| { | |
| "paper_id": "O08-1017", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T08:02:38.857776Z" | |
| }, | |
| "title": "\u7d44\u5408\u5f0f\u5012\u983b\u8b5c\u7d71\u8a08\u6b63\u898f\u5316\u6cd5\u65bc\u5f37\u5065\u6027\u8a9e\u97f3\u8fa8\uf9fc\u4e4b\u7814\u7a76 Associative Cepstral Statistics Normalization Techniques for Robust Speech Recognition", | |
| "authors": [ | |
| { | |
| "first": "Wen-Hsiang", | |
| "middle": [], | |
| "last": "\u675c\u6587\u7965", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Tu", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Kuang-Chieh", | |
| "middle": [], | |
| "last": "\u5433\u5149\u6770", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Wu", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| } | |
| ], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "The noise robustness property for an automatic speech recognition system is one of the most important factors to determine its recognition accuracy under a noise-corrupted environment. Among the various approaches, normalizing the statistical quantities of speech features is a very promising direction to create more noise-robust features. The related feature normalization approaches include cepsral mean subtraction (CMS), cepstral mean and variance normalization (CMVN), histogram equalization (HEQ), etc. In addition, the statistical quantities used in these techniques can be obtained in an utterance-wise manner or a codebook-wise manner. It has been shown that in most cases, the latter behaves better than the former. In this paper, we mainly focus on two issues. First, we develop a new procedure for developing the pseudo-stereo codebook, which is used in the codebook-based feature normalization approaches. The resulting new codebook is shown to provide a better estimate for the features statistics in order to enhance the performance of the codebook-based approaches. Second, we propose a series of new feature normalization approaches, including associative CMS (A-CMS), associative CMVN (A-CMVN) and associative HEQ (A-HEQ). In these approaches, two sources of statistic information for the features, the one from the utterance and the other from the codebook, are properly integrated. Experimental results show that these new feature normalization approaches perform significantly better than the conventional utterance-based and codebook-based ones. As the result, the proposed methods in this paper effectively improve the noise robustness of speech features.", | |
| "pdf_parse": { | |
| "paper_id": "O08-1017", | |
| "_pdf_hash": "", | |
| "abstract": [ | |
| { | |
| "text": "The noise robustness property for an automatic speech recognition system is one of the most important factors to determine its recognition accuracy under a noise-corrupted environment. Among the various approaches, normalizing the statistical quantities of speech features is a very promising direction to create more noise-robust features. The related feature normalization approaches include cepsral mean subtraction (CMS), cepstral mean and variance normalization (CMVN), histogram equalization (HEQ), etc. In addition, the statistical quantities used in these techniques can be obtained in an utterance-wise manner or a codebook-wise manner. It has been shown that in most cases, the latter behaves better than the former. In this paper, we mainly focus on two issues. First, we develop a new procedure for developing the pseudo-stereo codebook, which is used in the codebook-based feature normalization approaches. The resulting new codebook is shown to provide a better estimate for the features statistics in order to enhance the performance of the codebook-based approaches. Second, we propose a series of new feature normalization approaches, including associative CMS (A-CMS), associative CMVN (A-CMVN) and associative HEQ (A-HEQ). In these approaches, two sources of statistic information for the features, the one from the utterance and the other from the codebook, are properly integrated. Experimental results show that these new feature normalization approaches perform significantly better than the conventional utterance-based and codebook-based ones. As the result, the proposed methods in this paper effectively improve the noise robustness of speech features.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Abstract", | |
| "sec_num": null | |
| } | |
| ], | |
| "body_text": [ | |
| { | |
| "text": "\u672c\uf941\u6587\u6240\u8a0e\uf941\u53ca\u63d0\u51fa\u7684\u5f37\u5065\u5f0f\u6280\u8853\uff0c\u4e3b\u8981\u662f\u5728\u52a0\u6210\u6027\u96dc\u8a0a\u74b0\u5883\u4e0b\uff0c\u5c0d\u8a13\uf996\u8207\u6e2c\u8a66\u4e8c\u8005\u7684 \u8a9e\u97f3\u7279\u5fb5\uf96b\uf969\u7684\u7d71\u8a08\u7279\u6027\u52a0\u4ee5\u6b63\u898f\u5316\uff0c\u4ee5\ufa09\u4f4e\uf978\u74b0\u5883\u7684\uf967\u5339\u914d\u3002\u5176\u4e2d\u6211\u5011\uf9dd\u7528\u6885\u723e\u5012\u983b \u8b5c\u4fc2\uf969(mel-frequency cepstral coefficients, MFCC)\u505a\u70ba\u8a9e\u97f3\u7279\u5fb5\uff0c\u7d50\u5408\u8a9e\u97f3\u5075\u6e2c\u6280\u8853 (voice activity detection, VAD) [1] \u8207\u7279\u5fb5\u7d71\u8a08\u503c\u6b63\u898f\u5316\u7684\u8af8\u591a\u6280\u8853\uff0c\uf92d\u63d0\u5347\u8a9e\u97f3\u7279\u5fb5\u5728\u52a0 \u6210\u6027\u96dc\u8a0a\u74b0\u5883\u4e0b\u7684\u5f37\u5065\u6027\u3002\u672c\uf941\u6587\u4e2d\u6240\u8a0e\uf941\u7684\u7279\u5fb5\uf96b\uf969\u6b63\u898f\u5316\u6cd5\u5206\u5225\u70ba\uff1a (\u4e00)\u6574\u6bb5\u5f0f(utterance-based)\u7279\u5fb5\uf96b\uf969\u6b63\u898f\u5316\u6cd5 \u5373 \u50b3 \u7d71 \u7684 \u6574 \u6bb5 \u5f0f \u5012 \u983b \u8b5c \u5e73 \u5747 \u6d88 \u53bb \u6cd5 (utterance-based cepstral mean subtraction, U-CMS) [2] \u3001\u6574\u6bb5\u5f0f\u5012\u983b\u8b5c\u5e73\u5747\u503c\u8207\u8b8a\uf962\uf969\u6b63\u898f\u5316\u6cd5(utterance-based cepstral mean and variance normalization, U-CMVN) [3] \u8207\u6574\u6bb5\u5f0f\u7d71\u8a08\u5716\u7b49\u5316\u6cd5(utterance-based histogram equalization, U-HEQ) [4] \u3002\u6b64\uf9d0\u65b9\u6cd5\u662f\u4ee5\u4e00\u6574\u6bb5\u8a9e\uf906\u70ba\u57fa\u6e96\u53bb\u4f30\u7b97\u6bcf\u4e00\u7dad\u7279\u5fb5\uf96b\uf969\u7684\u7d71\u8a08 \u7279\u6027\uff0c\u4e26\u57f7\ufa08\u7279\u5fb5\uf96b\uf969\u6b63\u898f\u5316\u3002 (\u4e8c)\u78bc\u7c3f\u5f0f(codebook-based)\u7279\u5fb5\uf96b\uf969\u6b63\u898f\u5316\u6cd5 \u6b64\uf9d0\u65b9\u6cd5\u662f\u85c9\u7531\u78bc\u7c3f\uf92d\u5e6b\u52a9\u6211\u5011\u4f30\u7b97\u51fa\u4ee3\u8868\u8a13\uf996\u8a9e\u97f3\u7279\u5fb5\u8207\u6e2c\u8a66\u8a9e\u97f3\u7279\u5fb5\u7684\u7d71\u8a08 \u503c\uff0c\u85c9\u6b64\u57f7\ufa08\u8a9e\u97f3\u7279\u5fb5\u6b63\u898f\u5316\u3002\u5728\u904e\u53bb\u7684\u7814\u7a76\uf9e8 [5] [6] [7] \uff0c\u767c\u73fe\u6b64\uf9d0\u7684\u65b9\u6cd5\uff0c\u5305\u62ec\u78bc\u7c3f \u5f0f\u5012\u983b\u8b5c\u5e73\u5747\u6d88\u53bb\u6cd5(codebook-based cepstral mean subtraction, C-CMS)\u8207\u78bc\u7c3f\u5f0f\u5012\u983b\u8b5c \u5e73\u5747\u503c\u8207\u8b8a\uf962\uf969\u6b63\u898f\u5316\u6cd5(codebook-based cepstral mean and variance normalization, C-CMVN)\u7b49\uff0c\u5176\u6548\u679c\u90fd\u6bd4\u524d\u4e00\uf9d0\u4e4b\u6574\u6bb5\u5f0f\u7279\u5fb5\u6b63\u898f\u5316\u6cd5\uf92d\u7684\u597d\u3002 \u672c\uf941\u6587\u6839\u64da\u4ee5\u4e0a\u6240\u8ff0\u7684\u4e8c\uf9d0\u65b9\u6cd5\u63d0\u51fa\u4e00\u7cfb\uf99c\u6539\u9032\u7684\u6280\u8853\uff0c\u5206\u8ff0\u5982\u4e0b\uff1a \u2460 \u5728\u904e\u53bb\u78bc\u7c3f\u5f0f\u7279\u5fb5\u6b63\u898f\u5316\u6cd5\u4e2d [5] [6] [7] [2] \u3001\u6574 \u6bb5 \u5f0f \u5012 \u983b \u8b5c \u5e73 \u5747 \u503c \u8207 \u8b8a \uf962 \uf969 \u6b63 \u898f \u5316 \u6cd5 (utterance-based cepstral mean and variance normalization, U-CMVN) [3] \u8207 \u6574 \u6bb5 \u5f0f \u5012 \u983b \u8b5c \u7d71 \u8a08 \u5716 \u7b49 \u5316 \u6cd5 (utterance-based cepstral histogram equalization, U-HEQ) [4] ", | |
