| { | |
| "paper_id": "O09-1020", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T08:11:01.041706Z" | |
| }, | |
| "title": "", | |
| "authors": [], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "Cepstral statistics normalization techniques have been shown to be very successful at improving the noise robustness of speech features. In this paper, we propose a hybrid-based scheme to achieve a more accurate estimate of the statistical information of features in these techniques. By properly integrating codebook and utterance/segment knowledge, the", | |
| "pdf_parse": { | |
| "paper_id": "O09-1020", | |
| "_pdf_hash": "", | |
| "abstract": [ | |
| { | |
| "text": "Cepstral statistics normalization techniques have been shown to be very successful at improving the noise robustness of speech features. In this paper, we propose a hybrid-based scheme to achieve a more accurate estimate of the statistical information of features in these techniques. By properly integrating codebook and utterance/segment knowledge, the", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Abstract", | |
| "sec_num": null | |
| } | |
| ], | |
| "body_text": [ | |
| { | |
| "text": "resulting hybrid-based normalization methods significantly outperform conventional utterance-based, segment-based and codebook-based ones in recognition accuracy.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "For the Aurora-2 clean-condition training task, the proposed hybrid codebook/segment-based histogram equalization (CS-HEQ) achieves an average recognition accuracy of 90.66%, which is better than utterance-based HEQ (87.62%), segment-based HEQ (85.92%) and codebook-based HEQ (85.29%). Furthermore, the high-performance CS-HEQ can be implemented with a short delay and can thus be applied in real-time online systems. A similar performance promotion can be also found in the methods of hybrid-based cepstral mean subtraction (CMS), cepstral mean and variance normalization (CMVN), cepstral gain normalization (CGN) and higher-order cepstral moment normalization (HOCMN). [ ] x m \u6240\u5f97\u7684\u5404 \u7a2e\u7d71\u8a08\u503c\uff0c\u4e8b\u5be6\u4e0a\u8ddf [ ] x m \u65bc\u5e8f\uf99c\u4e4b\u9806\u5e8fm \u7121\u95dc\uff0c\u610f\u5373\u5728\u6574\uf906\u5f0f\u4f30\u6e2c\u6cd5\u800c\u8a00\uff0c\u6211\u5011\u53ea\u9808 \u8a08\u7b97\u4e00\u6b21\u7d71\u8a08\u503c\uff0c\u5c31\u53ef\u5c07\u6b64\u7d71\u8a08\u503c\u4f9b\u6574\uf906\uf9e8\u6bcf\u9805\u7279\u5fb5 [ ] x m \u4f5c\u6b63\u898f\u5316\u4f7f\u7528\u3002\u63db\u8a00\u4e4b\uff0c\uf967 \u540c\u9805\u7279\u5fb5 [ ] x m \u5171\u7528\u540c\u4e00\u7d44\u7d71\u8a08\u503c\u3002\u63a5\u4e0b\uf92d\uff0c\u6211\u5011\u5c07\u8a0e\uf941\u6574\uf906\u5f0f\u7279\u5fb5\u7d71\u8a08\u503c\u4f30\u6e2c\u6cd5\u904b\u7528 \u5728\u8a9e\u97f3\u7279\u5fb5\u6b63\u898f\u5316\u6280\u8853\u4e4b\u53ef\u80fd\u7684\u512a\u7f3a\u9ede\u3002 \u6574\uf906\u5f0f\u7279\u5fb5\u7d71\u8a08\u503c\u4f30\u6e2c\u6cd5\u904b\u7528\u5728\u8a9e\u97f3\u7279\u5fb5\u6b63\u898f\u5316\u6280\u8853\u4e4b\u512a\u7f3a\u9ede \u5728\u4ee5\u524d\u7684\u6587\u737b\uf9e8\uff0c\u5927\u591a\uf969\u5f37\u5065\u6027\u8a9e\u97f3\u7279\u5fb5\u6b63\u898f\u5316\u6280\u8853\u6240\u7528\u7684\u7d71\u8a08\u503c\uff0c\u7686\u662f\u85c9\u7531\u524d\u8ff0 \u7684\u6574\u6bb5\u8a9e\uf906\u4e4b\u8a9e\u97f3\u7279\u5fb5\u6240\u6c42\u5f97\uff0c\u96d6\u7136\u57f7\ufa08\u4e0a\u7c21\u55ae\u6709\u6548\uf961\uff0c\u800c\u4e14\u78ba\u5be6\u5c0d\u8a9e\u97f3\u7279\u5fb5\u6709\u660e\u986f\u63d0 \u5347\u5f37\u5065\u6027\u7684\u6548\u679c\uff0c\u4f46\u9084\u662f\u6709\u4e00\u4e9b\u6f5b\u5728\u7684\u7f3a\u9ede\uff0c\uf9b5\u5982\uff0c\u6574\uf906\u5f0f\u8a9e\u97f3\u7279\u5fb5\u6b63\u898f\u5316\u6280\u8853\u7121\u6cd5\u9054 \u5230\u5373\u6642\u8655\uf9e4(real-time processing)\u7684\u8981\u6c42\uff0c\u56e0\u70ba\u5c0d\u4e00\uf99a\uf905\u7684\u8a9e\u97f3\u7279\u5fb5\u5e8f\uf99c\u800c\u8a00\uff0c\u5fc5\u9808\u7b49\u5230 \u6700\u5f8c\u4e00\u500b\u8a9e\u97f3\u7279\u5fb5\u5f97\u5230\u4e4b\u5f8c\uff0c\u624d\u80fd\u6c42\u53d6\u7d71\u8a08\u503c\u3002\u9664\u6b64\u4e4b\u5916\uff0c\u96a8\u8457\u8a9e\uf906\u7684\uf967\u540c\uff0c\u800c\u7522\u751f\u7684", | |
