Performance Evaluation s of Speech Rec ognition System s under Noisy Conditions",
"authors": [
{
"first": "H",
"middle": [
"G"
],
"last": "Hirsch",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Pearce",
"suffix": ""
}
],
"year": 2000,
"venue": "Proc. of ISCA IIWR ASR2000",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "H. G . Hirsch and D. Pearce, \"The A URORA Experim ental Fram ework for the Performance Evaluation s of Speech Rec ognition System s under Noisy Conditions \", in Proc. of ISCA IIWR ASR2000, Paris, France, 2000",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Cepstral gain normalization for noise robust speech recognition",
"authors": [
{
"first": "S",
"middle": [],
"last": "Yoshizawa",
"suffix": ""
}
],
"year": 2004,
"venue": "Proc. IEEE International Conference on Acoustics, Speech and Signal Processing",
"volume": "1",
"issue": "",
"pages": "209--221",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "S. Yoshizawa et al., \"Cepstral gain normalization for noise robust speech recognition,\" in Proc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), vol. 1, pp. I-209-12, May 2004",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Higher orde r cepstral moment normalization (HOCMN) f or robust speech recognition",
"authors": [
{
"first": "C.-W",
"middle": [],
"last": "Hsu",
"suffix": ""
},
{
"first": "L.-S",
"middle": [],
"last": "Lee",
"suffix": ""
}
],
"year": 2004,
"venue": "Proc. IEEE International Conference on Acoustics, Speech and Signal Processing",
"volume": "17",
"issue": "",
"pages": "205--220",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "C.-W. Hsu and L.-S. Lee, \"Higher orde r cepstral moment normalization (HOCMN) f or robust speech recognition \", in Proc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), vol. 17, no. 2, pp. 205-220, Feb 2004",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "Cepstral shape norm alization (CSN) for robust speech recognition",
"authors": [
{
"first": "J",
"middle": [],
"last": "Du",
"suffix": ""
},
{
"first": "R.-H W",
"middle": [],
"last": "Ang",
"suffix": ""
}
],
"year": 2008,
"venue": "Proc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)",
"volume": "",
"issue": "",
"pages": "4389--4392",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "J. Du and R.-H W ang, \"Cepstral shape norm alization (CSN) for robust speech recognition\", in Proc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 4389-4392, April 2008",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "Entropy-based algor ithms for best basis selection",
"authors": [
{
"first": "R",
"middle": [
"R"
],
"last": "Coifm An",
"suffix": ""
},
{
"first": "M",
"middle": [
"V"
],
"last": "Ickerhauser",
"suffix": ""
}
],
"year": 1992,
"venue": "IEEE Trans. on Information Theory",
"volume": "38",
"issue": "2",
"pages": "713--718",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "R. R. Coifm an and M. V . W ickerhauser. \"Entropy-based algor ithms for best basis selection\", IEEE Trans. on Information Theory, vol. 38, no. 2, pp. 713-718, March 1992",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"uris": null,
"text": "L (z) : low-pass analysis filter H H (z) : high-pass analysis filter G L (z) : low-pass synthesis filter G H (z) : high-pass synthesis filter",
"num": null,
"type_str": "figure"
},
"TABREF0": {
"content": "<table><tr><td>\uf961(zero-crossing rate)\uf92d\u6c7a\u5b9a\u542b\u6709\u8a9e\u97f3\u6210\u5206\u7684\u4f4d\u7f6e\uff1b\u5728\u983b\u57df(frequency domain)\u4e0a\uff0c\u5247\u901a\u5e38 (cepstral mean subtraction, CMS)[8] \u3001\u5012\u983b\u8b5c\u5e73\u5747\u503c\u8207\u8b8a\uf962\uf969\u6b63\u898f\u5316\u6cd5(cepstral mean and</td></tr><tr><td>\u662f\u8a08\u7b97\u8a9e\u97f3\u983b\u8b5c\u7684\u71b5(entropy)\uf92d\u7372\u5f97\u8a9e\u97f3\u6210\u5206\u7684\u8cc7\u8a0a[3]\u3002\u800c\u5c0f\u6ce2\u5728\u6b64\u65b9\u5411\u4e0a\u6240\u63d0\u51fa\u7684\u6280 variance normalization, MVN)[9]\u3001\u5012\u983b\u8b5c\u5e73\u5747\u8207\u8b8a\uf962\uf969\u6b63\u898f\u5316\u6cd5\u7d50\u5408\u81ea\u52d5\u56de\u6b78\u52d5\u614b\u5e73\u5747</td></tr><tr><td>\u8853\u76f8\u5c0d\u8f03\u591a\uff0c\u8b6c\u5982\u5728\u6587\u737b[4]\u4e2d\u63d0\u5230\uf9ba\u4f7f\u7528\u5c0f\u6ce2\u8f49\u63db\u7684\u4fc2\uf969\u80fd\uf97e\u6bd4\uf9b5\u5224\u5b9a\u8a9e\u97f3\u53ca\u975e\u8a9e\u97f3 \uf984\u6ce2\u5668\u6cd5(cepstral m ean and variance norm alization p lus auto-regress ive m oving average</td></tr><tr><td>(non-speech)\u6210\u5206\uff0c\u6216\u662f\u5728\u53e6\u4e00[5]\u6587\u737b\uf9e8\u63d0\u51fa\u8a08\u7b97\u5c0f\u6ce2\u4fc2\uf969\u4e4b\u8b8a\uf962\uf969\uff0c\u5c07\u5176\u8996\u70ba\u4e00\u7d44\u96a8 filtering, MVA)[10]\u8207\u7d71\u8a08\u5716\u6b63\u898f\u5316\u6cd5(histogram equalization, HEQ)[11]\u7b49\u3002</td></tr><tr><td>\u6a5f\u8b8a\uf969(random variable) \u7d93\u7531\u6a5f\uf961\uf9e4\uf941\u4e4b\u7d50\u679c\u5224\u5b9a\uff0c\u6240\u5f97\u5206\uf9d0\u65b9\u6cd5\u76f8\u8f03\u65bc\u4e4b\u524d\u65b9\u5f0f\u80fd\uf901 \u4e0a\u8ff0\u5404\u7a2e\u7684\u6b63\u898f\u5316\u6280\u8853\u4e2d\uff0c\u7686\u662f\u628a\u55ae\u4e00\u7dad\u7279\u5fb5\u5e8f\uf99c\u4e4b\u6240\u6709\u7279\u5fb5\u8996\u70ba\u540c\u4e00\u500b\u96a8\u6a5f\u8b8a\uf969</td></tr><tr><td>\u7cbe\u78ba\u5224\u5225\u51fa\u8a9e\u97f3\u8ddf\u975e\u8a9e\u97f3\u4e4b\u6210\u4efd\u3002 \u7684\u53d6\u6a23(sample)\uff0c\u9032\u800c\u76f4\u63a5\u4f30\u6e2c\u6b64\u96a8\u6a5f\u8b8a\uf969\u4e4b\u7d71\u8a08\uf96b\uf969\uff0c\u8b6c\u5982\u671f\u671b\u503c(mean)\u3001\u8b8a\uf962\uf969</td></tr><tr><td>(\u4e09) \u5f37\u5065\u6027\u8a9e\u97f3\u7279\u5fb5\u64f7\u53d6(robust speech feature extraction) (variance)\u8207\u6a5f\uf961\u5206\u4f48(probability distribution)\u7b49\u3002\u96d6\u7136\u7a0b\u5e8f\u4e0a\uf9e0\u65bc\u5be6\u73fe\uff0c\u537b\u76f8\u5c0d\u5ffd\uf976\uf9ba\u4e00</td></tr><tr><td>\u6b64\uf9d0\u7684\u8a9e\u97f3\u8655\uf9e4\u6280\u8853\u65b9\u6cd5\u76ee\u7684\u662f\u64f7\u53d6\uf967\u5bb9\uf9e0\u53d7\u5230\u96dc\u8a0a\u5e72\u64fe\u7684\u8a9e\u97f3\u7279\u5fb5\uf96b\uf969\uff0c\u50b3\u7d71 \u6bb5\u8a9e\uf906\u4e4b\u4e2d\uff0c\u5176\u7279\u5fb5\u96a8\u6642\u9593\u8b8a\u5316\u7684\u7279\u6027\uff0c\uf9b5\u5982\u8abf\u8b8a\u983b\u8b5c\u7684\u8cc7\u8a0a\u3002\u5f9e\u53e6\u4e00\u89c0\u9ede\uf92d\u770b\uff0c\u9019\u4e9b</td></tr><tr><td>\u7684\u5f37\u5065\u6027\u8a9e\u97f3\u7279\u5fb5\u64f7\u53d6\u6280\u8853\u5927\u591a\uf969\u662f\u5728\u63a2\u8a0e\u8a9e\u97f3\u7279\u5fb5\u7684\u983b\u8b5c\u6027\u8cea\u9032\u800c\u767c\u5c55\u800c\u5f97\uff0c\u63db\uf906\u8a71 \u4f5c\u6cd5\u7b49\u540c\u65bc\u5c07\u5168\u90e8\u8abf\u8b8a\u983b\uf961\u4e4b\u6210\u4efd\u4e00\u4f75\u505a\u8655\uf9e4\u3002\u7136\u800c\u6839\u64da\u904e\u53bb\u8a31\u591a\u7684\u7814\u7a76\u767c\u73fe\uff0c\uf967\u540c\u7684</td></tr><tr><td>\uf96f\uff0c\u5176\u6240\u4f7f\u7528\u7684\u8f49\u63db\u6cd5\u70ba\u6709\u540d\u7684\u5085\uf9f7\uf96e\u8f49\u63db(Fourier transform)\u3002\u7136\u800c\u5c0f\u6ce2\u8655\uf9e4\u4e5f\u76f8\u7e7c\u61c9 \u8abf\u8b8a\u983b\u8b5c\u6210\u4efd\u5c0d\u65bc\u8a9e\u97f3\u8fa8\uf9fc\u64c1\u6709\uf967\u540c\u7684\u91cd\u8981\u6027\uff0c\uf901\u7cbe\u78ba\u5730\uf96f\uff0c\u5728 N.Kanedera \u5b78\u8005[12]</td></tr><tr><td>\u7528\u65bc\u5f37\u5065\u6027\u8a9e\u97f3\u7279\u5fb5\u64f7\u53d6\u6280\u8853\u4e0a\uff0c\uf9b5\u5982\uff0c\u5728[6]\u63d0\u51fa\u5c07\u539f\u59cb\u6885\u723e\u5012\u983b\u8b5c\u7279\u5fb5(mel-frequency \u8a73\u7d30\u6307\u51fa\u5927\u90e8\u5206\u7684\u8a9e\u97f3\u8fa8\uf9fc\u8cc7\u8a0a\u5206\u5e03\u5728 1 Hz \u548c 16 Hz \u7684\u8abf\u8b8a\u983b\uf961\u4e4b\u9593\uff0c\u4e14\u4e3b\u8981\u96c6\u4e2d\u5728 4</td></tr><tr><td>cepstral co efficients, M FCC)\u4e2d\u7684\uf9ea\u6563\u9918\u5f26\u8f49\u63db(discrete cosine trans form, DCT) \u7a0b\u5e8f\u6539\u8b8a Hz \u9644\u8fd1\u3002\u56e0\u6b64\uff0c\u8a31\u591a\u77e5\u540d\u4e14\u6210\u529f\u7684\u6642\u9593\u5e8f\uf99c\uf984\u6ce2\u5668(temporal filters)[13,14] \uff0c\u90fd\u662f\u7279\u5225</td></tr><tr><td>\u70ba\uf9ea\u6563\u5c0f\u6ce2\u8f49\u63db(discrete wavelet transform, DWT)\uff0c\u5176\uf941\u6587\u5448\u73fe\u7684\u5be6\u9a57\u7d50\u679c\u986f\u793a\u6240\u5f97\u5230\u7684 \u5f37\u8abf\u51fa\u9019\u4e9b\u91cd\u8981\u7684\u8abf\u8b8a\u983b\uf961\u6210\u5206\uff0c\u9032\u800c\u986f\u793a\u80fd\u6709\u6548\u6539\u5584\u96dc\u8a0a\u74b0\u5883\u4e0b\u8a9e\u97f3\u8fa8\uf9fc\u7684\u6548\u80fd\u3002</td></tr><tr><td>\u7279\u5fb5\u6bd4\u539f\u59cb MFCC \uf901\u5177\u6709\u96dc\u8a0a\u74b0\u5883\u4e4b\u5f37\u5065\u6027\u3002 \u800c\u524d\u9762\u4ecb\u7d39\u7684\u5404\u7a2e\u7279\u5fb5\u7d71\u8a08\u6b63\u898f\u5316\u6f14\u7b97\u6cd5\uff0c\u53ef\u80fd\u7f3a\u5931\u5728\u65bc\u7121\u6cd5\u6709\u6548\u7a81\u986f\uf967\u540c\u8abf\u8b8a\u983b</td></tr><tr><td>\u4e00\u3001\u7dd2\uf941 \u8fd1\uf98e\uf92d\uff0c\u8a9e\u97f3\u8655\uf9e4\u4e4b\uf9b4\u57df\u7684\u5b78\u8005\u6301\u7e8c\u5730\u958b\u767c\u7814\u7a76\uff0c\u4f7f\u8a9e\u97f3\u8655\uf9e4\u76f8\u95dc\uf9e4\uf941\u8207\u6280\u8853\uf967 \u65b7\u7cbe\u9032\u6210\u719f\uff0c\u9010\u6f38\u8da8\u65bc\u5be6\u969b\u61c9\u7528\u7684\u76ee\u7684\uff0c\u5c31\u8a9e\u97f3\u8fa8\uf9fc(speech recogn ition)\u800c\u8a00\uff0c\u5176\u7cfb\u7d71 \u5e38\u56e0\u6240\u5728\u74b0\u5883\u4e4b\u96dc\u8a0a\u5e72\u64fe\u6216\u662f\u50b3\u8f38\u901a\u9053\u7684\u6548\u61c9\uff0c\u800c\u4f7f\u8fa8\uf9fc\u6548\u80fd\u53d7\u5230\u660e\u986f\u5f71\u97ff\u3002\u91dd\u5c0d\u9019\u6a23 \u7684\u554f\u984c\uff0c\u8fd1\uf98e\uf92d\u7684\u7814\u7a76\u5b78\u8005\u63d0\u51fa\uf9ba\u4e00\u7cfb\uf99c\u7684\u74b0\u5883\u5f37\u5065\u6027(environmental robustness)\u6280\u8853\uff0c \uf961\u6210\u4efd\u5c0d\u65bc\u8a9e\u97f3\u8fa8\uf9fc\u7684\u91cd\u8981\u6027\uff0c\u56e0\u6b64\u6211\u5011\u5e0c\u671b\u80fd\u628a\u4e00\u7279\u5fb5\u6642\u9593\u5e8f\uf99c\u4e2d\u7684\uf967\u540c\u983b\uf961\u6210\u4efd\u5206 (\u56db) \u807d\u89ba\uf984\u6ce2\u5668\u8a2d\u8a08(auditory filter design) \uf9ea\u51fa\uf92d\uff0c\u9032\u800c\u500b\u5225\u8655\uf9e4\uff0c\u521d\u6b65\u7684\u69cb\u60f3\u662f\u80fd\u5c0d\u65bc\u8abf\u8b8a\u983b\uf961\u8f03\u91cd\u8981\u4e4b\u4f4e\u983b\u7684\u90e8\u4efd\u8f03\u7cbe\u7d30\u7684\u8655 \u4e00\u822c\u800c\u8a00\uff0c\u8a9e\u97f3\u8fa8\uf9fc\u4e2d\u7279\u5fb5\uf96b\uf969\u6c42\u53d6\u7a0b\u5e8f\uf9e8\u6240\u61c9\u7528\u7684\u8a9e\u97f3\u807d\u89ba\uf984\u6ce2\u5668\u7d44\u70ba\u6885\u723e\u5c3a \uf9e4\uff0c\u76f8\u5c0d\u6bd4\u8f03\uf967\u91cd\u8981\u4e4b\u9ad8\u983b\u7684\u90e8\u4efd\u5247\u4f7f\u7528\u8f03\u7c97\uf976\u7684\u65b9\u5f0f\u8655\uf9e4\u3002\u57fa\u65bc\u6b64\u76ee\u7684\uff0c\u6211\u5011\u767c\u73fe\u5c0f \ufa01(mel-scaled)\u7684\uf984\u6ce2\u5668\u7d44\uff0c\u9019\u4e9b\uf984\u6ce2\u5668\u5176\u5206\u4f48\u7279\u6027\u70ba\uff1a1 kHz \u983b\uf961\u4ee5\u4e0b\u70ba\u7dda\u6027\u5206\u4f48\uff0c1 kHz \u6ce2\u8f49\u63db\u662f\u500b\u5341\u5206\u6709\u7528\u7684\u5de5\u5177\uff0c\u512a\u9ede\u70ba\u5176\u80fd\u5c0d\u4e00\u983b\uf961\u5340\u57df\u4f5c\uf967\u7b49\u5206\u7684\ufa00\u5272\uff0c\u5373\u5c07\u8a0a\u865f\u5176\u8f03 \u4ee5\u4e0a\u983b\uf961\u70ba\u975e\u7dda\u6027\u5206\u4f48\uff0c\u5f7c\u6b64\u76f8\u4e92\u90e8\u5206\u91cd\u758a\uff0c\u5176\u53ef\u8fd1\u4f3c\u6a21\u64ec\u4eba\u8033\u807d\u89ba\u6548\u61c9\u3002\u76f8\u5c0d\u800c\u8a00\uff0c \u4f4e\u983b\uf961\u90e8\u5206\u4f7f\u7528\u8f03\u7a84\u7684\uf984\u6ce2\u5668\u904e\uf984\u51fa\uf92d\uff0c\u800c\u9ad8\u983b\u90e8\u5206\u5247\u7528\u8f03\u5bec\u7684\uf984\u6ce2\u5668\u5f97\u4e4b\uff0c\u4e4b\u5f8c\u5c0d\u65bc \u5c0f\u6ce2\u8655\uf9e4\u4e4b\u7814\u7a76\u5b78\u8005[7]\u4e5f\u63d0\u51fa\uf9ba\uf9dd\u7528\u5c0f\u6ce2\u5305(wavelet packet) \u7684\u7279\u6027\uf92d\u4eff\u6548\u4eba\u8033\u807d\u89ba\u6548 \u6bcf\u500b\u5b50\u983b\u5e36\u7684\u7279\u5fb5\u5e8f\uf99c\u4f5c\u7d71\u8a08\u6b63\u898f\u5316\u6cd5\u3002\u9019\u6a23\u7684\u7a0b\u5e8f\uff0c\u76f8\u8f03\u65bc\u50b3\u7d71\u7684\u5168\u983b\u5e36\u5f0f\u7684\u7279\u5fb5\u7d71 \u61c9\uff0c\u5176\u9069\u7576\u900f\u904e\u4e00\uf99a\uf905\u5c0f\u6ce2\u5305\u8f49\u63db\u6240\ufa00\u5272\u7684\u90e8\u4efd\u983b\u5e36\uff0c\u9078\u64c7\u51fa\u80fd\u8da8\u8fd1\u65bc\u4eba\u8033\u807d\u89ba\u7684\uf984\u6ce2 \u8a08\u6b63\u898f\u5316\u6cd5\uff0c\uf9e4\u61c9\u53ef\u4ee5\u9032\u4e00\u6b65\u63d0\u6607\u8655\uf9e4\u5f8c\u4e4b\u7279\u5fb5\u7684\u5f37\u5065\u6027\u3002\u4e4b\u5f8c\u4e00\u7cfb\uf99c\u7684\u7ae0\u7bc0\uff0c\u6211\u5011\u5c07 \u5668\u7d44\u6548\u61c9\uff0c\u800c\u7531\u65bc\u5c0f\u6ce2\u8655\uf9e4\u6240\u5f97\u4e4b\u5f7c\u6b64\u983b\u5e36\u9593\u90fd\u5047\u8a2d\u70ba\uf967\u76f8\u95dc\uff0c\u5373\u70ba\u4e92\uf967\u5f71\u97ff\uff0c\u56e0\u6b64\u6240 \u9010\u6b65\u4ecb\u7d39\u5c0f\u6ce2\u8f49\u63db\u4e4b\u5206\u983b\uf9e4\uf941\u4ee5\u53ca\u6240\u63d0\u51fa\u7684\u5206\u983b\u7279\u5fb5\u7d71\u8a08\u6b63\u898f\u5316\u6cd5\uff0c\u6700\u5f8c\u4ee5\u5be6\u9a57\u7d50\u679c\u8b49 \ufa00\u5272\u51fa\uf92d\u7684\u5404\u983b\uf961\u7bc4\u570d\u7684\u8a9e\u97f3\u4fe1\u865f\u90fd\u6db5\u84cb\uf9ba\u7368\uf9f7\u7684\u8fa8\uf9fc\u8cc7\u8a0a\uff0c\u5176\u4e2d\u7684\u5be6\u9a57\u7d50\u679c\u9a57\u8b49\uf9ba\u4ee5 \u5be6\u6b64\u5206\u983b\u5f0f\u6b63\u898f\u5316\u6cd5\u512a\u65bc\u50b3\u7d71\u4e4b\u5168\u983b\u5f0f\u6b63\u898f\u5316\u65b9\u6cd5\u3002 \u4e0a\u7684\u8655\uf9e4\u53ef\u4ee5\u512a\u65bc\u50b3\u7d71\u7684\u6885\u723e\uf984\u6ce2\u5668\u7d44\u8655\uf9e4\uff0c\u9054\u5230\u5c07\u8a9e\u97f3\u8fa8\uf9fc\u7cbe\u78ba\ufa01\u63d0\u5347\u7684\u76ee\u7684\u3002 \u85c9\u6b64\ufa09\u4f4e\u96dc\u8a0a\u6216\u901a\u9053\u5e72\u64fe\u6216\u51f8\u986f\u8a9e\u97f3\u7684\u7368\u7279\u6210\u4efd\uff0c\u800c\u9054\u5230\u660e\u986f\u7684\u6539\u9032\u6548\u679c\uff0c\u672c\uf941\u6587\u7684\u7814 \u7a76\u65b9\u5411\uff0c\u5373\u70ba\u958b\u767c\u51fa\u65b0\u7684\ufa09\u4f4e\u96dc\u8a0a\u8207\u901a\u9053\u5e72\u64fe\u4e4b\u76f8\u95dc\u7684\u8a9e\u97f3\u5f37\u5065\u6027\u6f14\u7b97\u6cd5\u3002\u7136\u800c\uff0c\u8ddf\u904e \u5728\u672c\uf941\u6587\u4e2d\uff0c\u6240\u767c\u5c55\u51fa\u7684\u65b0\u6280\u8853\uff0c\u4e26\uf967\u540c\u65bc\u4e0a\u8ff0\u6240\u63d0\u7684\u5e7e\u500b\u50b3\u7d71\u5c0f\u6ce2\u8655\uf9e4\u6240\u61c9\u7528 \u4e09\u3001\u5c0f\u6ce2\u8f49\u63db\u4e4b\u5206\u983b\u6280\u8853\uf9e4\uf941\u7684\u6982\u8ff0 \u7684\u65b9\u5411\uff0c\u800c\u662f\u8457\u91cd\u65bc\u5c07\u5c0f\u6ce2\u8655\uf9e4\u5176\u7279\u6b8a\u7684\u5206\u983b\u6280\u8853\u9069\u7576\u5730\u904b\u7528\u65bc\u8a9e\u97f3\u7279\u5fb5\u6642\u9593\u5e8f\uf99c \u53bb\u76f8\u95dc\u4e4b\u5f37\u5065\u6027\u6280\u8853\u8f03\u70ba\uf967\u540c\u7684\u662f\uff0c\u6211\u5011\u63a1\u7528\uf9ba\u5c0f\u6ce2\u8f49\u63db(wavelet transform)\uff0c\u5c0d\u65bc\u8a9e\u97f3 \u7279\u5fb5\u4e4b\u6642\u9593\u5e8f\uf99c(temporal trajectory)\u52a0\u4ee5\u8655\uf9e4\uff0c\uf92d\u6539\u5584\u8a9e\u97f3\u7279\u5fb5\u7684\u5f37\u5065\u6027\u3002 \u5728\u9019\u4e00\u7ae0\u4e2d\uff0c\u6211\u5011\u5c07\u5c08\u9580\u8a0e\uf941\u5c0f\u6ce2\u8f49\u63db\u904b\u7528\u65bc\uf9ea\u6563\u6642\u9593\u8a0a\u865f(discrete-time signal)\u7684 (temporal trajectory) \u4e0a\uff0c\u7d50\u5408\u5404\u7a2e\u7d71\u8a08\u6b63\u898f\u5316\u7684\u6280\u8853\uff0c\uf92d\u8655\uf9e4\u5c0f\u6ce2\u8f49\u5f8c\u5404\u5b50\u983b\u5e36\u7684\u7279\u5fb5 \u5206\u983b(frequency division) \u6280\u8853\uff0c\u6b64\u61c9\u7b97\u662f\u5c0f\u6ce2\u8f49\u63db\u6700\u5e38\u88ab\u7528\u4ee5\u8655\uf9e4\u8a0a\u865f\u7684\u65b9\u5411\u3002\u9996\u5148\u6211 \u6642\u9593\u5e8f\uf99c\uff0c\u5728\u4e4b\u5f8c\u7684\u7ae0\u7bc0\u4e2d\u6211\u5011\u5c07\u6703\u9010\u6b65\u4ecb\u7d39\u6b64\u65b0\u6280\u8853\uff0c\u5206\u6790\u5176\u4e3b\u8981\u89c0\uf9a3\u3001\u4f5c\u6cd5\u8207\u53ef\u80fd \u5c0f\u6ce2\u76f8\u95dc\uf9e4\uf941\u5728\u8a0a\u865f\u8655\uf9e4\u7684\u7bc4\u7587\u4e2d\u96d6\u5df2\u767c\u5c55\uf969\u5341\uf98e\uff0c\u7136\u800c\u76f8\u5c0d\u65bc\u5176\u4ed6\u8a31\u591a\uf9e4\uf941\u800c\u8a00\uff0c\u61c9 \u5011\u8003\u616e\u4e00\u7d44\u5178\u578b\u96d9\u901a\u9053\u7684\u6b63\u4ea4\u93e1\u50cf\uf984\u6ce2\u5668(quadrature-mirror filter bank, QMF)[15]\uff0c\u5982\u5716 \u512a\u65bc\u50b3\u7d71\u6280\u8853\u7684\u539f\u56e0\uff0c\u4e26\u4ee5\u4e00\u7cfb\uf99c\u7684\u5be6\u9a57\u8b49\u5be6\u6b64\u65b0\u6280\u8853\u76f8\u5c0d\u65bc\u50b3\u7d71\u76f8\u8fd1\u7684\u6280\u8853\u800c\u8a00\uff0c\uf901 \u7528\u65bc\u5728\u8a9e\u97f3\u5f37\u5065\u6027\u8655\uf9e4\u4e4b\uf9b4\u57df\u4e2d\u4ecd\u504f\u5c11\uf969\uff0c\u800c\u5176\u61c9\u7528\u7684\u65b9\u5411\u5927\u81f4\u4e0a\u4e3b\u8981\u5305\u542b\uf9ba\uff1a\u8a9e\u97f3\u5f37 \u672c\uf941\u6587\u5176\u9918\u7684\u7ae0\u7bc0\u6982\u8981\u5982\u4e0b\uff1a\u5728\u7b2c\u4e8c\u7ae0\uf9e8\uff0c\u4ecb\u7d39\u76ee\u524d\u5e38\u7528\u4e4b\u5f37\u5065\u6027\u7279\u5fb5\u7d71\u8a08\u6b63\u898f\u5316\u6cd5\u4e26 \u5316(speech enhancement)\u3001\u8a9e\u97f3\u7aef\u9ede\u5075\u6e2c(voice activity detection, VAD)\u3001\u5f37\u5065\u6027\u8a9e\u97f3\u7279\u5fb5 \u80fd\u6709\u6548\u63d0\u6607\u8a9e\u97f3\u8fa8\uf9fc\u5728\u96dc\u8a0a\u5e72\u64fe\u74b0\u5883\u4e0b\u7684\u7cbe\u78ba\u6027\u3002 \u4e09\u4e2d\u6240\u793a\uff1a</td></tr><tr><td>(robust speech feature)\u8207\u807d\u89ba\uf984\u6ce2\u5668\u8a2d\u8a08(auditory filter design)\u7b49\u3002\u6211\u5011\u5c07\u5b83\u5011\u7c21\u8ff0\u5982\u4e0b\uff1a \u63a2\u8a0e\u50b3\u7d71\u7d71\u8a08\u6b63\u898f\u5316\u6cd5\u4e4b\u53ef\u80fd\u7f3a\u5931\u3002\u5728\u7b2c\u4e09\u7ae0\uff0c\u6211\u5011\u5c07\u7c21\u8981\u4ecb\u7d39\uf9ea\u6563\u5c0f\u6ce2\u8f49\u63db\u4e4b\u5206\u983b\u6280 (\u4e00) \u8a9e\u97f3\u5f37\u5316(speech enhancement) \u8853\u7684\u5be6\u73fe\uff0c\u7b2c\u56db\u7ae0\u70ba\u672c\uf941\u6587\u7684\u91cd\u9ede\uff0c\u6211\u5011\u5c07\u5728\u6b64\u7ae0\u4e2d\u4ecb\u7d39\u6211\u5011\u6240\u63d0\u51fa\u7684\u65b0\u65b9\u6cd5\uff0c\u5373\uf978\u7a2e</td></tr><tr><td>\u8a9e\u97f3\u5f37\u5316\u4e3b\u8981\u76ee\u7684\uff0c\u901a\u5e38\u662f\u5728\u4e00\u6bb5\u8a0a\u865f\u4e2d\uff0c\u5c07\u96dc\u8a0a\u6291\u5236\uff0c\u4e26\u5c07\u8a9e\u97f3\u8a0a\u865f\u6210\u4efd\u5f37\u8abf\u51fa \u8abf\u8b8a\u983b\u8b5c\u57df\u7684\u5206\u983b\u7d71\u8a08\u7279\u5fb5\u88dc\u511f\u6cd5\uff1a\u5206\u983b\u5e36\u5e73\u5747\u503c\u8207\u8b8a\uf962\uf969\u6b63\u898f\u5316\u6cd5\u8207\u5206\u983b\u5e36\u7d71\u8a08\u5716\u7b49</td></tr><tr><td>\uf92d\uff0c\u5e38\u7528\u7684\u65b9\u5f0f\u662f\u5047\u8a2d\u96dc\u8a0a\u5728\u983b\u8b5c(spectrum)\u4e0a\u5177\u6709\u8f03\u70ba\u7a69\u614b(stationary)\u7684\u7279\u6027\uff0c\u5728\u983b\u57df \u5316\u6cd5\uff0c\u4e26\u5c0d\u5176\u521d\u6b65\u6548\u679c\u52a0\u4ee5\u4ecb\u7d39\u3002\u5728\u7b2c\u4e94\u7ae0\uff0c\u6211\u5011\u5c07\u57f7\ufa08\u4e00\u7cfb\uf99c\u7684\u8a9e\u97f3\u8fa8\uf9fc\u5be6\u9a57\uff0c\uf92d\u9a57</td></tr><tr><td>\u4e0a\u5c07\u96dc\u8a0a\u6210\u4efd\u6e1b\u4f4e\uff0c\uf9b5\u5982\u8a2d\u8a08\u4e00\uf984\u6ce2\u5668\uf92d\u904e\uf984\u96dc\u8a0a\u7b49\u3002\u800c\u4ee5\u76ee\u524d\u57fa\u65bc\u5c0f\u6ce2\u7684\u4fe1\u865f\u5f37\u5316\u65b9 \u7d50\uf941\uff0c\u53ca\u672a\uf92d\u53ef\u9032\u4e00\u6b65\u7814\u7a76\u7684\u65b9\u5411\u3002 \u6cd5\uff0c\u5176\u4e2d\u4e4b\u4e00\u70ba Donoho[1]\u5b78\u8005\u6240\u63d0\u51fa\u4f7f\u7528\u5c0f\u6ce2\u6536\u7e2e(wavelet shrinkage)\u7684\u65b9\u5f0f\uff0c\u5176\u65b9\u6cd5 \u8b49\u6240\u63d0\u4e4b\u65b0\u65b9\u6cd5\u8db3\u4ee5\u6709\u6548\u63d0\u6607\u8a9e\u97f3\u7279\u5fb5\u5728\u96dc\u8a0a\u74b0\u5883\u4e0b\u7684\u5f37\u5065\u6027\uff0c\u6700\u5f8c\uff0c\u7b2c\uf9d1\u7ae0\u5247\u70ba\u7c21\u8981 \u5716\u4e09 \u96d9\u901a\u9053 QMF \uf984\u6ce2\u5668\u7d44</td></tr><tr><td>\u662f\u7531\u5c0f\u6ce2\u8f49\u63db\u6240\u5f97\u4e4b\u4fc2\uf969\uff0c\u7d93\u7531\u9580\u6abb\u503c\u7684\u8a2d\u5b9a\u5c07\u96dc\u8a0a\u9069\ufa01\u5730\u6291\u5236\u3002\u5728\u5176\u76f8\u95dc\uf941\u6587\u4e4b\u5be6\u9a57 \u5176\u4e2d</td></tr><tr><td>\u7d50\u679c\u986f\u793a\uf9ba\uff0c\u900f\u904e\u5c0f\u6ce2\u8f49\u63db\u8655\uf9e4\u7684\u8a9e\u97f3\u5f37\u5316\u6548\u80fd\u6bd4\u8d77\u4e4b\u524d\u6240\u63d0\u51fa\u7684\u50b3\u7d71\u8a9e\u97f3\u5f37\u5316\u65b9\u6cd5 \u4e8c\u3001\u5404\u7a2e\u5f37\u5065\u6027\u6280\u8853\u4ecb\u7d39</td></tr><tr><td>[2]\u8981\uf92d\u7684\u597d\u3002 \u5728\u9019\uf9e8\u6211\u5011\u9996\u5148\u76ee\u524d\u5e38\u7528\u4e4b\u5f37\u5065\u6027\u7279\u5fb5\u7d71\u8a08\u6b63\u898f\u5316\u6cd5\uff0c\u4e4b\u5f8c\u63a2\u8a0e\u50b3\u7d71\u7d71\u8a08\u6b63\u898f\u5316\u6cd5 (\u4e8c) \u8a9e\u97f3\u7aef\u9ede\u5075\u6e2c(voice activity detection, VAD) \u4e4b\u53ef\u80fd\u7f3a\u5931\uff0c\u4e26\uf96f\u660e\u70ba\u4f55\u4f7f\u7528\u5c0f\u6ce2\u8f49\u63db(discrete wavelet transform, DWT)\u6539\u5584\u9019\u4e9b\u554f\u984c\u3002</td></tr><tr><td>\u7531\u65bc\u4e00\u6bb5\uf93f\u97f3(recording)\uf9e8\u53ef\u80fd\u5305\u542b\u6709\u975e\u8a9e\u97f3\u7684\u5340\u6bb5\uff0c\u5982\u679c\u4e00\u4f75\u8fa8\uf9fc\u6574\u6bb5\uf93f\u97f3\uff0c\u5c07 \u7531\u65bc\u8a9e\u97f3\u8fa8\uf9fc\u7cfb\u7d71\u5bb9\uf9e0\u53d7\u5230\u96dc\u8a0a\u74b0\u5883\u5f71\u97ff\u4f7f\u5f97\u5176\u8fa8\uf9fc\u6548\u80fd\ufa09\u4f4e\uff0c\u56e0\u6b64\u8a9e\u97f3\u8655\uf9e4\u76f8\u95dc</td></tr><tr><td>\u6703\u5f71\u97ff\u8fa8\uf9fc\u8655\uf9e4\u7684\u901f\ufa01\uff0c\u4e26\u53ef\u80fd\u9020\u6210\u8fa8\uf9fc\u7cbe\u78ba\ufa01\u660e\u986f\u4e0b\ufa09\u3002\u8a9e\u97f3\u7aef\u9ede\u5075\u6e2c(voice activity \u7814\u7a76\u7684\u5b78\u8005\u91dd\u5c0d\u6b64\u96dc\u8a0a\u5e72\u64fe\u7684\u554f\u984c\uff0c\u63d0\u51fa\u8af8\u591a\u7684\u5f37\u5065\u6027\u6280\u8853\uff0c\u9019\u4e9b\u6280\u8853\u4e2d\u6709\u4e00\u5927\uf9d0\u662f\u85c9</td></tr><tr><td>detection, endpoint detection)\u76f8\u95dc\u6280\u8853\u5373\u662f\u65bc\u6c7a\u5b9a\u51fa\u4e00\u6bb5\u8a0a\u865f\u4e2d\u771f\u6b63\u8a9e\u97f3\u5b58\u5728\u7684\u4f4d\u7f6e\u3002\u5728 \u7531\u6b63\u898f\u5316\u8a9e\u97f3\u7279\u5fb5\u7684\u7d71\u8a08\u7279\u6027\uff0c\uf92d\ufa09\u4f4e\u96dc\u8a0a\u5c0d\u8a9e\u97f3\u7279\u5fb5\u9020\u6210\u7684\u5931\u771f\u3002\u4ee5\u4e0b\u5c07\u4ecb\u7d39\u8fd1\uf98e\uf92d</td></tr><tr><td>\u50b3\u7d71\u7684\u4f5c\u6cd5\u4e0a\uff0c\u4ee5\u6642\u57df(time domain)\u800c\u8a00\uff0c\u900f\u904e\u8a08\u7b97\u4e00\u6bb5\u8a9e\u97f3\u4fe1\u865f\u7684\u80fd\uf97e(energy)\u6216\u904e\uf9b2 \u5728\u5f37\u5065\u6027\u8a9e\u97f3\u8fa8\uf9fc\u4e2d\u5e38\u7528\u7684\u5e7e\u7a2e\u8a9e\u97f3\u7279\u5fb5\u6b63\u898f\u5316\u6280\u8853\u3002\u5176\u4e2d\u5305\u542b\uf9ba\uff1a\u5012\u983b\u8b5c\u5e73\u5747\u6d88\u53bb\u6cd5</td></tr></table>",
"num": null,
"text": "\u95dc\u9375\u8a5e\uff1a\uf9ea\u6563\u5c0f\u6ce2\u8f49\u63db\u3001\u8a9e\u97f3\u8fa8\uf9fc\u3001\u5f37\u5065\u6027\u8a9e\u97f3\u7279\u5fb5\uf96b\uf969keywords: speech recognition, discrete wavelet transform, robust speech features",
"html": null,
"type_str": "table"
},
"TABREF1": {
"content": "<table><tr><td>\u7279\u5fb5\uf96b\uf96913\u7dad(c0~c12)\uff0c\u52a0\u4e0a\u4e00\u968e\u8207\u4e8c\u968e\u5dee\uf97e\uff0c\u7e3d\u5171\u70ba39\u7dad\u7279\u5fb5\uf96b\uf969\u3002\u5728\u8868\u4e09\u4e2d\uff0c (a) \uf967\u540c\u5f62\u5f0f\u4e4b\u5e73\u5747\u503c\u8207\u8b8a\uf962\uf969\u6b63\u898f\u5316\u6cd5\u7684\u8fa8\uf9fc\uf961\u6bd4\u8f03</td></tr><tr><td>HEQ\u3002\u76f8\u8f03\u65bc\u50b3\u7d71\u7684\u5168\u983b\u5e36\u7d71\u8a08\u88dc\u511f\u6cd5\uff0c\u6211\u5011\u6240\u63d0\u51fa\u4e4b\u5206\u983b\u5e36\u7d71\u8a08\u88dc\u511f \u6cd5\u6709\u4ee5\u4e0b\u5e7e\u9ede\u76f8\uf962\u4e4b\u8655\uff1a 1. \u50b3\u7d71\u7684\u5168\u983b\u5e36 MVN(FB-MVN)\u6cd5\u4e2d\uff0c\u4efb\u4e00\u7279\u5fb5\u5e8f\uf99c\u7684\u5e73\u5747\u503c\u8207\u8b8a\uf962\uf969\u901a\u5e38\u5206\u5225\u88ab\u6b63 \u898f\u5316\u70ba 0 \u8207 1\uff0c\u4f46\u5c0d\u65bc SB-MVN \u800c\u8a00\uff0c\uf967\u540c\u5206\u983b\u5e36\u7684\u7279\u5fb5\u5e8f\uf99c\u4e26\uf967\u64c1\u6709\u76f8\u540c\u7684\u76ee\u6a19 \u5e73\u5747\u503c\u8207\u8b8a\uf962\uf969\uff0c\u56e0\u6b64\uf967\u540c\u5206\u983b\u5e36\u7279\u5fb5\u5e8f\uf99c\u5373\u4f7f\u5728\u6b63\u898f\u5316\u5f8c\uff0c\u4ecd\u4fdd\u6709\u5f7c\u6b64\u7d71\u8a08\u7279\u6027 \u7684\u5dee\uf962\u3002\u76f8\u540c\u5730\uff0cSB-HEQ \u4e5f\u662f\u5177\u6709\u6b64\u7279\u6027\uff0c\uf967\u540c\u7684\u5206\u983b\u5e36\u7279\u5fb5\u5e8f\uf99c\u5c0d\u61c9\u81f3\uf967\u540c\u7684 \u76ee\u6a19\u6a5f\uf961\u5206\u4f48\u51fd\uf969\u3002 2. \u5728 SB-MVN \u8207 SB-HEQ \u4e2d\uff0c\u53ef\u4efb\u610f\u9078\u64c7\u67d0\u4e9b\u5206\u983b\u5e36\u5e8f\uf99c\uf92d\u4f5c\u6b63\u898f\u5316\u3002\u4e00\u822c\u800c\u8a00\uff0c\u5c0d \u65bc\u8a9e\u97f3\u8fa8\uf9fc\uf92d\uf96f\uff0c\u4f4e(\u8abf\u8b8a)\u983b\uf961\u7684\u6210\u5206\uff0c\u5305\u542b\u7684\u8a9e\u97f3\u9451\u5225\u8cc7\u8a0a\u8f03\u591a\uff0c\u56e0\u6b64\u6211\u5011\u901a\u5e38 \u512a \u5148 \u9078 \u64c7 \u4f4e \u983b \uf961 \u7684 \u5206 \u983b \u5e36 \u7279 \u5fb5 \u52a0 \u4ee5 \u6b63 \u898f \u5316 \u3002 \u4f46 \u662f \uff0c \u5982 \u679c \u6709 \u4e9b \u975e \u7a69 \u614b \u96dc \u8a0a (non-stationary noise) \u5b58\u5728\u65bc\u9ad8\u8abf\u8b8a\u983b\uf961\u7684\u5340\u57df\uff0c\u70ba\uf9ba\ufa09\u4f4e\u6b64\uf9d0\u96dc\u8a0a\u5e72\u64fe\uff0c\u5c31\u9808\u5c07\u9ad8 \u983b\u7684\u5206\u983b\u5e36\u8003\u616e\u9032\u53bb\u4e00\u540c\u8655\uf9e4\u3002 3. \u7531\u65bc DWT \u7a0b\u5e8f\u4e2d\u7684\ufa09\u4f4e\u53d6\u6a23(down-sampling)\u6b65\u9a5f\uff0c\u6211\u5011\u6240\u9700\u8655\uf9e4\u4e4b\u6240\u6709\u5206\u983b\u5e36\u5e8f \uf99c\u7684\u7279\u5fb5\u7e3d\uf969\u8fd1\u4f3c\u7b49\u540c\u65bc\u539f\u59cb\u5e8f\uf99c\u7684\u7279\u5fb5\u7e3d\uf969\uff0c\u56e0\u6b64\u8655\uf9e4\u4e0a\u4e26\uf967\u6703\u56e0\u70ba\u589e\u52a0\u5206\u983b\u5e36 \u7684\uf969\u76ee\u800c\u4f7f\u8a08\u7b97\u8907\u96dc\ufa01\u5927\u5e45\u63d0\u5347\u3002\u4f46\uf974\u4ee5\u50b3\u7d71\u7684\u5206\u983b\uf984\u6ce2\u5668\u7d44(filter-bank)\u4e4b\u65b9\u6cd5\uff0c \u6240\u9700\u8655\uf9e4\u7684\u7e3d\u7279\u5fb5\uf969\u6703\u660e\u986f\u96a8\u5206\u983b\u5e36\u7684\u500b\uf969\u800c\u589e\u52a0\uff0c\u76f8\u5c0d\u800c\u8a00\uff0c\u5176\u904b\u7b97\u7684\u8907\u96dc\ufa01\u6703 \u56e0\u6b64\u5927\u5e45\u63d0\u9ad8\u3002 (\u4e8c)\u5206\u983b\u5e36\u7279\u5fb5\u7d71\u8a08\u6b63\u898f\u5316\u6cd5\u7684\u521d\u6b65\u6548\u80fd\u8a0e\uf941 \u5728\u9019\uf9e8\uff0c\u6211\u5011\u5c07\u6240\u63d0\u51fa\u7684\u5206\u983b\u5e36\u7d71\u8a08\u6b63\u898f\u5316\u6cd5\u8ddf\u539f\u59cb\u4e4b\u5168\u983b\u5e36\u7d71\u8a08\u6b63\u898f\u5316\u6cd5\u4f5c\u521d\u6b65 \u7684\u6548\u80fd\u6bd4\u8f03\uff0c\u6839\u64da\u9019\u4e9b\u65b9\u6cd5\u5728\u4e00\u8a9e\u97f3\u7279\u5fb5\u5e8f\uf99c\u4e4b\u8abf\u8b8a\u983b\u8b5c\u7684\u5931\u771f\u6539\u5584\u7a0b\ufa01\uff0c\uf92d\u8a55\u4f30\u9019\u4e9b Hz]\u548c[25 Hz, 50 Hz])\u5247\u7dad\u6301\uf967\u52d5\uff0c\u800c SB-MVN (1,2,3,4) \u8207 SB-HEQ (1,2,3,4) \u8868\u793a\uf9ba\u5168\u90e8\u56db\u500b\u5206 \u983b\u5e36\u7686\u500b\u5225\u4ee5 MVN \u6216 HEQ \u8655\uf9e4\u3002 \u9996\u5148\uff0c\u6211\u5011\u5c0d\u65bc\u5168\u983b\u5e36\u8207\u5404\u7a2e\u5206\u983b\u5e36\u4e4b MVN \u6cd5\u7684\u8655\uf9e4\u7d50\u679c\u52a0\u4ee5\u8a0e\uf941\u3002\u5716\u4e03 (a)(b)(c)(d)\u5206\u5225\u8868\u793a\u70ba\u539f\u59cb\u672a\u8655\uf9e4\u4e4b\u7b2c\u4e00\u7dad MFCC( 1 c )\u7279\u5fb5\u5e8f\uf99c\u3001FB-MVN\u3001SB-MVN (1,2) \u8207 SB-MVN (1,2,3,4) \u8655\uf9e4\u5f8c\u4e4b 1 c \u5e8f\uf99c\u4e4b\u529f\uf961\u983b\u8b5c\u5bc6\ufa01(power spectral density, PSD)\u66f2\u7dda\u3002 \u5728\u5716\u4e03(a)\u4e2d\uff0c\u53ef\u770b\u51fa\uf967\u540c SNR \u503c\u4e0b(clean, 10 dB \u8207 0dB)\u4e4b\u672a\u8655\uf9e4\u904e\u7684 1 c \u5e8f\uf99c\uff0c\u5176 PSD \u66f2\u7dda\uff0c\u53d7\u5230\u52a0\u6210\u6027\u96dc\u8a0a(additive noise) \u7684\u5f71\u97ff\uff0c\u5b58\u5728\u56b4\u91cd\u7684\u5931\u771f\u60c5\u5f62\u3002\u800c\u7d93\u7531\u5716\u4e03 (b)\u53ef\u770b\u51fa\uff0cFB-MVN \u8655\uf9e4\u5f8c\u4e4b 1 c \u5e8f\uf99c\uff0c\u5728\u8f03\u4f4e\u7684\u8abf\u8b8a\u983b\uf961[0, 10 Hz]\u4e4b\u9593\uff0c\u5176 PSD \u5931\u771f \u7684\u60c5\u6cc1\u5f88\u660e\u986f\ufa09\u4f4e\uff0c\u4f46\u5728\u9ad8\u8abf\u8b8a\u983b\uf961\u7bc4\u570d[10Hz, 50 Hz]\uff0cPSD \u5931\u771f\u7684\u60c5\u5f62\u4e26\u6c92\u6709\u592a\u5927\u7684 \u6539\u5584\u3002\u5716\u4e03(c)\u70ba SB-MVN (1,2) \u6240\u5f97\u4e4b\u7279\u5fb5\u5e8f\uf99c\u4e4b PSD \u5716\uff0c\u5176\u6240\u8655\uf9e4\u7684\u983b\u5e36\u5206\u5225\u70ba[0, 6.25 Hz]\u548c[6.25 Hz, 12.5 Hz] \uff0c\u5f9e\u6b64\u5716\u53ef\u4ee5\u767c\u73fe\uff0c\u7d04\u5728\u8abf\u8b8a\u983b\uf961 20 Hz \u4ee5\u4e0b\uff0c\u5176 PSD \u5931\u771f\u60c5 \u5f62\u76f8\u5c0d\u6e1b\u4f4e\uff0c\u4f46\u5728\u672a\u8655\uf9e4\u7684\u8abf\u8b8a\u983b\uf961\u7bc4\u570d[12.5 Hz, 50 Hz]\uff0c\u540c\u6a23\u5b58\u6709\u660e\u986f\u7684\u5931\u771f\u60c5\u6cc1\u3002 \u5716\u4e03(d)\u70ba SB-MVN (1,2,3,4) \u6240\u5f97\u4e4b\u7279\u5fb5\u5e8f\uf99c\u4e4b PSD \u5716\uff0c\u5176\u6240\u8655\uf9e4\u7684\u983b\u5e36\u5206\u5225\u70ba[0, 6.25 Hz]\u3001[6.25 Hz, 12.5 Hz]\u3001[12.5 Hz, 25 Hz]\u8207[25 Hz, 50 Hz]\uff0c\u5f88\u660e\u986f\u53ef\u770b\u51fa\u5728\u5168\u90e8\u7684\u8abf\u8b8a \u983b\uf961\u7bc4\u570d\uff0c\u5176 PSD \u5931\u771f\u7684\u60c5\u6cc1\u7686\u6709\u6548\ufa09\u4f4e\u3002 \u63a5 \u4e0b \uf92d \uff0c \u5716 \u516b (a)(b)(c)(d) \u5206 \u5225 \u8868 \u793a \u70ba \u539f \u59cb \u672a \u8655 \uf9e4 \u4e4b \u7b2c \u4e00 \u7dad MFCC( 1 c ) \u7279 \u5fb5 \u5e8f \uf99c \u3001 FB-HEQ\u3001SB-HEQ (1,2) \u8207 SB-HEQ (1,2,3,4) \u8655\uf9e4\u5f8c\u4e4b 1 c \u5e8f\uf99c\u4e4b PSD \u66f2\u7dda\uff0c\u5176\u4e2d\u62ec\u5f27\u4e2d\u7684\uf969\u5b57 \u8868\u793a\u6240\u8655\uf9e4\u7684\u983b\u5e36\u3002\u6bd4\u8f03\u5716\u516b(a)\u8207\u5716\u516b(b)\u53ef\u77e5\uff0c\u5c0d\u65bc\u8f03\u4f4e\u7684\u8abf\u8b8a\u983b\uf961\u7bc4\u570d[0, 10 Hz] \uff0c \u7686\u5df2\u6709\u6548\ufa09\u4f4e\u3002\uf9d0\u4f3c\u4e4b\u524d\u7684\uf9fa\u6cc1\uff0c\u7576\u6bd4\u8f03\u5716\u516b(d)\u8207\u5716\u4e03(d)\u6642\uff0c\u53ef\u770b\u51fa SB-HEQ (1,2,3,4) \u5728 \ufa09\u4f4e PSD \u5931\u771f\u7684\u6027\u80fd\u4e0a\u512a\u65bc SB-MVN (1,2,3,4) \u3002 \u5716\u516b (a) \u539f\u59cb 1 c \u7279\u5fb5\u5e8f\uf99c\u3001(b)FB-HEQ\u3001(c)SB-HEQ (1,2) \u8207(d)SB-HEQ (1,2,3,4) \u4f5c\u7528\u5728\uf967\u540c\u8a0a \u96dc\u6bd4\u4e0b\u4e4b 1 c \u7279\u5fb5\u5e8f\uf99c\u4e4b\u529f\uf961\u983b\u8b5c\u5bc6\ufa01\u66f2\u7dda\u5716 \u4e94\u3001\u8abf\u8b8a\u983b\u8b5c\u5206\u983b\u5e36\u6b63\u898f\u5316\u6cd5\u7684\u8fa8\uf9fc\u5be6\u9a57\u7d50\u679c\u8207\u8a0e\uf941 \u672c\u7ae0\u4e3b\u8981\u5167\u5bb9\u70ba\u5448\u73fe\u4e26\u5206\u6790\u4e00\u7cfb\uf99c\u7684\u5f37\u5065\u6027\u7279\u5fb5\u6280\u8853\u6240\u5f97\u4e4b\u8a9e\u97f3\u8fa8\uf9fc\u7684\u6548\u679c\uff0c\u9019 \u4e9b\u6280\u8853\u5305\u62ec\uf9ba\u50b3\u7d71\u7684\u5168\u983b\u5f0f\u7279\u5fb5\u7d71\u8a08\u6b63\u898f\u5316\u6cd5\u3001\u6211\u5011\u6240\u65b0\u63d0\u51fa\u7684\u5206\u983b\u5f0f MVN(SB-MVN) \u6cd5\u8207\u5206\u983b\u5f0f HEQ(SB-HEQ)\u6cd5\u3002 (\u4e00)\u5be6\u9a57\u74b0\u5883\u8207\u67b6\u69cb\u8a2d\u5b9a \u672c \u8fa8 \uf9fc \u5be6 \u9a57 \u6240 \u63a1 \u7528 \u7684 \u8a9e \u97f3 \u8cc7 \uf9be \u5eab \u70ba \u6b50 \u6d32 \u96fb \u4fe1 \u6a19 \u6e96 \u5354 \u6703 (European \u8fa8\uf9fc\u7684\u5f71\u97ff\u3002\u7531\u65bc\u96dc\u8a0a\u7684\uf967\u540c\uff0c\u6e2c\u8a66\u74b0\u5883\u53ef\u5206\u70ba Set A\u3001Set B \u8207 Set C \u4e09\u7d44\u3002 \u5728\u8fa8\uf9fc\u4e2d\u6240\u4f7f\u7528\u7684\u8072\u5b78\u6a21\u578b\u662f\u7531\u96b1\u85cf\u5f0f\u99ac\u53ef\u592b\u6a21\u578b\u5de5\u5177(Hidden Markov Model Tool Kit, HTK)[17] (\u4e8c)\u5168\u983b\u5e36\u88dc\u511f\u6cd5\u8207\u5404\u7a2e\u5206\u983b\u5e36\u6b63\u898f\u5316\u6cd5\u4e4b\u5be6\u9a57\u7d50\u679c SB-MVN(1,2,3) SB-MVN(1,2,3,4) \u6211\u5011\u5448\u73fe\uf9ba\u57fa\u790e\u5be6\u9a57(baseline)\u8868\u4e09\u3001\u5404\u5206\u983b\u5e36\u65b9\u6cd5\u8207\u5168\u983b\u5e36\u65b9\u6cd5\u7684\u5e73\u5747\u8fa8\uf9fc\uf961(%)\u8207\u76f8\u5c0d\u932f\u8aa4\ufa09\u4f4e\uf961(%) Method Set A Set B Set C Avg. RR1 RR2 Baseline 71.92 68.22 77.61 71.58 --FB-MVN 85.03 85.56 85.60 85.36 48.49 -SB-MVN (1) 86.87 87.90 87.37 87.38 55.59 13.80 SB-MVN (1,2) 87.28 90.23 89.44 88.89 60.91 24.11 SB-MVN (1,2,3) 89.44 90.31 89.61 89.82 64.18 30.46 SB-MVN (1,2,3,4) 89.47 90.31 89.62 89.84 64.25 30.60 FB-HEQ 87.59 88.84 87.64 88.10 58.13 -SB-HEQ (1) 87.70 89.31 87.81 88.37 59.08 2.27 SB-HEQ (1,2) 89.22 90.55 90.23 89.95 64.64 15.55 SB-HEQ (1,2,3) 89.51 90.75 89.54 90.01 64.85 16.05 SB-HEQ (1,2,3,4) 89.51 90.83 89.57 90.05 64.99 16.39 \u8868\u56db\u3001\u6240\u6709\uf967\u540c SNR \u503c\u96dc\u8a0a\u74b0\u5883\u4e0b\u7684\u5e73\u5747\u8fa8\uf9fc\uf961(%) Method clean 20dB 15dB 10dB 5dB 0dB -5dB Baseline 99.79 95.80 88.15 73.81 56.32 43.82 40.13 FB-MVN 99.82 98.73 96.83 91.88 79.52 59.80 46.70 SB-MVN (1) 99.79 98.67 96.98 92.24 82.42 66.60 51.95 SB-MVN (1,2) 99.80 98.97 97.76 94.33 86.40 70.99 53.80 SB-HEQ (1,2,3,4) 99.64 98.85 97.69 94.74 87.15 71.83 54.01 84 85 86 87 88 89 90 91 Reccognition Accuracy(%) FB-MVN SB-MVN(1) SB-MVN(1,2) (b) \uf967\u540c\u5f62\u5f0f\u4e4b\u7d71\u8a08\u5716\u7b49\u5316\u6cd5\u7684\u8fa8\uf9fc\uf961\u6bd4\u8f03 84 85 86 87 88 89 90 91 Recognition Accuracy(%) FB-HEQ SB-HEQ (1) SB-HEQ (1,2) SB-HEQ(1,2,3) SB-HEQ (1,2,3,4) \u5716\u4e5d \u5404\u5206\u983b\u5e36\u65b9\u6cd5\u8207\u5168\u983b\u5e36\u65b9\u6cd5\u7684\u5e73\u5747\u8fa8\uf9fc\uf961(%)\u4e4b\u7d9c\u5408\u6bd4\u8f03\u5716 \u65b9\u6cd5\u7684\u6548\u80fd\u3002\u6211\u5011\u4f7f\u7528 AURORA-2 Hz, 25 \u5716\u4e03 (a) \u539f\u59cb 1 c \u7279\u5fb5\u5e8f\uf99c\u53ca(b)FB-MVN\u3001(c)SB-MVN (1,2) \u8207(d)SB-MVN (1,2,3,4) \u4f5c\u7528\u5728\uf967\u540c \u8a0a\u96dc\u6bd4\u4e0b\u4e4b 1 15 dB\u300110 dB\u30015 dB\u30010 dB \u8207-5 dB\uff0c\u56e0\u6b64\u6211\u5011\u53ef\u4ee5\u89c0\u5bdf\u5206\u6790\uf967\u540c\u96dc\u8a0a\u74b0\u5883\u4e0b\u5c0d\u65bc\u8a9e\u97f3 SB-HEQ (1,2,3) 99.66 98.84 97.70 94.68 87.09 71.74 53.78 c \u7279\u5fb5\u5e8f\uf99c\u4e4b\u529f\uf961\u983b\u8b5c\u5bc6\ufa01\u66f2\u7dda\u5716 Telecommunication S tandard Institute, ETSI) \u6240\u767c\ufa08\u7684\u8a9e\uf9be\u5eab AURORA-2[16]\uff0c\u5167\u5bb9\u662f\u4ee5 \u7f8e\u570b\u6210\uf98e\u7537\uf981\u6240\uf93f\u88fd\u7684\u4e00\u7cfb\uf99c\uf99a\u7e8c\u7684\u82f1\u6587\uf969\u5b57\u5b57\uf905\uff0c\u6e2c\u8a66\u8a9e\u97f3\u672c\u8eab\u52a0\u4e0a\u5404\u7a2e\u52a0\u6210\u6027\u96dc\u8a0a SB-MVN (1,2,3) 99.78 98.99 97.75 94.57 86.59 71.19 53.69 SB-MVN (1,2,3,4) 99.81 98.97 97.75 94.51 \u5f9e\u8868\u4e09\u3001\u8868\u56db\u548c\u5716\u4e5d\u53ef\u767c\u73fe\uff0c\u6211\u5011\u6240\u65b0\u63d0\u51fa\u7684\u5206\u983b\u5e36\u6b63\u898f\u5316\u6cd5\uff0c\u78ba\u5be6\u80fd\u6709\u6548\u63d0\u6607\u5176 86.60 71.34 53.72 \u96dc\u8a0a\u74b0\u5883\u4e0b\u7684\u5f37\u5065\u6027\uff0c\u5176\u8a73\u7d30\u73fe\u8c61\u5982\u4ee5\u4e0b\u5e7e\u9ede\uff1a \u6216\u901a\u9053\u6548\u61c9\u7684\u5e72\u64fe\u3002\u52a0\u6210\u6027\u96dc\u8a0a\u5171\u6709\u516b\u7a2e\uff0c\u5206\u5225\u662f\u5730\u4e0b\u9435(subway)\u3001\u4eba\u8072(babble)\u3001\u6c7d\uf902 \u6de8\u7121\u96dc\u8a0a\u7684\uf9fa\u614b(clean)\uff0c\u4ee5\u53ca\uf9d1\u7a2e\uf967\u540c\u96dc\u8a0a\u6bd4(signal to noise ratio, SNR)\uff0c\u5206\u5225\u662f 20 dB\u3001 SB-HEQ (1,2) 99.64 98.84 97.64 94.50 87.52 71.28 53.65 station)\u96dc\u8a0a\u7b49\uff1b\u800c\u901a\u9053\u6548\u61c9\u6709\uf978\u7a2e\uff0c\u5206\u5225\u70ba G712 \u548c MIRS\u3002\u96dc\u8a0a\u6bd4\uf9b5\u7684\u5927\u5c0f\u5305\u542b\uf9ba\u4e7e SB-HEQ (1) 99.72 98.72 97.31 93.34 83.98 68.49 52.70 (car)\u3001\u5c55\u89bd\u6703\u9928(exhibition)\u3001\u9910\u5ef3(restaurant)\u3001\u8857\u9053(street)\u3001\u98db\u6a5f\u5834(airport)\u548c\u706b\uf902\u7ad9(train FB-HEQ 99.77 99.01 97.76 94.22 84.30 65.21 48.96 1.</td></tr><tr><td>\u672c\u7ae0\u7bc0\u5be6\u9a57\u63a1\u7528\u6885\u723e\u5012\u983b\u8b5c\u4fc2\uf969(mel-frequency cepstral coefficients, MFCC)</td></tr></table>",
"num": null,