| "cite_spans": [ | |
| { | |
| "start": 172, | |
| "end": 175, | |
| "text": "[1]", | |
| "ref_id": "BIBREF0" | |
| }, | |
| { | |
| "start": 347, | |
| "end": 350, | |
| "text": "[2]", | |
| "ref_id": "BIBREF1" | |
| }, | |
| { | |
| "start": 436, | |
| "end": 439, | |
| "text": "[3]", | |
| "ref_id": "BIBREF2" | |
| }, | |
| { | |
| "start": 498, | |
| "end": 501, | |
| "text": "[4]", | |
| "ref_id": "BIBREF3" | |
| }, | |
| { | |
| "start": 635, | |
| "end": 638, | |
| "text": "[5]", | |
| "ref_id": "BIBREF4" | |
| }, | |
| { | |
| "start": 639, | |
| "end": 642, | |
| "text": "[6]", | |
| "ref_id": "BIBREF5" | |
| }, | |
| { | |
| "start": 643, | |
| "end": 646, | |
| "text": "[7]", | |
| "ref_id": "BIBREF6" | |
| }, | |
| { | |
| "start": 875, | |
| "end": 878, | |
| "text": "[5]", | |
| "ref_id": "BIBREF4" | |
| }, | |
| { | |
| "start": 879, | |
| "end": 882, | |
| "text": "[6]", | |
| "ref_id": "BIBREF5" | |
| }, | |
| { | |
| "start": 883, | |
| "end": 886, | |
| "text": "[7]", | |
| "ref_id": "BIBREF6" | |
| }, | |
| { | |
| "start": 887, | |
| "end": 890, | |
| "text": "[2]", | |
| "ref_id": "BIBREF1" | |
| }, | |
| { | |
| "start": 993, | |
| "end": 996, | |
| "text": "[3]", | |
| "ref_id": "BIBREF2" | |
| }, | |
| { | |
| "start": 1080, | |
| "end": 1083, | |
| "text": "[4]", | |
| "ref_id": "BIBREF3" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "\u4e00\u3001\u7dd2\uf941", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "\u3002 (\u4e00)\u6574\u6bb5\u5f0f\u5012\u983b\u8b5c\u5e73\u5747\u6d88\u53bb\u6cd5 (U-CMS) \u5012\u983b\u8b5c\u5e73\u5747\u6d88\u53bb\u6cd5(CMS)\u7684\u76ee\u7684\u662f\u5e0c\u671b\u4e00\u8a9e\u97f3\u7279\u5fb5\u5e8f\uf99c\u4e2d\uff0c\u6bcf\u4e00\u7dad\ufa01\u7684\u5012\u983b\u8b5c\u4fc2\uf969 \u9577\u6642\u9593\u5e73\u5747\u503c\u70ba0\u3002\u5047\u8a2d\u5176\u503c\uf967\u70ba0\u6642\uff0c\u6211\u5011\u5c31\u5c07\u6b64\u8996\u70ba\u901a\u9053\u96dc\u8a0a\u800c\u52a0\u4ee5\u6263\u9664\uff0c\u6b64\u7a2e\u65b9\u6cd5 \u5c0d\u65bc\ufa09\u4f4e\u901a\u9053\u96dc\u8a0a\u6548\u61c9\u662f\u4e00\u7a2e\u7c21\u55ae\u4e14\u6709\u7528\u7684\u6280\u8853\uff0c\u4f46\u662f\u6709\u6642\u5c0d\u65bc\ufa09\u4f4e\u52a0\u6210\u6027\u96dc\u8a0a\u4e0a\u4e5f\u6709 \u4e00\u5b9a\u7684\u6548\u679c\u3002\u5728\u591a\uf969\u7684\u4f5c\u6cd5\u4e0a\uff0c\u9996\u5148\u6211\u5011\u5c07\u6574\u6bb5\u8a9e\u97f3\u6bcf\u4e00\u7dad\u7684\u5012\u983b\u8b5c\u4fc2\uf969\u53d6\u5e73\u5747\u503c\uff0c\u7136 \u5f8c\u5c07\u6bcf\u4e00\u7dad\u7684\u4fc2\uf969\u6e1b\u6389\u5176\u5e73\u5747\u503c\uff0c\u5982\u6b64\u5373\u5f97\u5230\u88dc\u511f\u5f8c\u4e4b\u65b0\u7279\u5fb5\uff0c\u6b64\u7a31\u70ba\u6574\u6bb5\u5f0f\u5012\u983b\u8b5c\u5e73 \u5747\u6d88\u53bb\u6cd5(utterance-based cepstral mean subtraction, U-CMS)\u3002\u6839\u64da\u9019\u6a23\u7684\u539f\u5247\uff0c\u6211\u5011\u5047\u8a2d [ ] { } , 1, 2,..., X n n N = \u70ba\u4e00\u6bb5\u8a9e\u97f3\u6240\u64f7\u53d6\u5230\u7684\u67d0\u4e00\u7dad\u5012\u983b\u8b5c\u7279\u5fb5\uf96b\uf969\u5e8f\uf99c\uff0c\u5728\u7d93\u904e\u6574\u6bb5 \u5f0f \u5012 \u983b \u8b5c \u5e73 \u5747 \u6d88 \u53bb \u6cd5 (U-CMS) \u8655 \uf9e4 \u5f8c \uff0c \u5f97 \u5230 \u65b0 \u7684 \u7d93 \u904e \u88dc \u511f \u7684 \u7279 \u5fb5 \uf96b \uf969 \u5e8f \uf99c [ ] { } , 1, 2,..., U CMS X n n N \u2212 = \uff0c\u5176\uf969\u5b78\u5f0f\u5982\u4e0b\u6240\u793a\uff1a [ ] [ ] , 1, 2,..., . U CMS X X n X n n N \u03bc \u2212 = \u2212 = \u5f0f(2.1) \u5176\u4e2d [ ] 1 1 N X n X n N \u03bc = = \u2211 , N \u70ba\u6574\u6bb5\u8a9e\u97f3\u7684\u97f3\u6846\u500b\uf969\u3002 \u56e0\u6b64\uff0c\u5728 U-CMS \u6cd5\u4e2d\uff0c\u7528\u4ee5\u6b63\u898f\u5316\u7684\u5e73\u5747\u503c X \u03bc \u662f\u7531\u539f\u59cb\u6574\u6bb5\u7684\u7279\u5fb5\u5e8f\uf99c\u6240\u5f97\u3002 (\u4e8c)\u6574\u6bb5\u5f0f\u5012\u983b\u8b5c\u5e73\u5747\u503c\u8207\u8b8a\uf962\uf969\u6b63\u898f\u5316\u6cd5 (U-CMVN", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "\u4e00\u3001\u7dd2\uf941", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "{ } , 1, 2,..., X n n N = \u662f\u4e00\u6bb5\u8a9e\u97f3\u7684 \u67d0 \u4e00 \u7dad \u5012 \u983b \u8b5c \u7279 \u5fb5 \uf96b \uf969 \u5e8f \uf99c \uff0c \u5728 \u7d93 \u904e U-CMVN \u8655 \uf9e4 \u5f8c \uff0c \u5f97 \u5230 \u65b0 \u7684 \u7279 \u5fb5 \uf96b \uf969 [ ] { } , 1, 2,..., U CMVN X n n N \u2212 = \uff0c\u5176\uf969\u5b78\u5f0f\u5982\u4e0b\u6240\u793a\uff1a [ ] [ ] , 1, 2,..., X U CMVN X X n X n n N \u03bc \u03c3 \u2212 \u2212 = = . \u5f0f(2.2) \u5176\u4e2d 1 1 [ ] N X n X n N \u03bc = = \u2211 , ( ) 2 1 1 [ ] N X X n X n N \u03c3 \u03bc = = \u2212 \u2211 \u56e0\u6b64\uff0c\u5728U-CMVN\u4e2d\uff0c\u6240\u7528\u7684\u5e73\u5747\u503c X \u03bc \u8207\u6a19\u6e96\u5dee X \u03c3 \u7686\u7531\u6574\u6bb5\u8a9e\u97f3\u7684\u7279\u5fb5\u5e8f\uf99c\u800c\u5f97\u3002 (\u4e09)\u6574\u6bb5\u5f0f\u7d71\u8a08\u5716\u7b49\u5316\u6cd5(U-HEQ) \u7d71\u8a08\u5716\u7b49\u5316\u6cd5(HEQ)\u7684\u76ee\u7684\uff0c\u662f\u5e0c\u671b\u7528\u4ee5\u8a13\uf996\u8207\u6e2c\u8a66\u4e4b\u8a9e\u97f3\u7279\u5fb5\uf978\u8005\u80fd\u5920\u5177\u6709\u76f8\u540c \u7684\u7d71\u8a08\u5206\u4f48\u7279\u6027\uff0c\u85c9\u7531\u6b64\u5339\u914d\u7684\u8f49\u63db\u904e\u7a0b\uff0c\ufa09\u4f4e\u6e2c\u8a66\u7279\u5fb5\u8207\u8a13\uf996\u7279\u5fb5\u4e4b\u9593\u7531\u65bc\u96dc\u8a0a\u5f71\u97ff \u6240\u9020\u6210\u7684\uf967\u5339\u914d\u60c5\u5f62\u3002\u5176\u4f5c\u6cd5\u662f\u5c07\u6e2c\u8a66\u8a9e\u97f3\u7279\u5fb5\u8207\u8a13\uf996\u8a9e\u97f3\u7279\u5fb5\u7684\u6a5f\uf961\u5206\u4f48\u540c\u6642\u903c\u8fd1\u4e00 \uf96b\u8003\u6a5f\uf961\u5206\u4f48\u3002\u5728\u672c\uf941\u6587\u4e2d\u6240\u4f7f\u7528\u7684\uf96b\u8003\u6a5f\uf961\u5206\u4f48\u70ba\u4e00\u6a19\u6e96\u5e38\u614b\u5206\u4f48\u3002 \u6839\u64da\u4e0a\u8ff0\uff0c\u6211\u5011\u5047\u8a2d [ ] { } , 1, 2,..., X n n N = \u70ba\u4e00\u6bb5\u8a9e\u97f3\u67d0\u4e00\u7dad\u5012\u983b\u8b5c\u7279\u5fb5\uf96b\uf969\u5e8f \uf99c \uff1b ( ) X F x \u70ba [ ] X n \u7684 \u6a5f \uf961 \u5206 \u4f48 ( ) ( ) ( ) X F x P X x = \u2264 \uff0c \u5b83 \u662f \u7531 \u6574 \u6bb5 \u4e4b \u7279 \u5fb5 [ ] { } , 1, 2,..., X n n N = \u6c42 \u5f97 \uff1b ( ) N F x \u70ba \uf96b \u8003 \u6a5f \uf961 \u5206 \u4f48 \u3002 \u5247 \u6574 \u6bb5 \u5f0f \u7d71 \u8a08 \u5716 \u7b49 \u5316 \u6cd5 (utterance-based histogram equalization, U-HEQ)\u7684\uf969\u5b78\u8f49\u63db\u5f0f\u5982\u4e0b\u6240\u793a\uff1a [ ] [ ] ( ) ( ) 1 , U HEQ N X X n F F X n \u2212 \u2212 = \u5f0f(2.3) \u5176\u4e2d [ ]", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "\u4e00\u3001\u7dd2\uf941", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "\u5f8c\uff0c\u5efa\uf9f7\u4e00\u7d44\u5305\u542b M \u500b\u78bc\u5b57\u7684\u96c6\u5408\uff0c\u4ee5 [ ] { } | 1 x n n M \u2264 \u2264 \uf92d \u8868\u793a\uff0c\u540c\u6642\uff0c\u5176\u5c0d\u61c9\u7684\u6b0a\u91cd\u70ba{ } | 1 n w n M \u2264 \u2264 \u3002\u9019\u7d44\u5728\u4e2d\u4ecb\u7279\u5fb5\uf96b\uf969\u57df\u4e0a\u7684\u4e7e\u6de8\u8a9e\u97f3 \u78bc\u7c3f\u4e4b\u6240\u6709\u78bc\u5b57\uff0c\u518d\u7531 MFCC \u64f7\u53d6\uf9ca\u7a0b\u7684\u5f8c\u534a\u90e8\u8f49\u63db\u81f3\u5012\u983b\u8b5c\u57df\uff0c\u5982\u4e0b\u5f0f\u6240\u793a\uff1a ( ) [ ] [ ] x n f x n = \u223c \u5f0f(3.1) \u5176\u4e2d (.) f \u4ee3\u8868\u8f49\u63db\u7a0b\u5e8f\uff0c\u56e0\u6b64\uff0c [ ] { } , |1 n x n w n M \u2264 \u2264 \u70ba\u8f49\u63db\u81f3\u5012\u983b\u8b5c\u7684\u78bc\u7c3f\u53ca\u6b0a\u91cd \u503c\uff0c\u9019\u5c31\u662f\u4e7e\u6de8\u8a9e\u97f3\u7684\u5012\u983b\u8b5c\u78bc\u7c3f\u53ca\u6b0a\u91cd\u503c\u3002 \u5728\u96dc\u8a0a\u8a9e\u97f3\u65b9\u9762\uff0c\u6211\u5011\u85c9\u7531\u4e7e\u6de8\u8a9e\u97f3\u5728\u4e2d\u4ecb\u7279\u5fb5\uf96b\uf969\u57df\u4e0a\u7684\u78bc\u5b57\uff0c\uf92d\u5efa\uf9f7\u5c0d\u61c9\u81f3\u8a72 \u6bb5\u542b\u96dc\u8a0a\u4e4b\u6e2c\u8a66\u8a9e\u97f3\u7684\u78bc\u7c3f\u3002\u6211\u5011\u5c07\u6bcf\u4e00\u6e2c\u8a66\u8a9e\u97f3\u4f30\u6e2c\u5230\u7684\u7d14\u96dc\u8a0a\uff0c\u5728\u4e2d\u4ecb\u7279\u5fb5\uf96b\uf969\u57df (\u7dda\u6027\u983b\u8b5c\u57df)\u4e0a\u7528\u4e00\u7d44\u5411\uf97e [ ] { } | 1 n p p P \u2264 \u2264 \uf92d\u8868\u793a\uff0c\u7531\u65bc\u4e7e\u6de8\u8a9e\u97f3\u8207\u7d14\u96dc\u8a0a\u5728\u4e2d\u4ecb \u7279\u5fb5\uf96b\uf969\u57df\u4e0a\u5177\u6709\u7dda\u6027\u76f8\u52a0\u7684\u7279\u6027\uff0c\u56e0\u6b64\u96dc\u8a0a\u8a9e\u97f3\u7684\u78bc\u5b57\u53ef\u8868\u793a\u6210\u4e0b\u5f0f\uff1a [ ] [ ] [ ] { } ( 1) | , m n P p y m x n n p = \u2212 + = + \u5f0f(3.2) \u6700\u5f8c\uff0c\uf9d0\u4f3c\u5f0f(3.1)\uff0c\u6211\u5011\u5c07 [ ] y m \u7d93\u7531 MFCC \u64f7\u53d6\uf9ca\u7a0b\u5f8c\u534a\u90e8\u8f49\u63db\u81f3\u5012\u983b\u8b5c\u57df\uff0c\u5982\u4e0b\u5f0f \u6240\u793a\uff1a [ ] [ ] ( ), y m f y m = \u5f0f(3.3) \u6b64\u5916\uff0c\u6bcf\u500b [ ] y m \u7684\u6b0a\u91cd\u503c m v \u5247\u8a2d\u5b9a\u70ba\uff1a ( 1) , n m m n P p w v P = \u2212 + = \u5f0f(3.4) \u56e0\u6b64\uff0c [ ] y m \u4e4b\u6b0a\u91cd(\u5373 m v )\u662f\u5176\u5c0d\u61c9\u7684\u4e7e\u6de8\u8a9e\u97f3\u78bc\u5b57 [ ] x n \u4e4b\u6b0a\u91cd n w \u7684 1 P \uff0c\u5176\u4e2d P \u662f \u7d14\u96dc\u8a0a\u5411\uf97e { } [ ] n p \u7684\u500b\uf969\u3002\u6545 { } [ ], |1 m y m v m MP \u2264 \u2264 \uf965\u662f\u4ee3\u8868\u6b64\uf906\u96dc\u8a0a\u8a9e\u97f3\u5728\u5012\u983b\u8b5c \u57df\u4e0a\u7684\u78bc\u7c3f\u53ca\u6b0a\u91cd\u503c\u3002 [ ] { } , n x n w \u8207 [ ] { } , m y m v \u9019\uf978\u7d44\u5206\u5225\u4ee3\u8868\u4e7e\u6de8\u8a13\uf996\u8a9e\u97f3\u8207\u96dc\u8a0a\u6e2c \u8a66\u8a9e\u97f3\u7684\u78bc\u5b57\uff0c\u6211\u5011\u7a31\u4e4b\u70ba\u865b\u64ec\u96d9\u901a\u9053\u78bc\u7c3f\u3002\u6240\u8b02\u865b\u64ec\u7684\u610f\u601d\uff0c\u662f\u56e0\u70ba\u96dc\u8a0a\u8a9e\u97f3\u7684\u78bc\u7c3f \u4e26\uf967\u662f\u76f4\u63a5\u7531\u96dc\u8a0a\u8a9e\u97f3\u5f97\u5230\uff0c\u800c\u662f\u7d93\u7531\u4e7e\u6de8\u8a9e\u97f3\u78bc\u7c3f\u8207\u7d14\u96dc\u8a0a\u4f30\u7b97\u503c\u6240\u9593\u63a5\u5f97\u5230\u7684\u3002 (\u4e8c)\u78bc\u7c3f\u5f0f\u7279\u5fb5\uf96b\uf969\u6b63\u898f\u5316\u6280\u8853 \u9019\u4e00\u7bc0\u4e2d\uff0c\u6211\u5011\u5c07\u4ecb\u7d39\u78bc\u7c3f\u5f0f\u7279\u5fb5\uf96b\uf969\u6b63\u898f\u5316\u6280\u8853\u3002\u5728\u524d\u9762\u66fe\u63d0\u5230\uff0c\u6b64\uf9d0\u6b63\u898f\u5316\u6280 \u8853\uff0c\u662f\u540c\u6642\u91dd\u5c0d\u4e7e\u6de8\u8a9e\u97f3\u8207\u96dc\u8a0a\u8a9e\u97f3\u5012\u983b\u8b5c\u7279\u5fb5\uf96b\uf969\u4f5c\u8655\uf9e4\u3002\u800c\u5728\u9019\uf9e8\u7684\u78bc\u7c3f\u5f0f\u7279\u5fb5\uf96b \uf969\u6b63\u898f\u5316\u6280\u8853\uff0c\u662f\u85c9\u7531\u5728\u524d\u4e00\u7bc0\u4e2d\u63cf\u8ff0\u7684\u865b\u64ec\u96d9\u901a\u9053\u78bc\u7c3f\uff0c\uf92d\u5efa\uf9f7\u7279\u5fb5\u4e4b\u7d71\u8a08\uf97e\uff0c\u9032\u800c \u5c0d\u7279\u5fb5\u505a\u6b63\u898f\u5316\u3002\u9019\u4e09\u7a2e\u7279\u5fb5\uf96b\uf969\u6b63\u898f\u5316\u6280\u8853\u5206\u5225\u70ba\uff1a\u5012\u983b\u8b5c\u5e73\u5747\u6d88\u53bb\u6cd5(CMS)\u3001\u5012\u983b \u8b5c\u5e73\u5747\u503c\u8207\u8b8a\uf962\uf969\u6b63\u898f\u5316\u6cd5(CMVN)\u3001\u8207\u5012\u983b\u8b5c\u7d71\u8a08\u5716\u7b49\u5316\u6cd5(HEQ)\u3002\u5c0d\u65bcCMS\u8207CMVN \u800c\u8a00\uff0c\u6211\u5011\uf9dd\u7528\u524d\u4e00\u7bc0\u6240\u8ff0\u4e4b\u78bc\u7c3f\u8207\u6b0a\u91cd [ ] { } , m x m w \u8207 [ ] { } , m y m v \uff0c\u8a08\u7b97\u51fa\u5206\u5225\u4ee3\u8868\u4e7e \u6de8\u8a9e\u97f3\u8207\u96dc\u8a0a\u8a9e\u97f3\u7279\u5fb5\u7684\u8fd1\u4f3c\u7d71\u8a08\u503c\uff0c\u5982\u4e0b\u5f0f\u6240\u793a\uff1a , 1 ( [ ]) , N X i n i n w x n \u03bc = \u2248 \u2211 ( ) ( ) 2 2 2 , , 1 [ ] . N X i n X i i n w x n \u03c3 \u03bc = \u2248 \u2212 \u2211 \u5f0f(3.5) , 1 ( [ ]) , NP Y i m i m v y m \u03bc = \u2248 \u2211 ( ) ( ) 2 2 2 , , 1 [ ] . NP Y i m Y i i m v y m \u03c3 \u03bc = \u2248 \u2212 \u2211 \u5f0f(3.6) \u5176\u4e2d( ) i u \u4ee3\u8868\u4efb\u610f\u5411\uf97eu \u4e4b\u7b2ci \u7dad\uff0c , X i \u03bc \u8207 2 , X i \u03c3 \u5206\u5225\u4ee3\u8868\u4e7e\u6de8\u8a9e\u97f3\u7279\u5fb5\u5411\uf97e x \u7b2ci \u7dad\u7684\u5e73\u5747 \u503c\u8207\u8b8a\uf962\uf969\uff1b , Y i \u03bc \u8207 2 , Y i \u03c3 \u5206\u5225\u4ee3\u8868\u96dc\u8a0a\u8a9e\u97f3\u7279\u5fb5\u5411\uf97ey \u7b2ci", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "\u4e00\u3001\u7dd2\uf941", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "EQUATION", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [ | |
| { | |
| "start": 0, | |
| "end": 8, | |
| "text": "EQUATION", | |
| "ref_id": "EQREF", | |
| "raw_str": ") , ( ) ( ) . i i X i i i y i x x y y \u03bc \u03bc = \u2212 = \u2212 \u5f0f(3.7) \u5176\u4e2d x \u8207y \u5206\u5225\u70ba\u4e7e\u6de8\u8a9e\u97f3\u7279\u5fb5 x \u8207\u96dc\u8a0a\u8a9e\u97f3\u7279\u5fb5y \u5728\u7d93\u904e C-CMS \u8655\uf9e4\u5f8c\u7684\u65b0\u7279\u5fb5\u503c\u3002 \u800c \u78bc \u7c3f \u5f0f \u5012 \u983b \u8b5c \u5e73 \u5747 \u503c \u8207 \u8b8a \uf962 \uf969 \u6b63 \u898f \u5316 \u6cd5 (codebook-based cepstral mean and variance normalization, C-CMVN)\uff0c\u662f\u91dd\u5c0d\u5012\u983b\u8b5c\u7279\u5fb5\u4e4b\u5e73\u5747\u503c\u8207\u8b8a\uf962\uf969\u505a\u6b63\u898f\u5316\u8655\uf9e4\uff0c \u5176\uf969\u5b78\u8868\u793a\u5f0f\u5982\u4e0b\uff1a , ,", | |
| "eq_num": "( ) (" | |
| } | |
| ], | |
| "section": "\u4e00\u3001\u7dd2\uf941", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "EQUATION", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [ | |
| { | |
| "start": 0, | |
| "end": 8, | |
| "text": "EQUATION", | |
| "ref_id": "EQREF", | |
| "raw_str": ") ( ) , ( ) . i Xi i Yi i i X i Y i x y x y \u03bc \u03bc \u03c3 \u03c3 \u2212 \u2212 = = \u5f0f(3.8) \u5176\u4e2d x \u8207y \u5206\u5225\u70ba\u4e7e\u6de8\u8a9e\u97f3\u7279\u5fb5 x \u8207\u96dc\u8a0a\u8a9e\u97f3\u7279\u5fb5y \u7d93\u904e C-CMVN \u8655\uf9e4\u5f8c\u7684\u65b0\u7279\u5fb5\u503c\u3002 \u6700\u5f8c\uff0c\u78bc\u7c3f\u5f0f\u5012\u983b\u8b5c\u7d71\u8a08\u5716\u7b49\u5316\u6cd5(codebook-based cepsteral histogram equalization, C-HEQ)\uff0c\u5176\u57fa\u672c\u4f5c\u6cd5\u662f\uf9dd\u7528 [ ] { } , n x n w \u8207 [ ] { } , m y m v \uf978\u5957\u78bc\u7c3f\u5206\u5225\u8a08\u7b97\u51fa\u4e7e\u6de8\u8a9e\u97f3\u7279\u5fb5 \u8207\u96dc\u8a0a\u8a9e\u97f3\u7279\u5fb5\u4e4b\u6bcf\u4e00\u7dad\u4e4b\u8fd1\u4f3c\u7684\u6a5f\uf961\u5206\u4f48(probability distribution)\uff0c\u7136\u5f8c\u6c42\u4e00\u8f49\u63db\u51fd \uf969\uff0c\u4f7f\u4e8c\u8005\u4e4b\u6bcf\u4e00\u7dad\u7279\u5fb5\uf96b\uf969\u4e4b\u6a5f\uf961\u5206\u4f48\u7686\u903c\u8fd1\u65bc\u67d0\u4e00\u4e8b\u5148\u5b9a\u7fa9\u4e4b\uf96b\u8003\u6a5f\uf961\u5206\u4f48\u3002\u5177\u9ad4 \u4f5c\u6cd5\u5982\u4e0b\u63cf\u8ff0\uff1a \u5047\u8a2d\u6211\u5011\u73fe\u5728\u85c9\u7531\u78bc\u7c3f { } [ ], n x n w \u5efa\uf9f7\u7b2ci \u7dad\u4e7e\u6de8\u8a9e\u97f3\u7279\u5fb5( ) i x \u7684\uf94f\u7a4d\u5bc6\ufa01\u51fd\uf969\uff0c\u7531\u65bc\u78bc \u7c3f\u672c\u8eab\u610f\u5473\u8457\uf9ea\u6563\u7684\u5f62\u5f0f\uff0c\uf974\u6211\u5011\u5047\u8a2d\u7b2c ( ) i x \u672c\u8eab\u5c0d\u61c9\u4e4b\u96a8\u6a5f\u8b8a\uf969\u70ba i X \uff0c\u5247 i X \u7684\u6a5f\uf961 \u8cea\uf97e\u51fd\uf969(probability mass function)\u53ef\u7528\u4e0b\u5f0f\u8868\u793a\uff1a ( ) ( ) [ ] , i n i P X x n w = = \u5f0f(3.9) \u800c i X \u7684\u6a5f\uf961\u5bc6\ufa01\u51fd\uf969(probability density function, pdf)\uff0c\u5373\u53ef\u4ee5\u4e0b\u5f0f\u8868\u793a\uff1a ( ) ( ) 1 ( ) [ ] ; i M X n i n f x w x x n \u03b4 = = \u2212 \u2211 \u5f0f(3.10) \u5176\u4e2d ( ) \u03b4 \u22c5 \u70ba\u55ae\u4f4d\u8108\u885d(unit impulse)\u51fd\uf969\uff0c\u6545 i X \u4e4b\u6a5f\uf961\u5206\u4f48\uff0c\u6216\u7a31\u70ba\uf94f\u7a4d\u6a5f\uf961\u5bc6\ufa01\u51fd\uf969 (cumulative density function)\uff0c\u70ba\u4e0a\u5f0f ( ) i X f x \u4e4b\u7a4d\u5206\uff0c\u8868\u793a\u5982\u4e0b\uff1a ( ) ( ) ( ) 1 ( ) [ ] ; i M X i n i n F x P X x w u x x n = = \u2264 = \u2212 \u2211 \u5f0f(3.11) \u5176\u4e2d ( ) u x \u70ba\u55ae\u4f4d\u6b65\u968e\u51fd\uf969(unit step function)\uff0c\u5b9a\u7fa9\u70ba\uff1a ( ) 1, 0 0, 0 { x x u x \u2265 < = \u5f0f(3.12) \u56e0\u6b64\uff0c\u7b2ci \u7dad\u4e7e\u6de8\u8a9e\u97f3\u7279\u5fb5( ) i x \u4e4b\u6a5f\uf961\u5206\u4f48\u5247\u53ef\u7531\u5f0f(3.11)\u7684 ( ) i X F x \u8868\u793a\uff0c\u540c\uf9e4\uff0c\u85c9\u7531\u78bc \u7c3f { } [ ], m y m v \u5efa\uf9f7\u4e4b\u7b2ci \u7dad\u96dc\u8a0a\u8a9e\u97f3\u7279\u5fb5( ) i y \u7684\u6a5f\uf961\u5206\u4f48\u53ef\u7531\u4e0b\u5f0f\u8868\u793a\uff1a ( ) ( ) ( ) ( ) 1 [ ] ; i MP Y i m i m F y P Y y v u y y m = = \u2264 = \u2212 \u2211 \u5f0f(3.13) \u7531\u4e0a\u8ff0\u4f5c\u6cd5\u5f97\u5230 ( ) i X F x \u8207 ( ) i Y F y \u4e4b\u5f8c\uff0c\u6839\u64da\u5012\u983b\u8b5c\u7d71\u8a08\u5716\u7b49\u5316\u6cd5(HEQ)\u7684\u539f\uf9e4\uff0c\u6211\u5011\uf9dd \u7528\u4e0b\u9762\uf978\u5f0f\u5206\u5225\u6b63\u898f\u5316\u7b2ci \u7dad\u4e4b\u8a13\uf996\u4e7e\u6de8\u8a9e\u97f3\u7279\u5fb5 ( ) i x \u8207\u6e2c\u8a66\u96dc\u8a0a\u8a9e\u97f3\u7279\u5fb5( ) i y \uff1a ( ) ( ) ( ) ( ) 1 i i N X i x F F x \u2212 = , \u5f0f(3.14) ( ) ( ) ( ) 1 ( ) i N Y i i y F F y \u2212 = . \u5f0f(3.15) \u5176\u4e2d ( ) N F i \u70ba\u4e00\uf96b\u8003\u6a5f\uf961\u5206\u4f48(\u901a\u5e38\u70ba\u6a19\u6e96\u5e38\u614b\u5206\u4f48)\uff0c ( ) 1 N F \u2212 i \u70ba ( ) N F i \u7684\u53cd\u51fd\uf969\uff0cx", | |
| "eq_num": ", , ( ) (" | |
| } | |
| ], | |
| "section": "\u4e00\u3001\u7dd2\uf941", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "EQUATION", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [ | |
| { | |
| "start": 0, | |
| "end": 8, | |
| "text": "EQUATION", | |
| "ref_id": "EQREF", | |
| "raw_str": "\u9019\u4e00\u7bc0\u4e2d\u5c07\u5206\u5225\u4ecb\u7d39 A-CMS \u8207 A-CMVN \uf978\u7a2e\u7279\u5fb5\uf96b\uf969\u6b63\u898f\u5316\u6cd5\u3002\u6211\u5011\u85c9\u7531\u4e00\uf96b \uf969\u503c \u03b1 \u7684\u8abf\u6574\uff0c\u9069\u7576\u5730\u6574\u5408\u78bc\u7c3f\u8207\u6574\u6bb5\u7279\u5fb5\u4e4b\u7d71\u8a08\u8cc7\u8a0a\uff0c\u5e0c\u671b\u80fd\u9054\u5230\u8f03\u4f73\u4e4b\u8fa8\uf9fc\u6548\u679c\u3002 \u5c31\u6574\u6bb5\u8a9e\uf906(utterance)\u7684\u7279\u5fb5\u800c\u8a00\uff0c\u5047\u8a2d { } 1 2 , ,..., N X X X X = \u70ba\u4e00\u6bb5\u8a13\uf996\u7528\u6216\u6e2c\u8a66\u7528\u8a9e \u97f3\u5728\u6240\u64f7\u53d6\u5230\u7684\u67d0\u4e00\u7dad\u5012\u983b\u8b5c\u7279\u5fb5\uf96b\uf969\u5e8f\uf99c\uff0c\u5247\u5176\u6574\u6bb5\u5f0f\u4e4b\u7279\u5fb5\u7684\u5e73\u5747\u503c\u8207\u8b8a\uf962\uf969\u53ef\u7531 \u4e0b\uf978\u5f0f\u8a08\u7b97\u800c\u5f97\uff1a 1 1 , N u i i X N \u03bc = = \u2211 \u5f0f(4.1) ( ) 2 2 1 1 , N u i u i X N \u03c3 \u03bc = = \u2212 \u2211 \u5f0f(4.2) \u5176\u4e2d u \u03bc \u70ba\u6574\u6bb5\u5f0f\u4e4b\u7279\u5fb5\u5e73\u5747\u503c\uff0c 2 u \u03c3 \u70ba\u6574\u6bb5\u5f0f\u4e4b\u7279\u5fb5\u8b8a\uf962\uf969\uff0c N \u70ba\u6574\u6bb5\u8a9e\u97f3\u7684\u97f3\u6846\uf969\u3002 \u800c\u5728\u78bc\u7c3f\u4e0a\u7684\u7279\u5fb5\u65b9\u9762\uff0c\u5047\u8a2d { } 1 2 , ,..., M C C C C = \u70ba\u540c\u4e00\u6bb5\u8a9e\u97f3\u5c0d\u61c9\u5230\u7684\u5404\u78bc\u5b57 (codewords)\u7684\u67d0\u4e00\u7dad(\u8207\u524d\u4e00\u6bb5\u6240\u8ff0\u4e4b\u7dad\u503c\u76f8\u540c)\u4e4b\u96c6\u5408\uff0c\u5247\u6b64\u6bb5\u8a9e\u97f3\u7279\u5fb5\u4e4b\u78bc\u7c3f\u5f0f\u7684\u5e73 \u5747\u503c\u8207\u8b8a\uf962\uf969\u53ef\u7531\u4e0b\uf978\u5f0f\u8a08\u7b97\u800c\u5f97\uff1a 1 , M c j j j w C \u03bc = = \u2211 \u5f0f(4.3) 2 2 2 1 , M c j j c j w C \u03c3 \u03bc = = \u2212 \u2211 \u5f0f(4.4) \u5176\u4e2d c \u03bc \u70ba\u78bc\u7c3f\u5f0f\u4e4b\u7279\u5fb5\u5e73\u5747\u503c\uff0c 2 c \u03c3 \u70ba\u78bc\u7c3f\u5f0f\u4e4b\u7279\u5fb5\u8b8a\uf962\uf969\uff0c j w \u70ba\u6bcf\u4e00\u78bc\u5b57\u6240\u5c0d\u61c9\u5230\u7684 \u6b0a\u91cd\uff0c M \u70ba\u78bc\u5b57\uf969\u76ee\u3002 \u56e0\u6b64\uff0c\u7d44\u5408\u5f0f\u5012\u983b\u8b5c\u5e73\u5747\u6d88\u53bb\u6cd5(associative CMS, A-CMS)\u4e2d\uff0c\u6240\u4f7f\u7528\u7684\u7279\u5fb5\uf96b\uf969\u4e4b\u5e73\u5747 \u503c a \u03bc \uff0c\u53ef\u7531\u4e0b\u5f0f\u8a08\u7b97\u800c\u5f97\uff1a ( ) 1 a c u \u03bc \u03b1 \u03bc \u03b1 \u03bc = \u22c5 + \u2212 \u22c5 , \u5f0f(4.5) \u5176\u4e2d u \u03bc \u8207 c \u03bc \u5206\u5225\u5982\u5f0f(4.1)\u8207\u5f0f(4.3)\u6240\u793a\uff0c\u800c \u03b1 \u70ba\u4e00\u6b0a\u91cd\u503c\uff0c 0 1 \u03b1 \u2264 \u2264 \u3002 \u56e0\u6b64\uff0cA-CMS \u8655\uf9e4\u5f8c\u7684\u65b0\u7279\u5fb5\uf96b\uf969\uff0c\u53ef\u8868\u793a\u70ba\uff1a A-CMS: , 1 . i i a X X i N \u03bc = \u2212 \u2264 \u2264 \u5f0f", | |
| "eq_num": "(" | |
| } | |
| ], | |
| "section": "\u4e00\u3001\u7dd2\uf941", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "(codebook) \uf978 \u65b9 \u7684 \u7d71 \u8a08 \u8cc7 \u8a0a \uff0c \u7136 \u5f8c \u5efa \u69cb \u51fa \u4e00 \u4ee3 \u8868 \u6b64 \u8a9e \uf906 \u7279 \u5fb5 \u7684 \u6a5f \uf961 \u5206 \u4f48 ( ) ( ) X F x P X x = \u2264 \uff0c\u4ee5\u4f5c\u70ba HEQ \u6cd5\u7b49\u5316\u7279\u5fb5\u6240\u7528\u3002\u4ee5\u4e0b\uff0c\u6211\u5011\u63cf\u8ff0 A-HEQ \u57f7\ufa08\u6b65\u9a5f\uff1a \u5047\u8a2d\u67d0\u4e00\u5f85\u6b63\u898f\u5316\u7684\u539f\u8a9e\uf906\u4e4b\u7279\u5b9a\u4e00\u7dad\u7684\u7279\u5fb5\u5e8f\uf99c\u70ba { } 1 2 , ,..., N X X X \uff0c\u5176\u4e2d N \u70ba\u6b64 \u5e8f \uf99c \u4e4b \u7279 \u5fb5 \u7e3d \uf969 \uff0c \u800c \u5176 \u5c0d \u61c9 \u5230 \u4e4b \u540c \u4e00 \u7dad \u7684 \u78bc \u5b57 \uff0c \u8868 \u793a \u70ba { } 1 2 , ,..., M C C C \uff0c \u6b0a \u91cd \u70ba { } 1 2 , ,..., M w w w \uff0c\u5176\u4e2d M \u70ba\u78bc\u5b57\uf969\u76ee\u3002\u9996\u5148\uff0c\u6211\u5011\u8a2d\u5b9a\u4e00\uf96b\uf969 \u03b2 ( 0 \u03b2 \u2264 \u2264 \u221e)\uff0c\u6b64\uf96b\uf969 \u4ee3\u8868\uf9ba\u4f7f\u7528\u78bc\u7c3f\u5f0f\u8cc7\u8a0a\u76f8\u5c0d\u65bc\u4f7f\u7528\u6574\u6bb5\u5f0f\u8cc7\u8a0a\u7684\u6bd4\uf9b5\u3002\u63a5\u8457\uff0c\u6211\u5011\u7522\u751f\u4e00\u7d44\uf969\u76ee\u70ba N \u03b2 \u7684\u65b0\u7279\u5fb5 { } k C \uff0c\u6b64\u7d44\u65b0\u7279\u5fb5\u662f\u7531\u78bc\u5b57 { } m C \u6839\u64da\u5176\u6b0a\u91cd\u503c { } m w \u6240\u5efa\uf9f7\u7684\uff0c\u65b0\u7279\u5fb5 { } k C \u4e2d \u6709 [ ] m N w \u03b2 \u00d7 \u500b\u7279\u5fb5\u7684\u503c\u548c m C \u5b8c\u5168\u76f8\u540c\uff0c( [ ] m N w \u03b2 \u00d7 \u4ee3\u8868 m N w \u03b2 \u00d7 \u53d6\u56db\u6368\u4e94\u5165\u5f8c\u7684 \u503c)\uff0c\u63db\u8a00\u4e4b\uff0c\u65b0\u7279\u5fb5 { } k C \u70ba\u4e00\u7d44\u6574\u5408\uf9ba\u6b0a\u91cd\u503c\u7684\u65b0\u78bc\u5b57\uff0c\u7576\u539f\u78bc\u5b57 m C \u5176\u6b0a\u91cd\u503c\u70ba m w \u6642\uff0c\u5b83\u5c31\u6703\u5728\u65b0\u7279\u5fb5 { } k C \u4e2d\u51fa\u73fe[ ] m N w \u03b2 \u00d7 \u6b21\uff0c\uf9b5\u5982\uff0c\u5047\u8a2d\u539f\u78bc\u5b57\u96c6\u5408\u70ba { } 3,", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "\u4e00\u3001\u7dd2\uf941", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "4\u500b \u500b \u500b (\u5be6\u969b\u4e0a\uff0c\u7531\u65bc\u56db\u6368\u4e94\u5165\u7684\u95dc\u4fc2\uff0c\u6700\u5f8c\u5f97\u5230\u7684\u65b0\u7279\u5fb5 { } k C \u5176\u7e3d\uf969\u53ef\u80fd\uf967\u6703\u6070\u597d\u662f N \u03b2 \uff0c\u5373\u6070\u597d\u70ba\u539f\u8a9e\uf906\u7279\u5fb5\uf969\u76ee N \u7684 \u03b2 \u500d) \u3002 \u63a5\u4e0b\uf92d\uff0c\u6211\u5011\u5c31\u5c07\u539f\u8a9e\uf906\u7279\u5fb5 { } 1 2 , ,..., N X X X \u8207\u4ee3\u8868\u78bc\u5b57\u7684\u65b0\u7279\u5fb5 { } 1 2 , ,..., N C C C \u03b2 \uf905\uf997\u8d77\uf92d\uff0c\u5171\u540c\u6c7a\u5b9a\u4e00\u7d44\u4ee3\u8868\u6b64\u8a9e\uf906\u7279\u5fb5\u7684\u6a5f\uf961\u5206\u4f48\uff1a ( ) ( ) ( ) ( ) 1 1 1 , 1 N N X n k n k F x u x X u x C N \u03b2 \u03b2 = = \u239b \u239e \u239f \u239c = \u2212 + \u2212 \u239f \u239c \u239f \u239f \u239c + \u239d \u23a0 \u2211 \u2211 \u5f0f(4.10)", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "\u4e00\u3001\u7dd2\uf941", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "\u6700\u5f8c\uff0c\uf9dd\u7528 HEQ \u7684\u539f\uf9e4\uff0c\u6211\u5011\u5c07\u539f\u8a9e\uf906\u7279\u5fb5\u6b63\u898f\u5316\uff0c\u5982\u4e0b\u5f0f\u6240\u793a\uff1a A-HEQ: ", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "\u4e00\u3001\u7dd2\uf941", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "( ) ( ) 1 N X x F F x \u2212 = \u5f0f(4.11) \u5176\u4e2d N F \u70ba\uf96b\u8003\u4e4b\u6a5f\uf961\u5206\u4f48\uff0c x \u70ba\u539f\u59cb\u7279\u5fb5\uf96b\uf969(\u5373\u524d\u9762\u63d0\u5230\u7684 { } 1 2 , ,..., N X X X )\uff0c x \u5373\u70ba A-HEQ \u6cd5\u6240\u5f97\u4e4b\u65b0\u7279\u5fb5\uf96b\uf969\u3002 \u7531\u5f0f(4.10)\u53ef\u770b\u51fa\uff0c\u539f\u8a9e\u97f3\u7279\u5fb5\u4e4b\u6a5f\uf961\u5206\u4f48 ( ) X F x \u7531\u6574\uf906\u7279\u5fb5 { } n X \u8207\u65b0\u78bc\u5b57\u7279\u5fb5 { } k C \u5171 \u540c\u6c7a\u5b9a\uff0c\u524d\u8005\uf969\u76ee\u70ba N \uff0c\u5f8c\u8005\uf969\u76ee\u7d04\u70ba N \u03b2 \uff0c\u56e0\u6b64\uf96b\uf969 \u03b2 \u5927\u5c0f\u6c7a\u5b9a\uf9baA-HEQ\u4e2d\uff0c\u65b0\u78bc\u5b57 \u7279\u5fb5 { } k C \u5c0d ( ) X F x \u7684\u5f71\u97ff\u7a0b\ufa01\uff0c\u7576 0 \u03b2 = \u6642\uff0c\u76f8\u7576\u65bc\u78bc\u5b57\u65b9\u9762\u7684\u8cc7\u8a0a\u5b8c\u5168\u88ab\u5ffd\uf976\uff0cA-HEQ \u5373\u8b8a\u70ba\u539f\u5148\u6240\u4ecb\u7d39\u4e4b\u6574\u6bb5\u5f0fHEQ\u6cd5(U-HEQ)\uff0c\u800c\u7576 \u03b2 \u5f88\u5927( \u03b2 \u2192 \u221e )\u6642\uff0c\u539f\u5148\u8a9e\uf906\u7684\u7279 \u5fb5 { } n X \u4e4b\u8cc7\u8a0a\u5247\u5e7e\u4e4e\u88ab\uf96d\uf976\uff0c\u5247\u6b64\u6642A-", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "\u4e00\u3001\u7dd2\uf941", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "\u5f9e\u9019\u4e09\u500b\u8868\u683c\u7684\u7d50\u679c\uff0c\u6211\u5011\u53ef\u89c0\u5bdf\u5230\u4e0b\uf99c\u5e7e\u9ede\uff1a \u25cb 1 \u5c31 CMS \u6cd5\u800c\u8a00\uff0c\u539f\u59cb\u4e4b C-CMS(C-CMS*)\u76f8\u5c0d\u65bc\u57fa\u790e\u5be6\u9a57\u7d50\u679c\u9032\u6b65\u8f03\u5c0f(\u5982\u5728 N =256 \u4e0b\uff0c\u5728 Set A \u4e0b\u63d0\u5347\uf9ba 6.00%\uff0c\u5728 Set B \u4e0b\u63d0\u5347\uf9ba 7.41%)\uff0c\u5176\u6548\u679c\u751a\u81f3\u6bd4\u6574\u6bb5\u5f0f CMS(U-CMS)\uf92d\u7684\u5dee\uff0c\u7136\u800c\uff0c\u6211\u5011\u6240\u63d0\u51fa\u7684\u65b0 C-CMS\uff0c\u5247\u5e36\uf92d\u986f\u8457\u7684\u9032\u6b65(\u5982\u5728 N =256 \u4e0b\uff0c\u5728 Set A \u4e0b\u63d0\u5347\uf9ba 9.54%\uff0c\u5728 Set B \u4e0b\u63d0\u5347\uf9ba 13.70%)\uff0c\u7531\u6b64\u8b49\u5be6\uff0c\u6211\u5011\u6240\u7528\u7684\u65b0\u7684 \u78bc\u7c3f\u5efa\u69cb\u7a0b\u5e8f\u78ba\u5be6\u80fd\u6709\u6548\u63d0\u5347 C-CMS \u7684\u6548\u679c\uff0c\u800c\u4e14\u5176\u6548\u679c\u4e26\uf967\u6703\u96a8\u8457\u78bc\u5b57\uf969\u76ee\u7684\u5927 \u5c0f\uff0c\u800c\u6709\u660e\u986f\u7684\u8b8a\u5316\u3002\u5176\u6548\u679c\u5728 Set A \u4e0b\u512a\u65bc U-CMS\uff0c\u5728 Set B \u4e0b\u5247\uf976\u905c\u65bc U-CMS\uff0c \u9019\u53ef\u80fd\u539f\u56e0\u5728\u65bc\uff0cC-CMS \u4f7f\u7528\u4e00\u6bb5\u8a9e\u97f3\u524d\u5e7e\u500b\u97f3\u6846\u4f5c\u96dc\u8a0a\u4f30\u6e2c\uff0c\u9019\u5728 Set B \u6b64\u975e\u7a69\u5b9a (non-stationary)\u96dc\u8a0a\u74b0\u5883\u4e2d\u662f\u6bd4\u8f03\uf967\u7cbe\u78ba\u7684\u3002 \u25cb 2 \u5c31 CMVN \u6cd5\u800c\u8a00\uff0c\u539f\u59cb\u4e4b C-CMVN(\u5373 C-CMVN*)\u76f8\u5c0d\u65bc\u57fa\u790e\u5be6\u9a57\u7d50\u679c\u96d6\u5df2\u6709 \uf9ba\uf967\u932f\u7684\u8fa8\uf9fc\uf961\u63d0\u5347(\u5982\u5728 M =256 \u4e0b\uff0c\u5728 Set A \u4e0b\u63d0\u5347\uf9ba 14.75%\uff0c\u5728 Set B \u4e0b\u63d0\u5347\uf9ba 18.46%)\uff0c\u4f46\u662f\u76f8\u8f03\u65bc\u6574\u6bb5\u5f0f CMVN(U-CMVN)\u800c\u8a00\uff0c\u5728 M =16 \u8207M =64 \u4e0b\uff0c\u5176\u6548\u679c\u90fd \u6bd4 U-CMVN \u9084\u8981\u5dee\uff0c\u7136\u800c\uff0c\u6211\u5011\u6240\u63d0\u51fa\u4e4b\u65b0\u7684 C-CMVN\uff0c\u5247\u6709\u660e\u986f\u7684\u9032\u6b65\uff0c\u7121\uf941\u5728 M =16\u3001 M =64 \u6216 M =256 \u4e0b\uff0c\u5176\u6548\u679c\u90fd\u6bd4\u539f\u59cb\u7684 C-CMVN \u9084\u8981\u597d\uff0c\u4e14\u5e7e\u4e4e\u90fd\u512a\u65bc U-CMVN(\u50c5\u5728 M =16 \u6642\uff0cSet B \u4e4b\u5e73\u5747\u8fa8\uf9fc\uf961\uf976\u905c\u65bc U-CMVN)\uff0c\u7531\u6b64\u8b49\u5be6\uff0c\u6211\u5011\u6240\u7528 \u7684\u65b0\u7684\u78bc\u7c3f\u5efa\u69cb\u7a0b\u5e8f\u78ba\u5be6\u80fd\u6709\u6548\u63d0\u5347 C-CMVN \u7684\u6548\u679c\uff0c\u800c\u4e14\u5176\u6548\u679c\u4e26\uf967\u6703\u96a8\u8457\u78bc\u5b57\u7684 \u5927\u5c0f\uff0c\u800c\u6709\u660e\u986f\u7684\u8b8a\u5316\u3002 \u25cb 3 \u5c31", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "\u4e00\u3001\u7dd2\uf941", | |
| "sec_num": null | |
| } | |
| ], | |
| "back_matter": [], | |
| "bib_entries": { | |
| "BIBREF0": { | |
| "ref_id": "b0", | |
| "title": "Silence Energy Normalization for Robust Speech Recognition in Additive Noise Environments", | |
| "authors": [ | |
| { | |
| "first": "Chung-Fu", | |
| "middle": [], | |
| "last": "Tai", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Jeih-Weih", | |
| "middle": [], | |
| "last": "Hung", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": null, | |
| "venue": "2006 International Conference on Spoken Language Processing", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Chung-fu Tai and Jeih-weih Hung, \"Silence Energy Normalization for Robust Speech Recognition in Additive Noise Environments\", 2006 International Conference on Spoken Language Processing (Interspeech 2006-ICSLP)", | |
| "links": null | |
| }, | |
| "BIBREF1": { | |
| "ref_id": "b1", | |
| "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": "", | |
| "issue": "", | |
| "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, 1981", | |
| "links": null | |
| }, | |
| "BIBREF2": { | |
| "ref_id": "b2", | |
| "title": "Multiband and Adaptation Approaches to Robust Speech Recognition", | |
| "authors": [ | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Tiberewala", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Hermansky", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1997, | |
| "venue": "European Conference on Speech Communication and Technology", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "S. Tiberewala and H. Hermansky, \"Multiband and Adaptation Approaches to Robust Speech Recognition\", 1997 European Conference on Speech Communication and Technology (Eurospeech 1997)", | |
| "links": null | |
| }, | |
| "BIBREF3": { | |
| "ref_id": "b3", | |
| "title": "Non-Linear Transformations of the Feature Space for Robust Speech Recognition", | |
| "authors": [ | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Torre", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Segura", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "C", | |
| "middle": [], | |
| "last": "Benitez", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [ | |
| "M" | |
| ], | |
| "last": "Peinado", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [ | |
| "J" | |
| ], | |
| "last": "Rubio", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2002, | |
| "venue": "2002 International Conference on Acoustics, Speech and Signal Processing", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "A. Torre, J. Segura, C. Benitez, A. M. Peinado, and A. J. Rubio, \"Non-Linear Transformations of the Feature Space for Robust Speech Recognition\", 2002 International Conference on Acoustics, Speech and Signal Processing (ICASSP 2002)", | |
| "links": null | |
| }, | |
| "BIBREF4": { | |
| "ref_id": "b4", | |
| "title": "Feature Statistics Compensation for Robust Speech Recognition in Additive Noise Environments", | |
| "authors": [ | |
| { | |