| "cite_spans": [ | |
| { | |
| "start": 671, | |
| "end": 674, | |
| "text": "[ ]", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 694, | |
| "end": 697, | |
| "text": "[ ]", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 754, | |
| "end": 757, | |
| "text": "[ ]", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 780, | |
| "end": 783, | |
| "text": "[ ]", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "\u63d0 \u4e4b \u4e3b \u8981 \u4e09 \u7a2e \u7279 \u5fb5 \u7d71 \u8a08 \u503c \u4f30 \u6e2c \u6cd5 \uff0c \u5305 \u62ec \uf9ba \u6574 \uf906 \u5f0f (utterance-based)[8] \u3001 \u5206 \u6bb5 \u5f0f (segment-based)[8]\u8207\u78bc\u7c3f\u5f0f(codebook-based)[9]\u4e09\uf9d0\u65b9\u6cd5\uff0c\u53ca\u5b83\u5011\u53ef\u80fd\u7684\u512a\u9ede\u8207\u7f3a\u9ede\u3002 (\u4e00) \u6574\uf906\u5f0f\u7279\u5fb5\u7d71\u8a08\u503c\u4f30\u6e2c\u6cd5 \u5047\u8a2d\u67d0\u55ae\u4e00\u8a9e\uf906\u4e4b\u67d0\u4e00\u7dad\u7279\u5fb5\u5e8f\uf99c\u8868\u793a\u70ba [ ] { } ;1 x n n N \u2264 \u2264 (\u5f0f 2-1) \u5176\u4e2d N \u70ba\u7279\u5fb5\u5e8f\uf99c\u4e4b\u7279\u5fb5\u7e3d\u500b\uf969(\u5373\u97f3\u6846\u7e3d\uf969)\u3002\u5728\u6574\uf906\u5f0f\u7279\u5fb5\u7d71\u8a08\u503c\u4f30\u6e2c\u6cd5\uf9e8\uff0c\u6211\u5011\uf9dd \u7528(\u5f0f 2-1)\u6240\uf99c\u4e4b\u55ae\uf906\u6240\u6709\u7279\u5fb5\uff0c\u5171\u540c\u4f30\u6e2c\u7b2cm \u9805\u7279\u5fb5 [ ] x m \u7684\u7d71\u8a08\u503c\u3002\u63db\u8a00\u4e4b\uff0c\u6211\u5011\u5047 \u8a2d [ ] x m \u5c0d\u61c9\u81f3\u4e00\u96a8\u6a5f\u8b8a\uf969 [ ] X m \uff0c\u9032\u800c\u5047\u8a2d\u6574\uf906\u7279\u5fb5\u5e8f\uf99c [ ] { } ;1 x n n N \u2264 \u2264 \u70ba\u6b64\u96a8\u6a5f\u8b8a \uf969\u4e4b\u6a23\u672c(sample)\uff0c\u6839\u64da\u9019\u4e9b\u6a23\u672c\uff0c\u6211\u5011\u53ef\u4f30\u6e2c\u51fa [ ] X m \u6b64\u96a8\u6a5f\u8b8a\uf969\u7684\u5404\u7a2e\u7d71\u8a08\u503c\uff0c\uf9b5\u5982\uff1a 1. [ ] X m \u7684\u671f\u671b\u503c(\u5e73\u5747\u503c)\u70ba [ ] [ ] [ ] ,( ) 1 1 N X m u n m xn N \u03bc = = \u2211 , (\u5f0f 2-2) 2. [ ] X m \u7684\u8b8a\uf962\uf969(variance)\u70ba [ ] [ ] [ ] [ ] [ ] ( ) 2 2 ,( ) ,( ) 1 1 , N X m u X m u n m xn n N \u03c3 \u03bc = = \u2212 \u2211 (\u5f0f 2-3) 3. [ ] X m \u7684\u7b2cJ \u968e\u4e2d\u592e\u52d5\u5dee(central moment)\u70ba [ ] [ ] [ ] [ ] [ ] ( ) ( ) ,( ) ,( ) 1 1 , N J J X m u X m u n m xn n N \u03be \u03bc = = \u2212 \u2211 \u5176\u4e2dJ \u70ba\u4efb\u610f\u4e4b\u6b63\u5076\uf969 (\u5f0f 2-4) 4. [ ] X m \u7684\u52d5\u614b\u7bc4\u570d(dynamic range)\u70ba [ ] [ ] [ ] { } [ ] { } ,( ) 1 1 max min , X m u n N n N d m x n x n \u2264 \u2264 \u2264 \u2264 = \u2212 (\u5f0f 2-5) 5\u3001 [ ] X m \u7684\u6a5f\uf961\u5206\u4f48\u51fd\uf969(probability distribution function)\u70ba [ ] ( ) [ ] ( ) ,( ) 1 1 . N X m u n F z u z x n N = = \u2212 \u2211 (\u5f0f 2-6) \u5176\u4e2d\uff0c ( ) u i \u70ba\u55ae\u4f4d\u6b65\u968e\u51fd\uf969(unit step function)\uff0c\u5b9a\u7fa9\u70ba\uff1a ( ) 1, if 0 0, if 0 z u z z \u23a7 \u2265 \u23aa \u23aa = \u23a8 \u23aa < \u23aa \u23a9 \u5728\u4ee5\u4e0a\u4e94\u500b\u5f0f\u5b50\u4e2d\u7684\u5404\u7a2e\u7d71\u8a08\u503c\u7684\u4ee3\u865f\u4e2d\uff0c\u6211\u5011\u4ee5\u4e0b\u6a19 ( ) u \uf92d\u4ee3\u8868\u9019\u4e9b\u7d71\u8a08\u503c\u662f\u7531\u6574\uf906 (utterance)\u7684\u7279\u5fb5\u4f30\u6e2c\u800c\u5f97\uff0c\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u4ee5\u4e0a\u6240\u7b97\u4e4b\u91dd\u5c0d\u67d0\u4e00\u9805\u7279\u5fb5", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "\u95dc\u9375\u8a5e\uff1a\u8a9e\u97f3\u8fa8\uf9fc\u3001\u78bc\u7c3f\u3001\u7279\u5fb5\u7d71\u8a08\u503c\u4f30\u6e2c\u6cd5\u3001\u5f37\u5065\u6027\u8a9e\u97f3\u7279\u5fb5\uf96b\uf969", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "\u8a9e\u97f3\u7279\u5fb5\u5e8f\uf99c\u7684\u9577\ufa01(\u97f3\u6846\uf969)\u4e5f\uf967\u4e00\u6a23\uff0c\uf967\u540c\u8a9e\uf906\u6240\u5305\u542b\u7684\u97f3\u7d20\uf969\u76ee\u6216\u7a2e\uf9d0\uff0c\u53ca\u5176\u9577\ufa01 \u7684\u8b8a\u5316\u53ef\u80fd\u76f8\u5dee\u5f88\u5927\uff0c\u5c0e\u81f4\u5f71\u97ff\u5230\u6240\u4f30\u6e2c\u4e4b\u7d71\u8a08\u8cc7\u8a0a\u7684\u6e96\u78ba\u6027\u3002 (\u4e8c) \u5206\u6bb5\u5f0f\u7279\u5fb5\u7d71\u8a08\u503c\u4f30\u6e2c\u6cd5 \u9996\u5148\uff0c\u5047\u8a2d\u67d0\u55ae\u4e00\u8a9e\uf906\u7279\u5fb5 [ ] { } ;1 x n n N \u2264 \u2264 \u5176\u4e2d\u67d0\u4e00\u7247\u6bb5\u7279\u5fb5\u5e8f\uf99c\u8868\u793a\u70ba [ ] { } ; x k m L k m L \u2212 \u2264 \u2264 + (\u5f0f 2-7) \u5176\u4e2d 2 1 L + \u70ba\u7247\u6bb5\u7279\u5fb5\u5e8f\uf99c\u4e4b\u7279\u5fb5\u7e3d\uf969\uff0c(\u5f0f 2-7)\u6240\u8868\u793a\u7684\u5373\u70ba\u4ee5\u55ae\u9805\u7279\u5fb5 [ ] x m \u70ba\u4e2d\u5fc3 \u9ede\uff0c\u524d\u5f8c\u5404\u5ef6\u5c55 L \u9805\u7279\u5fb5\u6240\u5f97\u7684\u52d5\u614b\u7279\u5fb5\u7247\u6bb5\u3002\u5728\u5206\u6bb5\u5f0f\u7279\u5fb5\u7d71\u8a08\u503c\u4f30\u6e2c\u6cd5\uf9e8\uff0c\u7b2cm \u9805 \u7279\u5fb5 [ ] x m \u7684\u7d71\u8a08\u503c\uff0c\u662f\u85c9\u7531(\u5f0f 2-7)\u6240\uf99c\u4e4b\u7247\u6bb5\u8a9e\uf906\u6240\u6709\u7279\u5fb5\u4e2d\u6c42\u5f97\u3002\u63db\u8a00\u4e4b\uff0c\u6211\u5011\u5047 \u8a2d [ ] x m \u5c0d\u61c9\u81f3\u4e00\u96a8\u6a5f\u8b8a\uf969 [ ] X m \uff0c\u9032\u800c\u5047\u8a2d\u7247\u6bb5\u7279\u5fb5\u5e8f\uf99c [ ] { } ; x k m L k m L \u2212 \u2264 \u2264 + \u70ba \u6b64\u96a8\u6a5f\u8b8a\uf969\u4e4b\u6a23\u672c\uff0c\u7136\u5f8c\u6839\u64da\u6b64\u6a23\u672c\uff0c\u6211\u5011\u53ef\u4f30\u6e2c\u51fa\u96a8\u6a5f\u8b8a\uf969 [ ] X m \u7684\u5404\u7a2e\u7d71\u8a08\u503c\uff0c\u5982\uff1a 1. [ ] X m \u7684\u671f\u671b\u503c(mean)\u70ba [ ] [ ] [ ] ,( ) 1 , 2 1 m L X m s k m L m xk L \u03bc + = \u2212 = + \u2211 (\u5f0f 2-8) 2. [ ] X m \u7684\u8b8a\uf962\uf969(variance)\u70ba [ ] [ ] [ ] [ ] [ ] ( ) 2 2 ,( ) ,( ) 1 , 2 1 m L X m s X m s k m L m xk m L \u03c3 \u03bc + = \u2212 = \u2212 + \u2211 (\u5f0f 2-9) 3. [ ] X m \u7684\u7b2cJ \u968e\u4e2d\u592e\u52d5\u5dee(central moment)\u70ba [ ] [ ] [ ] [ ] [ ] ( ) ( ) ,( ) ,( ) 1 , 2 1 m L J J X m s X m s k m L m xk m L \u03be \u03bc + = \u2212 = \u2212 + \u2211 J \u70ba\u4efb\u610f\u4e4b\u6b63\u5076\uf969 (\u5f0f 2-10) 4. [ ] X m \u7684\u52d5\u614b\u7bc4\u570d(dynamic range)\u70ba [ ] [ ] [ ] { } [ ] { } ,( ) max min , X m s m L k m L m L k m L d m x k x k \u2212 \u2264 \u2264 + \u2212 \u2264 \u2264 + = \u2212 (\u5f0f 2-11) 5. [ ] X m \u7684\u6a5f\uf961\u5206\u4f48\u51fd\uf969(probability distribution function)\u70ba [ ] ( ) [ ] ( ) ,( ) 1 . 2 1 m L X m s k m L F z u z x k L + = \u2212 = \u2212 + \u2211 (\u5f0f 2-12) \u5728\u4ee5\u4e0a\u4e94\u500b\u5f0f\u5b50\u4e2d\u7684\u5404\u7a2e\u7d71\u8a08\u503c\u7684\u4ee3\u865f\u4e2d\uff0c\u6211\u5011\u4ee5\u4e0b\u6a19 ( ) s \uf92d\u4ee3\u8868\u9019\u4e9b\u7d71\u8a08\u503c\u662f\u7531\u7247\u6bb5 (segment)\u7684\u7279\u5fb5\u4f30\u6e2c\u800c\u5f97\uff0c\u5f9e\u4ee5\u4e0a\uf969\u5f0f\u4e4b\u5404\u7a2e\u7d71\u8a08\u503c\u6c42\u53d6\u6cd5\u5f97\u77e5\uff0c\u672c\u7bc0\u6240\u7528\u7684\u4f30\u6e2c\u6cd5\uf967 \u540c\u65bc\u4e0a\u4e00\u7bc0\u7684\u6574\uf906\u5f0f\u7d71\u8a08\u503c\u4f30\u6e2c\u6cd5\uff0c\u5b83\u6240\u91dd\u5c0d\u67d0\u4e00\u9805\u7279\u5fb5 [ ]", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "\u95dc\u9375\u8a5e\uff1a\u8a9e\u97f3\u8fa8\uf9fc\u3001\u78bc\u7c3f\u3001\u7279\u5fb5\u7d71\u8a08\u503c\u4f30\u6e2c\u6cd5\u3001\u5f37\u5065\u6027\u8a9e\u97f3\u7279\u5fb5\uf96b\uf969", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "x m \u6240\u5f97\u7684\u5404\u7a2e\u7d71\u8a08\u503c\uff0c\u4e8b \u5be6\u4e0a\u8207 [ ] x m \u4e2d\u7684\u5e8f\uf99c\u9806\u5e8fm \u6709\u95dc\uff0c\u4e5f\u5c31\u662f\uf96f\u5728\u5206\u6bb5\u5f0f\u7d71\u8a08\u503c\u4f30\u6e2c\u6cd5\u4e2d\uff0c\u6211\u5011\u5fc5\u9808\u8981\u500b \u5225\u8a08\u7b97\u6574\u6bb5\u8a9e\uf906\u4e2d\u6bcf\u4e00\u9805\u7279\u5fb5 [ ] x m ", | |