"text": "\u8cc7\uf9be\u5eab[20]\uf9e8\u7684 MAH_27O6571A \u8a9e\u97f3\u6a94\uff0c\u7136\u5f8c\u52a0\u5165 \uf967\u540c\u8a0a\u96dc\u6bd4(SNR)\u7684\u5730\u4e0b\u9435(subway)\u96dc\u8a0a\uff0c\u7e7c\u800c\u52a0\u4ee5\u8655\uf9e4\u3002 \u5728\u6211\u5011\u6240\u63d0\u51fa\u4e4b\u65b9\u6cd5\u4e2d\uff0c\u521d\u6b65\u4f7f\u7528\uf9ba\u4e09\u968e\u7684 DWT \u8f49\u63db\uff0c\u5c07\u6574\u500b\u8abf\u8b8a\u983b\u5e36[0, 50 Hz] \ufa00\u5272\u51fa\u56db\u7a2e\u5206\u983b\u5e36\u7bc4\u570d\uff0c\u5206\u5225\u662f[0, 6.25 Hz] \u3001[6.25 Hz, 12.5 Hz] \u3001[12.5 Hz, 25 Hz]\u548c[25 Hz, 50 Hz]\uff0c (\u7531\u65bc\u7279\u5fb5\u97f3\u6846\u53d6\u6a23\uf961\u70ba 100 Hz\uff0c\u56e0\u6b64\u7279\u5fb5\u5e8f\uf99c\u6db5\u84cb\u4e4b\u983b\uf961\u7bc4\u570d\u70ba[0, 50 Hz]) \u3002 \u5728\u4e4b\u5f8c\u8a0e\uf941\u7684\u6bcf\u500b\u983b\u5e36\u4e4b\u6b63\u898f\u6cd5\u4e2d\uff0c\u6211\u5011\u5728\u65b9\u6cd5\u540d\u7a31\u53f3\u4e0b\u65b9\u4f7f\u7528\u4e0b\u6a19\uf969\u5b57\uf92d\u8868\u793a\u88ab\u6b63\u898f \u5316\u7684\u983b\u5e36\uff0c\uf9b5\u5982 SB-MVN (1,2) \u8207 SB-HEQ (1,2) \u8868\u793a\uf9ba\u7b2c\u4e00\u500b\u5206\u983b\u5e36([0, 6.25 Hz] )\u8207\u7b2c\u4e8c \u500b\u5206\u983b\u5e36([6.25 Hz, 12.5 Hz])\u4f7f\u7528\uf9ba MVN \u6216 HEQ \u8655\uf9e4\uff0c\u5269\u9918\u7684\uf978\u500b\u9ad8\u983b\u5e36([12.5 FB-HEQ \u53ef\u6709\u6548\ufa09\u4f4e PSD \u4e4b\u5931\u771f\uff0c\u4f46\u5c0d\u65bc\u5176\u4ed6\u8abf\u8b8a\u983b\uf961\u7bc4\u570d[10 Hz, 50 Hz] \uff0cPSD \u5931\u771f \u7684\u60c5\u5f62\u4e26\u6c92\u6709\u7372\u5f97\u592a\u5927\u7684\u6539\u5584\u3002\u5716\u516b(c)\u70ba SB-HEQ (1,2) \u6240\u5f97\u4e4b\u7279\u5fb5\u5e8f\uf99c\u4e4b PSD \u5716\uff0c\u5176\u6240 \u8655\uf9e4\u7684\u983b\u5e36\u5206\u5225\u70ba[0, 6.25 Hz] \u8207[6.25 Hz, 12.5 Hz] \uff0c\u5728\u6b64\u5716\u4e2d\uff0c\u53ef\u4ee5\u767c\u73fe\u7d04\u5728\u8abf\u8b8a\u983b\uf961 20 Hz \u4ee5\u4e0b\u4e4b PSD \u5931\u771f\u73fe\u8c61\u76f8\u5c0d\u88ab\u6e1b\u4f4e\uff0c\u4f46\u5728\u5176\u4ed6\u8abf\u8b8a\u983b\uf961\u7bc4\u570d\uff0c\u4ecd\u6709\u660e\u986f\u7684\u5931\u771f\u60c5 \u6cc1\u3002\u8ddf\u4e4b\u524d\u5716\u4e03(c)SB-MVN (1,2) \u7684\u6548\u679c\u6bd4\u8f03\uff0c\u53ef\u770b\u51fa SB-HEQ (1,2) \u512a\u65bc SB-MVN (1,2) \uff0c\uf901 \u6709\u6548\ufa09\u4f4e\u7d04\u5728\u983b\uf961 20 Hz \u4ee5\u4e0b\u7684 PSD \u5931\u771f\ufa01\u3002\u5716\u516b(d)\u70ba SB-HEQ (1,2,3,4) \u6240\u5f97\u4e4b\u7279\u5fb5\u5e8f\uf99c \u4e4b PSD \u5716\uff0c\u5176\u6240\u8655\uf9e4\u7684\u983b\u5e36\u500b\u5225\u70ba[0, 6.25 Hz]\u3001[6.25 Hz, 12.5 Hz]\u3001[12.5 Hz, 25 Hz]\u8207 [25 Hz, 50 Hz] \uff0c\u5f9e\u6b64\u5716\u5f88\u660e\u986f\u53ef\u770b\u51fa\u5168\u90e8\u7684\u8abf\u8b8a\u983b\uf961\u7bc4\u570d\u4e4b PSD \u66f2\u7dda\uff0c\u5176\u5931\u771f\u7684\u60c5\u6cc1 \u8a13\uf996\u800c\u5f97\uff0c\u5305\u62ec\uf9ba 11 \u500b\uf969\u5b57\u6a21\u578b(zero, one, two,\u2026, nine \u53ca oh)\u4ee5\u53ca\u975c\u97f3 (silence)\u6a21\u578b\uff0c\u6bcf\u500b\uf969\u5b57\u6a21\u578b\u5247\u6709 16 \u500b\uf9fa\u614b\uff0c\u5404\uf9fa\u614b\u5305\u542b 20 \u500b\u9ad8\u65af\u5bc6\ufa01\u6df7\u5408\u3002 \u3001\u5404\u7a2e\u5206\u983b\u5f0f SB-MVN \u8207 SB-HEQ\u3001\u5168\u983b\u5f0f FB-MVN \u548c FB-HEQ \u4f5c\u7528\u5728\u539f\u59cb MFCC \u7279\u5fb5\u4e0a\u6240\u5f97\u7684\u5e73\u5747\u8fa8\uf9fc\u7d50\u679c(\uf967\u540c\u7a2e\u8fa8\uf9fc\u74b0\u5883\u7684\u5e73\u5747\u8fa8\uf9fc \uf961\u53ca\u76f8\u5c0d\u6539\u5584\uf961) \uff0c\u5176\u4e2d RR1\u548c RR2\u5206\u5225\u70ba\u76f8\u8f03\u65bc\u57fa\u790e\u5be6\u9a57\u548c\u5168\u983b\u5e36\u6cd5\u4e4b\u76f8\u5c0d\u932f\u8aa4\ufa09\u4f4e \uf961(relative error rate reductions)\u3002\u8868\u56db\uf99c\u51fa\u5728\u5404\u7a2e\uf967\u540c\u7684 SNR \u503c\u4e0b\u7684\u5404\u7a2e\u65b9\u6cd5\u7684\u5e73\u5747\u8fa8 \uf9fc\uf961\uff0c\u800c\u5716\u4e5d\u7c21\u8981\u756b\u51fa\u5404\u65b9\u6cd5\u5e73\u5747\u8fa8\uf9fc\uf961\u7684\u6bd4\u8f03\u5716\u3002 \u7121\uf941\u5168\u983b\u5e36\u8207\u5206\u983b\u5e36\u6b63\u898f\u5316\u65b9\u6cd5\uff0c\u76f8\u8f03\u65bc\u57fa\u672c\u5be6\u9a57\u800c\u8a00\uff0c\u90fd\u6709\uf97c\u597d\u7684\u6539\u5584\u6548\u80fd\uff0c\u76f8 \u5c0d\u932f\u8aa4\ufa09\u4f4e\uf961\u90fd\u5728 48%\u4ee5\u4e0a\uff0c\u9664\u6b64\u4e4b\u5916\uff0c\u6bcf\u4e00\u7a2e HEQ \u7684\u6548\u679c\u90fd\u6bd4\u5176\u76f8\u540c\u5f62\u5f0f\u7684 MVN \uf92d\u7684\u597d\u3002\u76f8\u8f03\u65bc MVN\uff0cHEQ \u984d\u5916\u5c0d\u65bc\u7279\u5fb5\u9ad8\u968e\u52d5\u5dee\u505a\u88dc\u511f\u8655\uf9e4\uff0c\u6240\u4ee5\u6574\u9ad4 \uf92d\uf96f\uff0cHEQ \uf901\u6709\u52a9\u65bc\u6539\u5584\u96dc\u8a0a\u74b0\u5883\u6240\u9020\u6210\u7684\u7279\u5fb5\u5931\u771f\u3002 2. SB-MVN \u7684\u56db\u7a2e\u5206\u983b\u6a21\u5f0f\u6548\u80fd\u90fd\u512a\u65bc\u539f\u59cb\u5168\u983b\u5f0f\u7684 FB-MVN\uff0c\u6b64\u60c5\u6cc1\u5728 SB-HEQ \u8207 FB-HEQ \u4e4b\u9593\u7684\u6bd4\u8f03\u4e5f\u662f\u5982\u6b64\u3002\u800c SB-MVN \u548c SB-HEQ \u76f8\u8f03\u65bc\u539f\u59cb FB-MVN \u548c FB-HEQ \u7684\u76f8\u5c0d\u932f\u8aa4\ufa09\u4f4e\uf961\u5206\u5225\u9ad8\u9054 30.60%\u8207 16.39%\uff0c\u6b64\u7d50\u679c\u986f\u793a\u6240\u63d0\u51fa\u7684\u65b0\u5206 \u983b\u8655\uf9e4\u6280\u8853\u512a\u65bc\u50b3\u7d71\u5168\u983b\u5e36\u7684\u8655\uf9e4\uff0c\u56e0\u6b64\u6211\u5011\u6210\u529f\u7684\u9a57\u8b49\uf9ba\u4e4b\u524d\u7ae0\u7bc0\u7684\u63a8\uf941\uff0c\u5373\uf967 \u540c\u7684\u8abf\u8b8a\u983b\u8b5c\u6210\u4efd\u5c0d\u65bc\u8a9e\u97f3\u8fa8\uf9fc\u6709\uf967\u540c\u7684\u91cd\u8981\u6027\uff0c\u5c0d\uf967\u540c\u983b\u5e36\u5206\u5225\u4f5c\u88dc\u511f\u53ef\u5e36\uf92d\uf901 \u597d\u7684\u6548\u80fd\u3002 3. \u5f9e\u8868\u56db\u4e2d\u89c0\u5bdf\u5728\uf967\u540c SNR \u503c\u60c5\u6cc1\u4e0b\u7684\u5e73\u5747\u8fa8\uf9fc\uf961\uff0c\u6211\u5011\u53ef\u77e5\u5728\uf967\u53d7\u4efb\u4f55\u96dc\u8a0a\u5e72\u64fe",
"html": null,
"type_str": "table"
}
}
}
} |