| "first": "Tsung-Hsueh", | |
| "middle": [], | |
| "last": "Hsieh", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2007, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Tsung-hsueh Hsieh, \"Feature Statistics Compensation for Robust Speech Recognition in Additive Noise Environments\" , M.S. thesis, National Chi Nan University, Taiwan, 2007", | |
| "links": null | |
| }, | |
| "BIBREF5": { | |
| "ref_id": "b5", | |
| "title": "Speech Feature Compensation Based on Pseudo Stereo Codebooks for Robust Speech Recognition in Additive Noise Environments", | |
| "authors": [ | |
| { | |
| "first": "Jeih-Weih", | |
| "middle": [], | |
| "last": "Tsung-Hsueh Hsieh", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Hung", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2007, | |
| "venue": "European Conference on Speech Communication and Technology", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Tsung-hsueh Hsieh and Jeih-weih Hung, \"Speech Feature Compensation Based on Pseudo Stereo Codebooks for Robust Speech Recognition in Additive Noise Environments\", 2007 European Conference on Speech Communication and Technology (Interspeech 2007-Eurospeech)", | |
| "links": null | |
| }, | |
| "BIBREF6": { | |
| "ref_id": "b6", | |
| "title": "Cepstral Statistics Compensation and Normalization Using Online Pseudo Stereo Codebooks for Robust Speech Recognition in Additive Noise Environments", | |
| "authors": [ | |
| { | |
| "first": "Jeih-Weih", | |
| "middle": [], | |
| "last": "Hung", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2008, | |
| "venue": "IEICE Transactions on Information and Systems", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Jeih-weih Hung, \"Cepstral Statistics Compensation and Normalization Using Online Pseudo Stereo Codebooks for Robust Speech Recognition in Additive Noise Environments\", IEICE Transactions on Information and Systems, 2008", | |
| "links": null | |
| }, | |
| "BIBREF7": { | |
| "ref_id": "b7", | |
| "title": "The AURORA Experimental Framework for the Performance Evaluations of Speech Recognition Systems under Noisy Conditions", | |
| "authors": [ | |
| { | |
| "first": "H", | |
| "middle": [ | |
| "G" | |
| ], | |
| "last": "Hirsch", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Pearce", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2000, | |
| "venue": "Proceedings of ISCA IIWR ASR2000", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "H. G. Hirsch and D. Pearce, \"The AURORA Experimental Framework for the Performance Evaluations of Speech Recognition Systems under Noisy Conditions,\" Proceedings of ISCA IIWR ASR2000, Paris, France, 2000", | |
| "links": null | |
| }, | |
| "BIBREF8": { | |
| "ref_id": "b8", | |
| "title": "Transmission Performance Characteristics of Pulse Code Modulation Channels", | |
| "authors": [], | |
| "year": 1996, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "ITU recommendation G.712, \"Transmission Performance Characteristics of Pulse Code Modulation Channels,\" Nov. 1996", | |
| "links": null | |
| } | |
| }, | |
| "ref_entries": { | |
| "FIGREF0": { | |
| "text": "\u7dad\u7684\u5e73\u5747\u503c\u8207\u8b8a\uf962\uf969\uff0c\u548c\u4e4b\u524d \u6587\u737b[5-7]\u4e2d\u7684\u65b9\u6cd5\u660e\u986f\u5dee\uf962\u5728\u65bc\uff0c\u6b64\u523b\u6211\u5011\u6240\u7528\u7684\u7d71\u8a08\u503c(\u5e73\u5747\u503c\u8207\u8b8a\uf962\uf969)\u662f\u4ee5\u52a0\u6b0a\u5e73 \u5747(weighted average)\u7684\u5f62\u5f0f\u6240\u6e2c\u5f97\uff0c\u800c\u975e[5-7]\u4e2d\u4e4b\u5747\u52fb\u5e73\u5747(uniform average)\u7684\u5f62\u5f0f\u3002 \u78bc\u7c3f\u5f0f\u5012\u983b\u8b5c\u5e73\u5747\u6d88\u53bb\u6cd5(codebook-based cepstral mean subtraction, C-CMS)\uff0c\u662f\u5c0d \u5012\u983b\u8b5c\u7279\u5fb5\u4e4b\u5e73\u5747\u503c\u4f5c\u6b63\u898f\u5316\u8655\uf9e4\uff0c\u5176\uf969\u5b78\u8868\u793a\u5f0f\u5982\u4e0b\uff1a , ,", | |
| "num": null, | |
| "type_str": "figure", | |
| "uris": null | |
| }, | |
| "FIGREF1": { | |
| "text": "\u3001\u5f0f(4.7)\u8207\u5f0f(4.8)\u53ef\u660e\u986f\u770b\u51fa\uff0c \u03b1 \u7684\u5927\u5c0f\u6c7a\u5b9a\uf9ba\u7d44\u5408\u5f0f\u65b9\u6cd5\u4e2d\uff0c\u4f7f\u7528\u78bc\u7c3f\u5f0f\u7d71 \u8a08\uf97e\u8207\u6574\u6bb5\u5f0f\u7d71\u8a08\uf97e\u7684\u6bd4\uf9b5\u3002\u7576 1 \u03b1 = \u6642\uff0cA-CMS\u6216A-CMVN \u5373\u70ba\u539f\u59cb\u4e4b\u78bc\u7c3f\u5f0f CMS(C-CMS)\u6216\u78bc\u7c3f\u5f0fCMVN(C-CMVN)\uff0c\u76f8\u53cd\u5730\uff0c\u7576 0 \u03b1 = \u6642\uff0cA-CMS\u6216A-CMVN\u5373 \u70ba\u539f\u59cb\u4e4b\u6574\u6bb5\u5f0fCMS(U-CMS)\u6216\u6574\u6bb5\u5f0fCMVN(U-CMVN)\u3002 (\u4e8c)\u7d44\u5408\u5f0f\u5012\u983b\u8b5c\u7d71\u8a08\u5716\u7b49\u5316\u6cd5(associative HEQ, A-HEQ) \u5728\u9019\u4e00\u7bc0\u4e2d\uff0c\u6211\u5011\u5c07\u4ecb\u7d39\u7d44\u5408\u5f0f\u7d71\u8a08\u5716\u7b49\u5316\u6cd5(associative histogram equalization, A-HEQ)\uff0c\uf9d0\u4f3c\u4e4b\u524d\u7684\u89c0\uf9a3\uff0c\u6211\u5011\u8a66\u8457\u6574\u5408\u55ae\u4e00\u8a9e\uf906(utterance)\u7279\u5fb5\u53ca\u5176\u5c0d\u61c9\u4e4b\u78bc\u5b57\u7d44\u5408", | |
| "num": null, | |
| "type_str": "figure", | |
| "uris": null | |
| }, | |
| "FIGREF2": { | |
| "text": "3, 3, 5, 5, 5,..., 5, 7, 7, 7,..., 7}", | |
| "num": null, | |
| "type_str": "figure", | |
| "uris": null | |
| }, | |
| "TABREF6": { | |
| "html": null, | |
| "num": null, | |
| "type_str": "table", | |
| "text": "HEQ \u6cd5\u800c\u8a00\uff0cC-HEQ \u540c\u6a23\u4e5f\u80fd\u6709\u6548\u63d0\u6607\u8fa8\uf9fc\uf961\uff0c\u4f46\u7121\uf941\u5728 A \u7d44\u96dc\u8a0a\u74b0\u5883\u4e0b\u6216 \u4e0b\uff0c\u7522\u751f\u6700\u4f73\u8fa8\uf9fc\uf961\u7684 \u03b1 \u503c(\u5982\u5f0f(4.5)\u3001(4-7)\u8207 (4.8)\u4e2d\u6240\u793a)\u6216 \u03b2 \u503c(\u5982\u5f0f(4.10)\u4e2d\u6240\u793a)\uf967\u76e1\u76f8\u540c\uff0c\u56e0\u6b64\u5728\u4ee5\u4e0b\u7684\u5be6\u9a57\u8fa8\uf9fc\u7d50\u679c\u4e2d\uff0c\u6211\u5011 \u53ea\u5448\u73fe\u5728\uf967\u540c\u7684 N \u503c\u6642\uff0c\u6240\u7522\u751f\u6700\u4f73\u5e73\u5747\u8fa8\uf9fc\uf961\u4e4b \u03b1 \u503c\u6216 \u03b2 \u503c\u4e4b\u7d50\u679c\u3002\u9996\u5148\uff0c\u8868\u4e94\u70ba A-CMS \u5728\u78bc\u5b57\uf969\u76ee N \u5206\u5225\u70ba 16\u300164 \u8207 256 \u4e0b\uff0c\u6240\u5f97\u5230\u7684\u6700\u4f73\u8fa8\uf9fc \u7d50\u679c\uff0c\u70ba\uf9ba\u6bd4\u8f03\u8d77\ufa0a\uff0c\u6211\u5011\u4e5f\u5c07\u8868\u4e8c\u4e2d\u7684\u57fa\u672c\u5be6\u9a57\u3001C-CMS( M =256)\u8207 U-CMS \u7684\u5e73\u5747 \u8fa8\uf9fc\uf961\uf99c\u5728\u8868\u4e2d\u3002\u5f9e\u6b64\u8868\u4e2d\uff0c\u6211\u5011\u53ef\u4ee5\u89c0\u5bdf\u5230\u4ee5\u4e0b\u5e7e\u7a2e\u60c5\u5f62\uff1a \u25cb 1 \u7d44\u5408\u5f0f\u5012\u983b\u8b5c\u5e73\u5747\u6d88\u53bb\u6cd5(A-CMS)\u76f8\u8f03\u65bc\u57fa\u672c\u5be6\u9a57\u800c\u8a00\uff0c\u7121\uf941\u5728\u78bc\u5b57\uf969M =16\u3001 64 \u8207 256 \u4e0b\uff0c\u5176\u5e73\u5747\u8fa8\uf9fc\uf961\u7686\u6709\u5927\u5e45\u7684\u9032\u6b65\uff0c\u4e09\u8005\u5728 A \u7d44\u96dc\u8a0a\u74b0\u5883\u4e0b\u5206\u5225\u6709 11.86%\u3001 11.30%\u8207 10.98%\u7684\u8fa8\uf9fc\uf961\u63d0\u5347\uff0c\u5728 B \u7d44\u96dc\u8a0a\u74b0\u5883\u4e0b\u5206\u5225\u6709 17.76%\u300116.82%\u8207 16.83% \u7684\u8fa8\uf9fc\uf961\u63d0\u5347\uff0c\u7531\u6b64\u53ef\u770b\u51fa A-CMS \u5177\u6709\uf967\u932f\u4e4b\u7279\u5fb5\u5f37\u5065\u5316\u6548\u679c\u3002 \u25cb 2 A-CMS \u5728\u5404\u7a2e\uf967\u540c\u7684\u78bc\u5b57\uf969 N \u4e4b\u4e0b\uff0c\u5176\u5e73\u5747\u8fa8\uf9fc\uf961\u7686\u6bd4 C-CMS \u8207 U-CMS \uf92d \u5f97\u597d\uff0c\u5176\u4e2d\u5728 N =16 \u6642\u80fd\u6709\u6700\u4f73\u7684\u6548\u679c\uff0c\u5728 A \u7d44\u96dc\u8a0a\u74b0\u5883\u8207 B \u7d44\u96dc\u8a0a\u74b0\u5883\u4e0b\u4e4b\u5e73\u5747\u8fa8 \uf9fc\uf961\u5206\u5225\u70ba 83.78%\u8207 85.55%\uff0c\u76f8\u8f03\u65bc C-CMS \u53d6 M =256 \u6240\u5f97\u4e4b\u6700\u4f73\u8fa8\uf9fc\uf961\uff0cA-CMS \u5728 A \u7d44\u96dc\u8a0a\u74b0\u5883\u8207 B \u7d44\u96dc\u8a0a\u74b0\u5883\u4e0b\u5206\u5225\u9032\u6b65\uf9ba 2.32%\u548c 4.06%\uff0c\u9019\u4e9b\u9032\u6b65\u90fd\u986f\u793a\uf9ba A-CMS \u512a\u65bc C-CMS\u3002\u6700\u5f8c\u76f8\u8f03\u65bc U-CMS\uff0cA-CMS \u5728 A \u7d44\u96dc\u8a0a\u74b0\u5883\u8207 B \u7d44\u96dc\u8a0a\u74b0\u5883\u4e0b \u5176\u8fa8\uf9fc\uf961\u5206\u5225\u53ef\u4ee5\u63d0\u5347 4.41%\u548c 3.08%\u3002\u56e0\u6b64\u7531\u5be6\u9a57\uf969\u64da\u4e2d\u53ef\u4ee5\u8b49\u660e\uff0c\u76f8\u5c0d\u65bc C-CMS \u8207 U-CMS \u800c\u8a00\uff0cA-CMS \u90fd\u53ef\u4ee5\u5f97\u5230\u8f03\u597d\u7684\u8fa8\uf9fc\u7d50\u679c\uff0c\u9019\u53ef\u80fd\u662f\u56e0\u70ba A-CMS \u540c\u6642\u6574\u5408 \uf9ba C-CMS \u8207 U-CMS \u6240\u7528\u7684\u7d71\u8a08\u8cc7\u8a0a\uff0c\u6240\u4ee5\u5b83\uf901\u80fd\u6709\u6548\u6539\u5584\u8a9e\u97f3\u5728\u96dc\u8a0a\u4e0b\u7684\u5f37\u5065\u6027\u3002 \u8868\u4e94\u3001U-CMS\u3001\u65b0C-CMS\u8207A-CMS\u7684\u5e73\u5747\u8fa8\uf9fc\uf961", | |