| "cite_spans": [ | |
| { | |
| "start": 19, | |
| "end": 22, | |
| "text": "[ ]", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 75, | |
| "end": 78, | |
| "text": "[ ]", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "\u95dc\u9375\u8a5e\uff1a\u8a9e\u97f3\u8fa8\uf9fc\u3001\u78bc\u7c3f\u3001\u7279\u5fb5\u7d71\u8a08\u503c\u4f30\u6e2c\u6cd5\u3001\u5f37\u5065\u6027\u8a9e\u97f3\u7279\u5fb5\uf96b\uf969", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "\u7684\u7d71\u8a08\u503c\uff0c\u63a5\u8457\u5c07\u6bcf\u9805\u7279\u5fb5 [ ] x m \u7684\u7d71\u8a08\u503c\u4f9b\u7576\u4e0b\u7684 \u7279\u5fb5 [ ] x m \u4f5c\u6b63\u898f\u5316\u8655\uf9e4\u3002\u63db\u8a00\u4e4b\uff0c\uf967\u540c\u9805\u7279\u5fb5 [ ] x m \u6240\u7528\u7684\u7d71\u8a08\u503c\u6703\uf967\u4e00\u6a23\u3002\u4ee5\u4e0b\uff0c\u6211 \u5011\u5c07\u8a0e\uf941\u5206\u6bb5\u5f0f\u7279\u5fb5\u7d71\u8a08\u503c\u4f30\u6e2c\u6cd5\u904b\u7528\u5728\u8a9e\u97f3\u7279\u5fb5\u6b63\u898f\u5316\u6280\u8853\u4e4b\u53ef\u80fd\u7684\u512a\u7f3a\u9ede\u3002 \u5206\u6bb5\u5f0f\u7279\u5fb5\u7d71\u8a08\u503c\u4f30\u6e2c\u6cd5\u904b\u7528\u5728\u8a9e\u97f3\u7279\u5fb5\u6b63\u898f\u5316\u6280\u8853\u4e4b\u512a\u7f3a\u9ede \u5206\u6bb5\u5f0f\u8a9e\u97f3\u7279\u5fb5\u6b63\u898f\u5316\u6280\u8853\u53ef\u4ee5\u5f4c\u88dc\u6574\u6bb5\u5f0f\u6280\u8853\u7684\u7f3a\u9ede\uff0c\u4f7f\u5176\u53ef\u4ee5\u9054\u5230\u8fd1\u4f3c\u5373\u6642\u8655 \uf9e4\u7684\u6548\u679c\u3002\u5047\u5982\u7247\u6bb5(\u5373\u52d5\u614b\u7684\u8996\u7a97)\u9577\ufa01\u8d8a\u77ed\uff0c\u5373\u6642\u8655\uf9e4\u7684\u512a\u9ede\u8d8a\u660e\u986f\uff0c\u4e14\u56e0\u7247\u6bb5\u9577\ufa01 \u8a2d\u70ba\u56fa\u5b9a\u5e38\uf969\uff0c\u5176\u5305\u542b\u7684\u97f3\u7d20\uf969\u76ee\u76f8\u5c0d\u800c\u8a00\u8f03\u5c11\uff0c\uf967\u540c\u7247\u6bb5\u6240\u5305\u542b\u7684\u97f3\u7d20\uf969\u76ee\u8f03\u70ba\u4e00 \u81f4\uff0c\u6240\u4f30\u6e2c\u4e4b\u7d71\u8a08\u503c\u7684\u6e96\u78ba\u6027\u53d7\u5230\u4e00\u7247\u6bb5\u7279\u5fb5\u4e2d\u7684\u97f3\u7d20\uf969\u76ee\u5f71\u97ff\u8f03\u5c0f\uff0c\u56e0\u6b64\ufa09\u4f4e\uf9ba\u76f8\u540c \u97f3\u7d20\u5728\uf967\u540c\u8a9e\uf906\u4e4b\u9593\u7684\u8b8a\uf962\u6027\u3002\u7136\u800c\u5176\u7f3a\u9ede\u70ba\uff0c\uf974\u7247\u6bb5\u9577\ufa01\uf967\u5920\u9577\uff0c\u4ee3\u8868\u80fd\u7528\u4ee5\u4f30\u6e2c\u7684 \u6a23\u672c\uf969\u8f03\u5c11\uff0c\u5247\u4f30\u6e2c\u5230\u7684\u7d71\u8a08\u503c\u53ef\u80fd\u6703\u8f03\uf967\u7cbe\u78ba\uff0c\u5c0e\u81f4\u7279\u5fb5\u7d71\u8a08\u6b63\u898f\u5316\u7684\u6548\u679c\u8b8a\u5dee\uff0c\u9019 \u610f\u5473\u8457\u5728\u9019\u5206\u6bb5\u5f0f\u6280\u8853\u4e2d\uff0c\u53ef\u80fd\u7121\u6cd5\u540c\u6642\u9054\u6210\u5373\u6642\u8655\uf9e4\u7684\u6548\u679c\u8207\u5927\u5e45\u6b63\u898f\u5316\u7684\u8fa8\uf9fc\u6e96\u78ba \u6027\uff0c\u56e0\u6b64\u901a\u5e38\u5fc5\u9808\u5728\u5373\u6642\u8655\uf9e4\u8207\u5f37\u5065\u6027\u6548\u80fd\u9019\uf978\u8005\u512a\u9ede\u4e4b\u9593\u4f5c\u53d6\u6368(trade-off)\u3002 (\u4e09) \u78bc\u7c3f\u5f0f\u7279\u5fb5\u7d71\u8a08\u503c\u4f30\u6e2c\u6cd5 \u5728\u9019\u4e00\u7bc0\u4e2d\uff0c\u6211\u5011\u7c21\u8981\u4ecb\u7d39\u5982\u4f55\uf9dd\u7528\u78bc\u7c3f\u8cc7\u8a0a\uf92d\u4f30\u6e2c\u7279\u5fb5\u7684\u5404\u7a2e\u7d71\u8a08\u503c\uff0c\u800c\u78bc\u7c3f\u69cb \u6210 \u7684 \u8a73 \u7d30 \u7a0b \u5e8f \u8acb \uf96b \u7167 \u6587 \u737b [7,9] \u3002 \u9996 \u5148 \uff0c \u5047 \u8a2d \u67d0 \u55ae \u4e00 \u8a9e \uf906 \u4e4b \u67d0 \u4e00 \u7dad \u7279 \u5fb5 \u5e8f \uf99c [ ] { } ;1 x n n N \u2264 \u2264 \u6240\u69cb\u6210\u7684\u4e00\u7d44\u78bc\u7c3f\uff0c\u8868\u793a\u70ba [ ] { } , ;1 r y r w r M \u2264 \u2264 (\u5f0f 2-13) \u5176\u4e2d r w \u70ba\u6bcf\u4e00\u78bc\u5b57 [ ] y r \u6240\u5c0d\u61c9\u7684\u6b0a\u91cd\u503c\uff0c\u800c M \u70ba\u78bc\u5b57\u7e3d\uf969\u3002\u5728\u78bc\u7c3f\u5f0f\u7279\u5fb5\u7d71\u8a08\u503c\u4f30\u6e2c \u6cd5\u65b9\u9762\uff0c\u6211\u5011\uf9dd\u7528(\u5f0f 2-13)\u6240\u793a\u4e4b\u6574\u7d44\u78bc\u5b57\uff0c\u53bb\u6c42\u5f97\u6bcf\u4e00\u9805\u7279\u5fb5 [ ] x m \u6240\u5c0d\u61c9\u4e4b\u96a8\u6a5f\u8b8a\uf969 [ ] X m \u5176\u7d71\u8a08\u503c\u5982\u4e0b\uff1a 1. [ ] X m \u7684\u671f\u671b\u503c(mean)\u70ba [ ] [ ] [ ] ,( ) 1 M X m c r r m wy r \u03bc = = \u2211 , (\u5f0f 2-14) 2. [ ] X m \u7684\u8b8a\uf962\uf969(variance)\u70ba [ ] [ ] [ ] [ ] [ ] ( ) 2 2 ,( ) ,( ) 1 M X m c r X m c r m w y r m \u03c3 \u03bc = = \u2212 \u2211 , (\u5f0f 2-15) 3. [ ] X m \u7684\u7b2cJ \u968e\u4e2d\u592e\u52d5\u5dee(central moment)\u70ba [ ] [ ] [ ] [ ] [ ] ( ) ( ) ,( ) ,( ) 1 M J J X m c r X m c r m w y r m \u03be \u03bc = = \u2212 \u2211 ,J \u70ba\u4efb\u610f\u4e4b\u6b63\u5076\uf969 (\u5f0f 2-16) 4. [ ] X m \u7684\u52d5\u614b\u7bc4\u570d(dynamic range)\u70ba [ ] [ ] [ ] { } [ ] { } ,( ) 1 1 max min X m c r M r M d m y r y r \u2264 \u2264 \u2264 \u2264 = \u2212 , (\u5f0f 2-17) 5. [ ] X m \u7684\u6a5f\uf961\u5206\u4f48\u51fd\uf969(probability distribution function)\u70ba [ ] ( ) [ ] ( ) ,( ) 1 . M X m c r r F z wu z y r = = \u2212 \u2211 (\u5f0f 2-18) \u5f9e\u4e0a\u8ff0\u5404\u5f0f\u6240\u793a\uff0c\u6211\u5011\u5f97\u77e5\u67d0\u4e00\u9805\u7279\u5fb5 [ ] x m \u6240\u5c0d\u61c9\u7684\u5404\u7a2e\u7d71\u8a08\u503c\uff0c\u5176\u5be6\u8207 [ ] x m \u4e2d\u7684\u5e8f \uf99c \u9806 \u5e8f m \u7121 \u95dc \uff0c \u4e5f \u5c31 \u662f \uf96f \u5728 \u78bc \u7c3f \u5f0f \u7d71 \u8a08 \u503c \u4f30 \u6e2c \u6cd5 \u65b9 \u9762 \uff0c \u6211 \u5011 \u5f9e \u4e00 \u7d44 \u78bc \u5b57 [ ] { } , ;1 r y r w r M \u2264 \u2264 \u4e2d\uff0c\u53ea\u8a08\u7b97\u4e00\u6b21\u7d71\u8a08\u503c\uff0c\u5c31\u53ef\u4f9b\u6574\u6bb5\u8a9e\uf906\u4e2d\u7684\u6bcf\u9805\u7279\u5fb5 [ ] x m \u4f5c\u6b63 \u898f\u5316\u8655\uf9e4\u3002\u63db\u8a00\u4e4b\uff0c\uf967\u540c\u9805\u7279\u5fb5 [ ] x m \u5c07\u5171\u7528\u540c\u4e00\u7d44\u7d71\u8a08\u503c\uff0c\u6240\u4ee5\u672c\u7bc0\u4f30\u6e2c\u6cd5\uf9d0\u4f3c\u65bc\u6574", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "\u95dc\u9375\u8a5e\uff1a\u8a9e\u97f3\u8fa8\uf9fc\u3001\u78bc\u7c3f\u3001\u7279\u5fb5\u7d71\u8a08\u503c\u4f30\u6e2c\u6cd5\u3001\u5f37\u5065\u6027\u8a9e\u97f3\u7279\u5fb5\uf96b\uf969", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "EQUATION", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [ | |
| { | |
| "start": 0, | |
| "end": 8, | |
| "text": "EQUATION", | |
| "ref_id": "EQREF", | |
| "raw_str": "{ } ;1 x n n N \u2264 \u2264 \uff0c\u5247\u7d93\u904e CMS \u8655\uf9e4\u5f8c\u7684\u8f38\u51fa\u7279\u5fb5\uf96b\uf969\u8868\u793a\u5f0f\u5982\u4e0b\uff1a [ ] [ ] [ ] , 1 x n x n n n N \u03bc = \u2212 \u2264 \u2264 , (\u5f0f 3-1) \u800c\u7d93\u904e CMVN \u8655\uf9e4\u5f8c\u7684\u7279\u5fb5\uf96b\uf969\u8868\u793a\u5f0f\u5982\u4e0b\uff1a [ ] [ ] [ ] ( ) [ ] / , 1 x n x n n n n N \u03bc \u03c3 = \u2212 \u2264 \u2264 , (\u5f0f 3-2) \u5176\u4e2d N \u70ba\u6574\u6bb5\u5e8f\uf99c\u4e4b\u7279\u5fb5\u7e3d\uf969\uff0c\u800c [ ] n \u03bc \u8207 [ ] n \u03c3 \u5206\u5225\u70ba\u7279\u5fb5 [ ] x n \u7684\u5e73\u5747\u503c\u8207\u6a19\u6e96\u5dee\u3002 \u5728\u7b2c\u4e00\u7a2e\u4f75\u5408\u5f0f\u7279\u5fb5\u7d71\u8a08\u4f30\u6e2c\u6cd5\u4e2d\uff0c [ ] n \u03bc \u8207 [ ] n \u03c3 \u53ef\u7531\u4e0b\uf99c\uf978\u516c\u5f0f\u4f30\u6e2c\u800c\u5f97\uff1a CU-CMS/CU-CMVN\uff1a [ ] [ ] ( ) [ ] ( , ) ( ) ( ) 1 , c u c u n n n \u03bc \u03b1 \u03bc \u03b1 \u03bc = + \u2212 (\u5f0f 3-3) [ ] [ ] [ ] ( ) [ ] [ ] [ ] 2 2 2 2 2 2 ( , ) ( ) ( ) ( ) ( ) ( , ) 1 c u c c u u c u n n n n n n \u03c3 \u03b1\u03c3 \u03bc \u03b1 \u03c3 \u03bc \u03bc \u23a1 \u23a4 \u23a1 \u23a4 = + + \u2212 + \u2212 \u23a2 \u23a5 \u23a2 \u23a5 \u23a3 \u23a6 \u23a3 \u23a6 , (\u5f0f 3-4) \u5176\u4e2d\u4e0b\u6a19\"( ) c \"\u3001\"( ) u \"\u8207\"( , ) c u \"\u5206\u5225\u4ee3\u8868\u4f7f\u7528\u78bc\u7c3f\u5f0f\u3001\u6574\uf906\u5f0f\u8207\u4f75\u5408\u78bc\u7c3f/\u6574\uf906\u5f0f\u7d71\u8a08\u503c \u4f30\u6e2c\u6cd5\uff0c\u800c [ ] ( ) c