| "content": "<table><tr><td>B \u7d44\u96dc\u8a0a\u74b0\u5883\u4e0b\uff0c\u5176\u5e73\u5747\u8fa8\uf9fc\uf961\u90fd\u6bd4 U-HEQ \uf92d\u5f97\u5dee\uff0c\u6211\u5011\u63a8\u6e2c\u5176\u539f\u56e0\u53ef\u80fd\u5728\u65bc\uff0cC-HEQ \u5728\u7d14\u96dc\u8a0a\u7684\u4f30\u6e2c\u4e0a\uff0c\u662f\u4ee5\u6bcf\u4e00\u6bb5\u6e2c\u8a66\u8a9e\u97f3\u7684\u524d\u5e7e\u500b\u97f3\u6846\u4f5c\u70ba\u7d14\u96dc\u8a0a\u97f3\u6846\u7684\u4ee3\u8868\uff0c\u56e0\u800c\u9020 \u6210\u7d14\u96dc\u8a0a\u8cc7\u8a0a\uf967\u8db3\uff0c\u5c0e\u81f4\u6240\u5f97\u7684\u96dc\u8a0a\u8a9e\u97f3\u78bc\u7c3f\uf967\u5920\u7cbe\u6e96\uff0c\u6700\u7d42\u9020\u6210 C-HEQ \u8fa8\uf9fc\uf961\u6bd4 U-HEQ \u9084\u8981\u5dee\u7684\u7d50\u679c\u3002 \u8868\u4e8c\u3001U-CMS\u3001\u539f\u59cbC-CMS(C-CMS*)\u3001\u8207\u65b0C-CMS\u7684\u5e73\u5747\u8fa8\uf9fc\uf961(%) Method Set A Set B average AR RR Baseline 71.92 67.79 69.86 \u23af \u23af U-CMS 79.37 82.47 80.92 11.07 36.71 C-CMS*(M=16) 74.21 70.81 72.51 2.65 8.81 C-CMS*(M =64) 74.03 70.74 72.39 2.53 8.39 C-CMS*(M =256) 77.92 75.20 76.56 6.71 22.24 C-CMS(M =16) 79.04 79.56 79.30 9.45 31.33 C-CMS(M =64) 80.79 80.19 80.49 10.64 35.28 C-CMS(M=256) 81.46 81.49 81.48 11.62 38.55 \u8868\u4e09\u3001U-CMVN\u3001\u539f\u59cbC-CMVN(C-CMVN*)\u3001\u8207\u65b0C-CMVN\u7684\u5e73\u5747\u8fa8\uf9fc\uf961 Method Set A Set B average AR RR Baseline 71.92 67.79 69.86 \u23af \u23af U-CMVN 85.03 85.56 85.30 15.44 51.22 C-CMVN*(M =16) 84.44 82.40 83.42 13.57 45.00 C-CMVN*(M=64) 84.13 81.53 82.83 12.98 43.04 C-CMVN*(M=256) 86.67 86.25 86.46 16.61 55.08 C-CMVN(M=16) 85.41 85.21 85.31 15.46 51.27 C-CMVN(M=64) 86.92 86.81 86.87 17.01 56.43 C-CMVN(M=256) 87.10 87.32 87.21 17.36 57.57 \u63a5\u8457\uff0c\u8868\uf9d1\u70baA-CMVN\u5728\u78bc\u5b57\uf969\u76eeM \u5206\u5225\u70ba16\u300164\u8207256\u4e0b\uff0c\u6240\u5f97\u5230\u7684\u6700\u4f73\u8fa8\uf9fc\u7d50 \u8868\u4e03\u3001U-HEQ\u3001\u65b0C-HEQ\u8207A-HEQ\u7684\u5e73\u5747\u8fa8\uf9fc\uf961 \u8868\u56db\u3001U-HEQ\u8207\u65b0C-HEQ\u7684\u5e73\u5747\u8fa8\uf9fc\uf961 Set A Set B average AR RR Baseline 71.92 67.79 69.86 \u23af U-CMS 79.37 82.47 80.92 11.07 C-CMS(M=256) 81.46 81.49 81.48 11.62 A-CMS(M=16, \u03b1 =0.5) 83.78 85.55 84.67 14.81 A-CMS(M=64, \u03b1 =0.6) 83.22 84.61 83.92 14.06 82.90 84.62 83.76 13.91 46.13 U-HEQ\uf901\u80fd\u63d0\u5347\u96dc\u8a0a\u74b0\u5883\u4e0b\u8a9e\u97f3\u8fa8\uf9fc\u7684\u7cbe\u78ba\ufa01\u3002 A-CMS(M=256, \u03b1 =0.6) \u6211\u5011\u518d\u6b21\u9a57\u8b49\uf9ba\u7d44\u5408\u5f0f\u7684\u65b9\u6cd5\u512a\u65bc\u78bc\u7c3f\u5f0f\u8207\u6574\u6bb5\u5f0f\u7684\u65b9\u6cd5\uff0c\u5373A-HEQ\u6bd4C-HEQ\u8207 46.64 \u63d0\u5347\uf9ba3.07%\u82072.54%\uff0c\u5176\u76f8\u5c0d\u6539\u5584\uf961\u5206\u5225\u70ba23.62%\u820721.76%\u3002\uf9d0\u4f3c\u4e4b\u524d\u7684\u7d50\u679c\uff0c\u9019\uf9e8 49.13 \u65bcC-HEQ\uff1b\u800c\u8ddfU-HEQ\u6bd4\u8f03\u6642\uff0cA-HEQ\u5728A\u7d44\u96dc\u8a0a\u74b0\u5883\u8207B\u7d44\u96dc\u8a0a\u74b0\u5883\u4e0b\u5176\u8fa8\uf9fc\uf961\u5206\u5225 38.55 \u5883\u8207B\u7d44\u96dc\u8a0a\u74b0\u5883\u4e0b\u5176\u8fa8\uf9fc\uf961\u5247\u5206\u5225\u9032\u6b65\uf9ba3.85%\u82074.80%\uff0c\u9019\u4e9b\u9032\u6b65\u90fd\u986f\u793a\uf9baA-HEQ\u512a 36.71 \u70ba90.07%\u548c90.87%\uff0c\u76f8\u8f03\u65bcC-HEQ\u53d6 M =256\u6240\u5f97\u4e4b\u6700\u4f73\u8fa8\uf9fc\uf961\uff0cA-HEQ\u5728A\u7d44\u96dc\u8a0a\u74b0 \u23af A-CMS\uff0cA-CMVN\u540c\u6642\u6574\u5408\uf9baC-CMVN\u8207U-CMVN\u6240\u7528\u7684\u7d71\u8a08\u8cc7\u8a0a\uff0c\u56e0\u6b64\u6211\u5011\u9810\u671f\u5b83 \u5177\u5099\uf9ba\uf901\u4f73\u7684\u8a9e\u97f3\u7279\u5fb5\u5f37\u5065\u5316\u7684\u6548\u679c\uff0c\u5be6\u9a57\uf969\u64da\u4e5f\u78ba\u5be6\u9a57\u8b49\uf9baA-CMVN\u7684\u8868\u73fe\u660e\u986f\u512a\u65bc C-CMVN\u8207U-CMVN\u3002 \u8868\uf9d1\u3001U-CMVN\u3001\u65b0C-CMVN\u8207A-CMVN\u7684\u5e73\u5747\u8fa8\uf9fc\uf961 Method Set A Set B Average AR RR Baseline 71.92 67.79 69.86 \u23af \u23af U-CMVN 85.03 85.56 85.30 15.44 51.22 C-CMVN(M=256) 87.10 87.32 87.21 17.36 57.57 A-CMVN(M =16, \u03b1 =0.7) 88.11 88.97 88.54 18.69 61.98 A-CMVN(M=64, \u03b1 =0.8) 88.00 88.56 88.28 18.43 61.12 A-CMVN(M=256, \u03b1 =0.8) 87.35 88.05 87.70 17.85 59.20 \u6700\u5f8c\uff0c\u8868\u4e03\u70baA-HEQ\u5728\u78bc\u5b57\uf969\u76ee M \u5206\u5225\u70ba16\u300164\u8207256\u4e0b\uff0c\u6240\u5f97\u5230\u7684\u6700\u4f73\u8fa8\uf9fc\u7d50\u679c\uff0c \u70ba\uf9ba\u6bd4\u8f03\u8d77\ufa0a\uff0c\u6211\u5011\u4e5f\u5c07\u8868\u56db\u4e2d\u7684\u57fa\u672c\u5be6\u9a57\u3001C-HEQ(M=256)\u8207U-HEQ\u7684\u5e73\u5747\u8fa8\uf9fc\uf961\uf99c \u5728\u8868\u4e2d\u3002\u5f9e\u6b64\u8868\u4e2d\uff0c\u6211\u5011\u53ef\u4ee5\u89c0\u5bdf\u5230\u4ee5\u4e0b\u5e7e\u7a2e\u60c5\u5f62\uff1a \u25cb 1 \u5c0d\u65bc\u7d44\u5408\u5f0f\u7d71\u8a08\u5716\u7b49\u5316\u6cd5(A-HEQ)\u800c\u8a00\uff0c\u7121\uf941\u5728\u78bc\u5b57\uf969 M =16\u300164\u8207256\u4e0b\uff0c\u5176 \u5e73\u5747\u8fa8\uf9fc\uf961\u76f8\u8f03\u65bc\u57fa\u672c\u5be6\u9a57\u800c\u8a00\uff0c\u90fd\u6709\u5927\u5e45\u7684\u6539\u9032\uff0c\u4e09\u8005\u5728A\u7d44\u96dc\u8a0a\u74b0\u5883\u4e0b\u5206\u5225\u6709 18.15%\u300117.28%\u820715.76%\u7684\u8fa8\uf9fc\uf961\u63d0\u5347\uff0c\u5728B\u7d44\u96dc\u8a0a\u74b0\u5883\u4e0b\u5206\u5225\u670923.08%\u300122.36%\u8207 \u25cb 2 A-HEQ\u5728\u5404\u7a2e\u78bc\u5b57\uf969M\u7684\u60c5\u5f62\u4e0b\uff0c\u5176\u5e73\u5747\u8fa8\uf9fc\uf961\u7686\u6bd4C-HEQ\u8207U-HEQ\uf92d\u5f97\u597d\uff0c \u5176\u4e2d\u4ee5M=16\u6240\u5f97\u7684\u5e73\u5747\u8fa8\uf9fc\uf961\u70ba\u6700\u4f73\uff0c\u5728A\u7d44\u96dc\u8a0a\u74b0\u5883\u8207B\u7d44\u96dc\u8a0a\u74b0\u5883\u4e0b\u4e4b\u8fa8\uf9fc\uf961\u5206\u5225 \u8fd1\uf98e\uf92d\uff0c\u672c\u5be6\u9a57\u5ba4\u767c\u5c55\uf9ba\u78bc\u7c3f\u5f0f(codebook-based)\u7279\u5fb5\uf96b\uf969\u6b63\u898f\u5316\u6280\u8853\uff0c\u5206\u5225\u70baC-CMS \u8207C-CMVN\u3002\u9867\u540d\u601d\u7fa9\uff0c\u5728\u9019\u4e9b\u65b9\u6cd5\u4e2d\uff0c\u6240\u4f7f\u7528\u7684\u7279\u5fb5\u7d71\u8a08\uf97e\u662f\u7531\u78bc\u7c3f\u8a08\u7b97\u800c\u5f97\uff0c\u5be6\u9a57 \u8b49\u5be6\u9019\u4e9b\u78bc\u7c3f\u5f0f\u7279\u5fb5\uf96b\uf969\u6b63\u898f\u5316\u6280\u8853\u5176\u8868\u73fe\u5927\u81f4\u4e0a\u7686\u512a\u65bc\u6574\u6bb5\u5f0f\u7279\u5fb5\uf96b\uf969\u6b63\u898f\u5316\u6280 \u8853\u3002\u7136\u800c\u6211\u5011\u767c\u73fe\uff0c\u5b83\u5011\u4ecd\u7136\u6709\u9032\u4e00\u6b65\u7684\u6539\u5584\u7a7a\u9593\u3002\u56e0\u6b64\uff0c\u672c\uf941\u6587\u4e2d\u6211\u5011\u63d0\u51fa\uf9ba\u4e00\u5957\u6539 \uf97c\u5f0f\u7684\u78bc\u7c3f\u5efa\uf9f7\u7a0b\u5e8f\uff0c\u76f8\u5c0d\u65bc\u539f\u7a0b\u5e8f\u7684\uf967\u540c\u4e4b\u8655\uff0c\u5728\u65bc\u6211\u5011\u61c9\u7528\uf9ba\u8a9e\u97f3\u5075\u6e2c\u6280\u8853\u8655\uf9e4\u4e7e \u6de8\u8a9e\u97f3\u8a0a\u865f\uff0c\u7136\u5f8c\uf9dd\u7528\u7d14\u8a9e\u97f3\u5340\u6bb5\u7684\u8a9e\u97f3\u7279\u5fb5\uf92d\u8a13\uf996\u78bc\u5b57\uff1b\u6b64\u5916\uff0c\u9019\u4e9b\u78bc\u5b57\u6839\u64da\u5176\u6db5\u84cb \u7684\u7279\u5fb5\uf969\u76ee\u8ce6\u4e88\uf967\u540c\u7684\u6b0a\u91cd(weight)\uff0c\u6b64\u6539\uf97c\u6cd5\u5728\u7b2c\u4e09\u7ae0\u6709\u8a73\u7d30\u7684\uf96f\u660e\u3002 \u9664\uf9ba\u63d0\u51fa\u4e0a\u8ff0\u6539\uf97c\u5f0f\u7684\u78bc\u7c3f\u5efa\uf9f7\u7a0b\u5e8f\u4e4b\u5916\uff0c\u672c\uf941\u6587\u53e6\u4e00\u91cd\u9ede\u5728\u65bc\uff0c\u6211\u5011\u63d0\u51fa\uf9ba\u4e00\u7cfb \uf99c\u7d44\u5408\u5f0f(associative)\u7279\u5fb5\uf96b\uf969\u6b63\u898f\u5316\u6280\u8853\uff0c\u5206\u5225\u70baA-CMS\u3001A-CMVN\u8207A-HEQ\uff0c\u9019\u4e9b \u6280\u8853\u4e2d\uff0c\u6211\u5011\u6574\u5408\uf9ba\u6574\u6bb5\u5f0f\u6280\u8853\u8207\u78bc\u7c3f\u5f0f\u6280\u8853\u6240\u7528\u7684\u7279\u5fb5\u7d71\u8a08\u8cc7\u8a0a\uff0c\u7528\u6b64\u6574\u5408\u5f8c\u4e4b\u7d71\u8a08 \uf97e\uf92d\u57f7\ufa08CMS\uff0cCMVN\u6216HEQ\uff0c\u5176\u8a73\u8ff0\u65bc\u7b2c\u56db\u7ae0\u4e2d\uff0c\u9019\u6a23\u7684\u6280\u8853\u53ef\u4ee5\u6709\u6548\u5730\u88dc\u511f\u78bc\u7c3f \u5f0f\u6280\u8853\u4e2d\uff0c\u7d14\u96dc\u8a0a\u8cc7\u8a0a\uf967\u8db3\u7684\u7f3a\u9ede\uff0c\u7b2c\u4e94\u7ae0\u4e2d\u7684\u5be6\u9a57\u7d50\u679c\u8b49\u5be6\uff0c\u7d44\u5408\u5f0f\u7684\u7279\u5fb5\uf96b\uf969\u6b63\u898f \u5316\u6280\u8853\u6bd4\u6574\u6bb5\u5f0f\u8207\u78bc\u7c3f\u5f0f\u7279\u5fb5\uf96b\uf969\u6b63\u898f\u5316\u6280\u8853\uff0c\u5747\u80fd\uf901\u660e\u986f\u5730\u63d0\u5347\u8fa8\uf9fc\u7cbe\u78ba\ufa01\u3002 \u96d6\u7136\u7d44\u5408\u5f0f\u7279\u5fb5\uf96b\uf969\u6b63\u898f\u5316\u6280\u8853\u6548\u679c\u5341\u5206\u986f\u8457\uff0c\u4f46\u5176\u6700\u4f73\u8868\u73fe\u6709\u8cf4\u65bc\u67d0\u4e9b\u81ea\u7531\uf96b\uf969 (\u5373\u5f0f(4.5)\u3001\u5f0f(4.8)\u4e2d\u7684 \u03b1 \u53ca\u5f0f(4.10)\u4e2d\u7684 \u03b2 )\u7684\u624b\u52d5\u8abf\u6574\uf92d\u6574\u5408\u78bc\u7c3f\u5f0f\u8207\u6574\u6bb5\u5f0f\u4e4b\u7d71\u8a08\u8cc7 \u8a0a\uff0c\u56e0\u6b64\u5728\u672a\uf92d\u7684\u767c\u5c55\u4e0a\uff0c\u6211\u5011\u5e0c\u671b\u80fd\u81ea\u52d5\u5730\u6c42\u53d6\u51fa\u6700\u4f73\u7684 \u03b1 \u8207 \u03b2 \u7b49\uf96b\uf969\u503c\uff0c\uf92d\u5c0d\uf978\u65b9 \u7684\u7d71\u8a08\u8cc7\u8a0a\u4f5c\uf901\u7cbe\u78ba\u7684\u6574\u5408\uff0c\u540c\u6642\uff0c\u5728\u5efa\u69cb\u96dc\u8a0a\u8a9e\u97f3\u78bc\u7c3f\u7684\u7a0b\u5e8f\u4e0a\uff0c\u6211\u5011\u4e5f\u5e0c\u671b\u80fd\uf96b\u8003 \u8a31\u591a\u96dc\u8a0a\u4f30\u6e2c\u7684\u65b9\u6cd5\uff0c\uf901\u7cbe\u78ba\u6e2c\u5f97\u4e00\u6bb5\u8a9e\u97f3\u4e2d\u7d14\u96dc\u8a0a\u7684\u7d71\u8a08\u7279\u6027\uff0c\u671f\u5f85\uf901\u6709\u6548\u5730\u63d0\u5347\u78bc \u7c3f\u5f0f\u7279\u5fb5\uf96b\uf969\u6b63\u898f\u5316\u6280\u8853\u7684\u6548\u80fd\u3002 \u898f\u5316\u6280\u8853\u4e2d\uff0c\u7531\u65bc\u5728\uf967\u540c\u7684\u78bc\u5b57\uf969\u76ee N Method 21.10%\u7684\u8fa8\uf9fc\uf961\u63d0\u5347\uff0c\u986f\u793a\uf9baA-HEQ\u5728\u8a9e\u97f3\u7279\u5fb5\u5f37\u5065\u6027\u7684\u6548\u80fd\uff0c\u4e14\u76f8\u8f03\u65bc\u4e4b\u524d\u6240\u8ff0\u7684 \uf978\u7a2e\u7d44\u5408\u5f0f\u7279\u5fb5\u6b63\u898f\u5316\u6cd5A-CMS\u8207A-CMVN\uff0cA-HEQ\u7684\u8868\u73fe\uf901\u70ba\u512a\uf962\u3002 \uf96b\u8003\u6587\u737b</td></tr><tr><td>Method Baseline \u679c\uff0c\u5728\u8868\u4e2d\uff0c\u6211\u5011\u4e5f\uf99c\u51fa\u539f\u8868\u4e09\u4e2d\u7684\u57fa\u672c\u5be6\u9a57\u3001C-CMVN( M =256)\u8207U-CMVN\u7684\u5e73\u5747 Set A Set B average AR RR 71.92 67.79 69.86 \u23af Method Set A Set B Average AR RR \u23af \u8fa8\uf9fc\uf961\u4ee5\u4f9b\u6bd4\u8f03\u3002\u5f9e\u6b64\u8868\u4e2d\uff0c\u6211\u5011\u53ef\u4ee5\u89c0\u5bdf\u5230\u4ee5\u4e0b\u5e7e\u7a2e\u60c5\u5f62\uff1a Baseline 71.92 67.79 69.86 \u23af \u23af</td></tr><tr><td>U-HEQ C-HEQ(M=16) \u25cb 1 \u7d44\u5408\u5f0f\u5012\u983b\u8b5c\u5e73\u5747\u503c\u8207\u8b8a\uf962\uf969\u6b63\u898f\u5316\u6cd5(A-CMVN)\u5728\u78bc\u5b57\uf969\u76ee M =16\u300164 \u8207 256 87.00 88.33 87.67 17.81 59.08 84.03 84.46 84.25 14.39 U-HEQ 87.00 88.33 87.67 17.81 59.08 47.74 C-HEQ(M=64) 86.32 85.90 86.11 16.26 \u4e0b\uff0c\u76f8\u8f03\u65bc\u57fa\u672c\u5be6\u9a57\u800c\u8a00\uff0c\u5176\u5e73\u5747\u8fa8\uf9fc\uf961\u7686\u6709\u5927\u5e45\u7684\u6539\u9032\uff0c\u9019\u4e09\u7a2e A-CMVN \u5728 A \u7d44\u96dc C-HEQ(M=256) 86.22 86.07 86.15 16.29 54.04 53.92 C-HEQ(M=256) 86.22 86.07 86.15 16.29 \u8a0a\u74b0\u5883\u4e0b\u5206\u5225\u6709 16.19%\u300116.08%\u8207 15.43%\u7684\u8fa8\uf9fc\uf961\u63d0\u5347\uff0c\u5728 B \u7d44\u96dc\u8a0a\u74b0\u5883\u4e0b\u5206\u5225\u6709 A-HEQ(M=16, \u03b2 =0.9) 90.07 90.87 90.47 20.62 68.39 54.04 21.18%\u300120.77%\u8207 20.26%\u7684\u8fa8\uf9fc\uf961\u63d0\u5347\uff0c\u7531\u6b64\u53ef\u4ee5\u767c\u73fe A-CMVN \u78ba\u5be6\u80fd\ufa09\u4f4e\u52a0\u6210\u6027\u96dc A-HEQ(M=64, \u03b2 =0.9) 89.20 90.15 89.68 19.82 65.75</td></tr><tr><td>\u8a0a\u5c0d\u8a9e\u97f3\u7279\u5fb5\u7684\u5e72\u64fe\uff0c\u800c\u63d0\u5347\u8fa8\uf9fc\u7cbe\u78ba\ufa01\u3002 A-HEQ(M=256, \u03b2 =1) 87.68 88.89 \u25cb 2 A-CMVN\u5728\u5404\u7a2e\u78bc\u5b57\uf969N\u7684\u60c5\u5f62\u4e0b\uff0c\u5176\u5e73\u5747\u8fa8\uf9fc\uf961\u7686\u6bd4C-CMVN\u3001U-CMVN\uf92d 88.29 18.43 61.14 2.\u7d44\u5408\u5f0f\u7279\u5fb5\uf96b\uf969\u6b63\u898f\u5316\u6cd5\u4e4b\u8fa8\uf9fc\u7d50\u679c \u5f97\u597d\uff0c\u5176\u4e2d\u4ee5 N =16\u6642\u8868\u73fe\u70ba\u6700\u4f73\uff0c\u5728A\u7d44\u96dc\u8a0a\u74b0\u5883\u8207B\u7d44\u96dc\u8a0a\u74b0\u5883\u4e0b\u4e4b\u5e73\u5747\u8fa8\uf9fc\uf961\u5206 \uf9d1\u3001\u7d50\uf941\u8207\u672a\uf92d\u5c55\u671b \u5728\u9019\u4e00\u7bc0\u4e2d\uff0c\u6211\u5011\u5c07\u4ecb\u7d39\u672c\uf941\u6587\u6240\u63d0\u51fa\u4e4b\u7d44\u5408\u5f0f(associative)\u7279\u5fb5\uf96b\uf969\u6b63\u898f\u5316\u6280\u8853\u4e4b \u5225\u70ba88.11%\u548c88.97%\uff0c\u76f8\u8f03\u65bcC-CMVN\u53d6 M =256\u6240\u5f97\u4e4b\u6700\u4f73\u8fa8\uf9fc\uf961\uff0cA-CMVN\u5728A\u7d44 \u5728\u672c\uf941\u6587\u4e2d\uff0c\u6211\u5011\u4e3b\u8981\u8a0e\uf941\u7684\u7279\u5fb5\uf96b\uf969\u6b63\u898f\u5316\u6280\u8853\uff0c\u5206\u5225\u70ba\u5012\u983b\u8b5c\u5e73\u5747\u6d88\u53bb\u6cd5 \u8fa8\uf9fc\u7d50\u679c\uff0c\u9019\u4e09\u7a2e\u6280\u8853\u5206\u5225\u70ba\u7d44\u5408\u5f0f\u5012\u983b\u8b5c\u5e73\u5747\u6d88\u53bb\u6cd5(associative CMS, A-CMS)\u3001\u7d44\u5408 \u96dc\u8a0a\u74b0\u5883\u8207B\u7d44\u96dc\u8a0a\u74b0\u5883\u5247\u5206\u5225\u9032\u6b65\uf9ba1.01%\u82071.65%\uff0c\u9019\u4e9b\u9032\u6b65\u90fd\u986f\u793a\uf9baA-CMVN\u512a\u65bc (CMS)\u3001\u5012\u983b\u8b5c\u5e73\u5747\u503c\u8207\u8b8a\uf962\uf969\u6b63\u898f\u5316\u6cd5(CMVN)\u8207\u5012\u983b\u8b5c\u7d71\u8a08\u5716\u7b49\u5316\u6cd5(HEQ)\uff0c\u9019\u4e09 \u5f0f\u5012\u983b\u8b5c\u5e73\u5747\u503c\u8207\u8b8a\uf962\uf969\u6b63\u898f\u5316\u6cd5(associative CMVN, A-CMVN)\u8207\u7d44\u5408\u5f0f\u7d71\u8a08\u5716\u7b49\u5316 C-CMVN\uff1b\u800c\u8ddfU-CMVN\u6bd4\u8f03\u6642\uff0cA-CMVN\u5728A\u7d44\u96dc\u8a0a\u74b0\u5883\u8207B\u7d44\u96dc\u8a0a\u74b0\u5883\u4e0b\uff0c\u5176\u8fa8\uf9fc \u7a2e\u6280\u8853\u7686\u9808\u4f7f\u7528\u5230\u7279\u5fb5\u7684\u7d71\u8a08\uf97e\u3002\u50b3\u7d71\u4e0a\uff0c\u9019\u4e9b\u7d71\u8a08\uf97e\u662f\u7d93\u7531\u4e00\u6574\u6bb5\u7684\u8a9e\u97f3\u7279\u5fb5\u4f30\u6e2c\u800c \u6cd5(associative histogram equalization, A-HEQ)\u3002\u5728 A-CMS\u3001A-CMVN \u8207 A-HEQ \u4e09\u7a2e\u6b63 \uf961\u5206\u5225\u53ef\u4ee5\u63d0\u53473.08%\u548c3.41%\uff0c\u5176\u76f8\u5c0d\u6539\u5584\uf961\u5206\u5225\u70ba20.55%\u820723.62%\u3002\uf9d0\u4f3c\u4e4b\u524d\u7684 \u5f97\u3002\u56e0\u6b64\uff0c\u5176\u5c0d\u61c9\u7684\u6280\u8853\uff0c\u6211\u5011\u7d71\u7a31\u70ba\u6574\u6bb5\u5f0f(utterance-based)\u7279\u5fb5\uf96b\uf969\u6b63\u898f\u5316\u6280\u8853\u3002\u5728</td></tr></table>" | |
| } | |
| } | |
| } | |
| } |