n \u03bc \u3001 [ ] ( ) u n \u03bc \u3001 [ ] 2 ( ) c n \u03c3 \u8207 [ ] 2 ( ) u n \u03c3 \u5206\u5225\u5b9a\u7fa9\u65bc\u524d\u4e00\u7ae0\u7684(\u5f0f 2-14)\u3001(\u5f0f 2-2)\u3001(\u5f0f 2-15)\u8207(\u5f0f 2-3)\uff0c \u03b1 \u70ba\u6b0a\u91cd\u503c\uff0c\u4ecb\u65bc 0 \u5230 1 \u4e4b\u9593\uff0c\u88ab\u7528\uf92d\u8abf\u6574\u78bc\u7c3f\u5f0f\u7d71\u8a08\u8cc7\u8a0a\u8207\u6574\u6bb5\u5f0f \u7d71\u8a08\u8cc7\u8a0a\u4e4b\u9593\u7684\u6bd4\uf9b5\u3002\u85c9\u7531(\u5f0f 3-3)\u8207(\u5f0f 3-4)\u6240\u4f30\u6e2c\u4e4b\u5e73\u5747\u503c\u8207\u8b8a\uf962\uf969\u800c\u6210\u7684 CMS \u8207 CMVN\uff0c\u6211\u5011\u5206\u5225\u7a31\u70ba\u4f75\u5408\u78bc\u7c3f/\u6574\uf906\u5f0f CMS(hybrid codebook/utterance-based CMS, CU-CMS) \u8207 \u4f75 \u5408 \u78bc \u7c3f / \u6574 \uf906 \u5f0f CMVN(hybrid codebook/utterance-based CMVN, CU-CMVN)\uff0c\u7531(\u5f0f 3-3)\u8207(\u5f0f 3-4)\u53ef\u660e\u986f\u770b\u51fa\uff0cCU-CMS \u8207 CU-CMVN \u6240\u4f7f\u7528\u7684\u5e73\u5747\u503c \u8207\u8b8a\uf962\uf969\u662f\u5c07\u524d\u4e00\u7ae0\u6240\u8ff0\u4e4b\u8a9e\u97f3\u7279\u5fb5\u78bc\u7c3f\u8207\u6574\u6bb5\u8a9e\u97f3\u7279\u5fb5\u7684\u5e73\u5747\u503c\u8207\u8b8a\uf962\uf969\u4f5c\u4e00\u7dda\u6027 \u7684\u7d44\u5408\u3002\u5982\u679c\u6b0a\u91cd\u503c 1 \u03b1 = \u6642\uff0cCU-CMS \u548c CU-CMVN \u5c07\u5206\u5225\u7b49\u540c\u65bc\u78bc\u7c3f\u5f0f CMS (C-CMS) \u548c\u78bc\u7c3f\u5f0f CMVN (C-CMVN)\uff0c\u53e6\u4e00\u65b9\u9762\uff0c\u5982\u679c 0 \u03b1 = \u6642\uff0cCU-CMS \u548c CU-CMVN \u5c07\u5206 \u5225\u7b49\u540c\u65bc\u6574\uf906\u5f0f CMS (U-CMS)\u548c\u6574\uf906\u5f0f CMVN (U-CMVN)\u3002 \u5728\u7b2c\u4e8c\u7a2e\u4f75\u5408\u5f0f\u7279\u5fb5\u7d71\u8a08\u4f30\u6e2c\u6cd5\u4e2d\uff0c [ ] n \u03bc \u8207 [ ] n \u03c3 \u53ef\u7531\u4e0b\uf99c\uf978\u516c\u5f0f\u4f30\u6e2c\u800c\u5f97\uff1a CS-CMS/CS-CMVN\uff1a [ ] [ ] ( ) [ ]", | |
| "eq_num": "( , ) ( ) ( ) 1" | |
| } | |
| ], | |
| "section": "\u95dc\u9375\u8a5e\uff1a\u8a9e\u97f3\u8fa8\uf9fc\u3001\u78bc\u7c3f\u3001\u7279\u5fb5\u7d71\u8a08\u503c\u4f30\u6e2c\u6cd5\u3001\u5f37\u5065\u6027\u8a9e\u97f3\u7279\u5fb5\uf96b\uf969", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "EQUATION", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [ | |
| { | |
| "start": 0, | |
| "end": 8, | |
| "text": "EQUATION", | |
| "ref_id": "EQREF", | |
| "raw_str": ", c s c s n n n \u03bc \u03b1 \u03bc \u03b1 \u03bc = + \u2212 (\u5f0f 3-5) [ ] [ ] [ ] ( ) [ ] [ ] [ ] 2 2 2 2 2 2 ( , )", | |
| "eq_num": "( ) ( ) ( ) ( ) ( , ) 1 c" | |
| } | |
| ], | |
| "section": "\u95dc\u9375\u8a5e\uff1a\u8a9e\u97f3\u8fa8\uf9fc\u3001\u78bc\u7c3f\u3001\u7279\u5fb5\u7d71\u8a08\u503c\u4f30\u6e2c\u6cd5\u3001\u5f37\u5065\u6027\u8a9e\u97f3\u7279\u5fb5\uf96b\uf969", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "EQUATION", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [ | |
| { | |
| "start": 0, | |
| "end": 8, | |
| "text": "EQUATION", | |
| "ref_id": "EQREF", | |
| "raw_str": "= + + \u2212 + \u2212 \u23a2 \u23a5 \u23a2 \u23a5 \u23a3 \u23a6 \u23a3 \u23a6 (\u5f0f 3-6) \u5176\u4e2d\u4e0b\u6a19\"( ) c \"\u3001\"( ) s \"\u8207\"( , ) c s \"\u5206\u5225\u4ee3\u8868\u4f7f\u7528\u78bc\u7c3f\u5f0f\u3001\u5206\u6bb5\u5f0f\u8207\u4f75\u5408\u78bc\u7c3f/\u5206\u6bb5\u5f0f\u7d71\u8a08\u503c \u4f30\u6e2c\u6cd5\uff0c\u800c [ ] ( ) c n \u03bc \u3001 [ ] ( ) s n \u03bc \u3001 [ ] 2 ( ) c n \u03c3 \u8207 [ ] 2 ( ) s n \u03c3 \u5206\u5225\u5b9a\u7fa9\u65bc\u524d\u4e00\u7ae0\u7684(\u5f0f 2-14)\u3001(\u5f0f 2-8)\u3001 (\u5f0f 2-15)\u8207(\u5f0f 2-9)\uff0c \u03b1 \u70ba\u6b0a\u91cd\u503c\uff0c\u4ecb\u65bc 0 \u5230 1 \u4e4b\u9593\uff0c\u88ab\u7528\uf92d\u8abf\u6574\u78bc\u7c3f\u5f0f\u7d71\u8a08\u8cc7\u8a0a\u8207\u5206\u6bb5\u5f0f \u7d71\u8a08\u8cc7\u8a0a\u4e4b\u9593\u7684\u6bd4\uf9b5\u3002\u85c9\u7531(\u5f0f 3-5)\u8207(\u5f0f 3-6)\u6240\u4f30\u6e2c\u4e4b\u5e73\u5747\u503c\u8207\u8b8a\uf962\uf969\u800c\u6210\u7684 CMS \u8207 CMVN \u6cd5\uff0c\u6211\u5011\u5206\u5225\u7a31\u70ba\u4f75\u5408\u78bc\u7c3f/\u5206\u6bb5\u5f0f CMS(hybrid codebook/segment-based CMS, CS-CMS) \u8207 \u4f75 \u5408 \u78bc \u7c3f / \u5206 \u6bb5 \u5f0f CMVN(hybrid codebook/segment-based CMVN, CS-CMVN)\uff0c\uf9d0\u4f3c\u524d\u9762\u6240\u63d0\u4e4b CU-CMS \u8207 CU-CMVN\uff0c\u5f9e(\u5f0f 3-5)\u8207(\u5f0f 3-6)\u53ef\u770b\u51fa\uff0c CS-CMS \u8207 CS-CMVN \u6240\u4f7f\u7528\u7684\u5e73\u5747\u503c\u8207\u8b8a\uf962\uf969\u662f\u5c07\u524d\u4e00\u7ae0\u6240\u8ff0\u4e4b\u8a9e\u97f3\u7279\u5fb5\u78bc\u7c3f\u8207\u7247\u6bb5 \u8a9e\u97f3\u7279\u5fb5\u7684\u5e73\u5747\u503c\u8207\u8b8a\uf962\uf969\u4f5c\u4e00\u7dda\u6027\u7684\u7d44\u5408\u3002\u5982\u679c\u6b0a\u91cd 1 \u03b1 = \u6642\uff0cCS-CMS \u548c CS-CMVN \u5c07\u5206\u5225\u7b49\u540c\u65bc\u78bc\u7c3f\u5f0f CMS (C-CMS)\u548c\u78bc\u7c3f\u5f0f CMVN (C-CMVN)\uff0c\u7136\u800c\uff0c\uf974 0 \u03b1 = \u6642\uff0c CS-CMS \u548c CS-CMVN \u5c07\u5206\u5225\u7b49\u540c\u65bc\u5206\u6bb5\u5f0f CMS (S-CMS)\u548c\u5206\u6bb5\u5f0f CMVN (S-CMVN) \u3002 (\u4e8c) \u4f75\u5408\u5f0f\u9ad8\u968e\u5012\u983b\u8b5c\u52d5\u5dee\u6b63\u898f\u5316\u6cd5 \u5c0d\u4e00\u7279\u5fb5\u6642\u9593\u5e8f\uf99c [ ] { } ;1 x n n N \u2264 \u2264 \u800c\u8a00\uff0c\u7d93\u9ad8\u968e\u5012\u983b\u8b5c\u52d5\u5dee\u6b63\u898f\u5316\u6cd5(HOCMN) \u8655\uf9e4\u5f8c\u6240\u5f97\u7684\u65b0\u7279\u5fb5\u6642\u9593\u5e8f\uf99c\u5982\u4e0b\u5f0f\uff1a [ ] [ ] [ ] ( ) [ ] ( ) 1 ( ) / , 1 J J x n x n n n n N \u03bc \u03be = \u2212 \u2264 \u2264 (\u5f0f 3-7) \u5176\u4e2d N \u70ba\u6574\u6bb5\u5e8f\uf99c\u4e4b\u7279\u5fb5\uf969\uff0c [ ] n \u03bc \u8207 [ ] ( ) J n \u03be \u5206\u70ba\u7279\u5fb5 [ ] x n \u7684\u5e73\u5747\u503c\u8207\u7b2cJ \u968e\u4e2d\u592e\u52d5\u5dee\u3002 \uf9d0\u4f3c\u4e0a\u4e00\u7bc0\u6240\u8ff0\uff0c\u9019\uf9e8\u6211\u5011\u6709\uf978\u7a2e\u65b9\u5f0f\uf92d\u4f30\u6e2c [ ] n \u03bc \u8207 [ ] ( ) J n \u03be \uff0c\u6240\u5c0d\u61c9\u7684 HOCMN \u6cd5 \u6211 \u5011 \u5206 \u5225 \u7a31 \u70ba \u4f75 \u5408 \u78bc \u7c3f / \u6574 \uf906 \u5f0f HOCMN(CU-HOCMN) \u8207 \u4f75 \u5408 \u78bc \u7c3f / \u5206 \u6bb5 \u5f0f HOCMN(CS-HOCMN)\uff0c\u5b83\u5011\u5728 [ ] n \u03bc \u8207 [ ] ( ) J n \u03be \u7684\u4f30\u6e2c\u904b\u7b97\u5982\u4e0b\uf99c\uf969\u5f0f\uff1a CU-HOCMN (hybrid codebook/utterance-based HOCMN)\uff1a [ ] [ ] ( ) [ ] ( , ) ( ) ( ) 1 , c u c u n n n \u03bc \u03b1 \u03bc \u03b1 \u03bc = + \u2212 (\u5f0f 3-8) [ ] [ ] [ ] ( ) ( ) [ ] [ ] ( ) ( ) ( , )", | |
| "eq_num": "( , ) ( , )" | |
| } | |
| ], | |
| "section": "\u95dc\u9375\u8a5e\uff1a\u8a9e\u97f3\u8fa8\uf9fc\u3001\u78bc\u7c3f\u3001\u7279\u5fb5\u7d71\u8a08\u503c\u4f30\u6e2c\u6cd5\u3001\u5f37\u5065\u6027\u8a9e\u97f3\u7279\u5fb5\uf96b\uf969", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "+ = = \u2212 \u239b \u239e \u239b \u239e \u239f \u239f \u239c \u239c = \u2212 + \u2212 \u2212 \u239f \u239f \u239c \u239c \u239f \u239f \u239f \u239f \u239c \u239c + \u239d \u23a0 \u239d \u23a0 \u2211 \u2211 (\u5f0f 3-11) \u5176\u4e2d [ ] ( ) c n \u03bc \u3001 [ ] ( ) u n \u03bc \u3001 [ ] ( ) s n \u03bc \u5206\u5225\u5b9a\u7fa9\u65bc\u524d\u4e00\u7ae0\u7684(\u5f0f 2-14)\u3001(\u5f0f 2-2)\u8207(\u5f0f 2-8)\uff0c \u03b1 \u70ba\u6b0a \u91cd\u503c\uff0c\u4ecb\u65bc 0 \u5230 1 \u4e4b\u9593\uff0c\u7528\uf92d\u8abf\u6574\u78bc\u7c3f\u7d71\u8a08\u8cc7\u8a0a\u8207\u6574\u6bb5\u6216\u7247\u6bb5\u7279\u5fb5\u7d71\u8a08\u8cc7\u8a0a\u4e4b\u9593\u7684\u6bd4\uf9b5\u3002 (\u4e09) \u4f75\u5408\u5f0f\u5012\u983b\u8b5c\u589e\u76ca\u6b63\u898f\u5316\u6cd5 \u5c0d\u4e00\u7279\u5fb5\u6642\u9593\u5e8f\uf99c [ ] { } ;1 x n n N \u2264 \u2264 \u800c\u8a00\uff0c\u7d93\u5012\u983b\u8b5c\u589e\u76ca\u6b63\u898f\u6cd5(CGN)\u8655\uf9e4\u5f8c\u6240\u5f97 \u7684\u65b0\u7279\u5fb5\u6642\u9593\u5e8f\uf99c\u5982\u4e0b\u5f0f\uff1a [ ] [ ] [ ] ( ) [ ] / , 1 x n x n n d n n N \u03bc = \u2212 \u2264 \u2264 , (\u5f0f 3-12) \u5176\u4e2d N \u70ba\u6574\u6bb5\u5e8f\uf99c\u4e4b\u7279\u5fb5\u7e3d\uf969\uff0c [ ] n \u03bc \u8207 [ ] d n \u5206\u5225\u70ba\u7279\u5fb5 [ ] x n \u7684\u5e73\u5747\u503c\u8207\u52d5\u614b\u7bc4\u570d\u3002 \uf9d0\u4f3c\u4e0a\uf978\u7bc0\u7684\u65b9\u6cd5\uff0c\u9019\uf9e8\u6211\u5011\u6709\uf978\u7a2e\u65b9\u5f0f\uf92d\u4f30\u6e2c [ ] n \u03bc \u8207 [ ] d", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "\u95dc\u9375\u8a5e\uff1a\u8a9e\u97f3\u8fa8\uf9fc\u3001\u78bc\u7c3f\u3001\u7279\u5fb5\u7d71\u8a08\u503c\u4f30\u6e2c\u6cd5\u3001\u5f37\u5065\u6027\u8a9e\u97f3\u7279\u5fb5\uf96b\uf969", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "\u5176\u4e2d ( ) c Y \u3001 ( ) u X \u8207 ( ) s X \u4ee3\u8868\uf9ba(\u5f0f 2-13)\u3001(\u5f0f 2-1)\u8207(\u5f0f 2-7)\u3002 [ ] ( ) c n \u03bc \u3001 [ ] ( ) u n \u03bc \u8207 [ ] ( ) s n \u03bc \u5206\u5225 \u5b9a\u7fa9\u65bc\u524d\u4e00\u7ae0\u7684(\u5f0f 2-14)\u3001(\u5f0f 2-2)\u8207(\u5f0f 2-8)\uff0c\u5176\u4e2d \u03b1 \u70ba\u4e00\u500b\u4ecb\u65bc 0 \u5230 1 \u4e4b\u9593\u7684\u6b0a\u91cd\u503c\uff0c \u4ee3\u8868\uf9ba\u78bc\u7c3f\u5f0f\u7d71\u8a08\u8cc7\u8a0a\u8207\u6574\uf906\u5f0f\u6216\u5206\u6bb5\u5f0f\u7d71\u8a08\u8cc7\u8a0a\u9019\uf978\u8005\u4e4b\u9593\u6240\u4f7f\u7528\u7684\u6bd4\uf9b5\uff0c ( ) max \u22c5 \u8207 ( ) min \u22c5 \u5206\u5225\u70ba\u53d6\u6700\u5927\u503c\u8207\u6700\u5c0f\u503c\u7684\u51fd\uf969\uff0c\u800c\u300e \u222a \u300f\u70ba\uf997\u96c6\u7b26\u865f\uff0c\u610f\u6307\u5c07\u78bc\u7c3f\u8207\u6574\uf906(\u6216 \u7247\u6bb5)\u8a9e\uf906\u7684\u7279\u5fb5\uf905\u5728\u4e00\u8d77\u3002 (\u56db) \u4f75\u5408\u5f0f\u5012\u983b\u8b5c\u7d71\u8a08\u5716\u7b49\u5316\u6cd5 \u5c0d\u4e00\u7279\u5fb5\u6642\u9593\u5e8f\uf99c [ ] { } ;1 x n n N \u2264 \u2264 \u800c\u8a00\uff0c\u7d93\u5012\u983b\u8b5c\u7d71\u8a08\u5716\u7b49\u5316\u6cd5(HEQ)\u8655\uf9e4\u5f8c\u6240 \u5f97\u7684\u65b0\u7279\u5fb5\u6642\u9593\u5e8f\uf99c\u5982\u4e0b\u5f0f\uff1a [ ] [ ] ( ) ( ) 1 , 1 ref X x n F F x n n N \u2212 = \u2264 \u2264 , (\u5f0f 3-17) \u5176\u4e2d N \u70ba\u6574\u6bb5\u5e8f\uf99c\u4e4b\u7279\u5fb5\u7e3d\uf969\uff0c ( ) ref F \u22c5 \u70ba\u9810\u5148\u5b9a\u7fa9\u7684\uf96b\u8003\u6a5f\uf961\u5206\u5e03\u51fd\uf969\uff0c\u800c ( ) X F \u22c5 \u5247\u70ba \u7279\u5fb5 [ ] x n \u7684\u6a5f\uf961\u5206\u5e03\u51fd\uf969\u3002 \uf9d0\u4f3c\u524d\u9762\u5e7e\u7bc0\u6240\u8ff0\uff0c\u9019\uf9e8\u6211\u5011\u6709\uf978\u7a2e\u65b9\u5f0f\uf92d\u4f30\u6e2c\u6a5f\uf961\u5206\u5e03\u51fd\uf969 ( ) X F \u22c5 \uff0c\u5206\u5225\u4f7f\u7528\u5728 HEQ \u4e0a\uff0c\u56e0\u6b64\u6211\u5011\u5206\u5225\u7a31\u70ba\u4f75\u5408\u78bc\u7c3f/\u6574\uf906\u5f0f HEQ(CU-HEQ)\u8207\u4f75\u5408\u78bc\u7c3f/\u5206\u6bb5\u5f0f HEQ(CS-HEQ)\uff0c\u5b83\u5011\u5c0d ( ) X F \u22c5 \u7684\u4f30\u6e2c\u8868\u793a\u5f0f\u5982\u4e0b\u6240\u793a\uff1a CU-HEQ(hybrid codebook/utterance-based HEQ)\uff1a ( ) ( ) ( ) ( ) ,( , ) ,( ) ,( ) 1 , X c u X c X u F z F z F z \u03b1 \u03b1 = + \u2212 (\u5f0f 3-18)", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "\u95dc\u9375\u8a5e\uff1a\u8a9e\u97f3\u8fa8\uf9fc\u3001\u78bc\u7c3f\u3001\u7279\u5fb5\u7d71\u8a08\u503c\u4f30\u6e2c\u6cd5\u3001\u5f37\u5065\u6027\u8a9e\u97f3\u7279\u5fb5\uf96b\uf969", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "CS-HEQ(hybrid codebook/segment-based HEQ)\uff1a ", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "\u95dc\u9375\u8a5e\uff1a\u8a9e\u97f3\u8fa8\uf9fc\u3001\u78bc\u7c3f\u3001\u7279\u5fb5\u7d71\u8a08\u503c\u4f30\u6e2c\u6cd5\u3001\u5f37\u5065\u6027\u8a9e\u97f3\u7279\u5fb5\uf96b\uf969", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "( ) ( ) ( ) ( ) ,( , ) ,( ) ,( ) 1 X c s X c X s F z F z F z \u03b1 \u03b1 = + \u2212 , (\u5f0f 3-19) \u5176\u4e2d \u03b1 \u70ba\u4e00\u500b\u4ecb\u65bc 0 \u5230 1 \u4e4b\u9593\u7684\u6b0a\u91cd\u503c\uff0c\u4ee3\u8868\uf9ba\u78bc\u7c3f\u5f0f\u7d71\u8a08\u8cc7\u8a0a\u8207\u6574\u6bb5\u5f0f\u6216\u5206\u6bb5\u5f0f\u7d71\u8a08 \u8cc7\u8a0a\u9019\uf978\u8005\u4e4b\u9593\u6240\u4f7f\u7528\u7684\u6bd4\uf9b5\uff0c\u800c ( ) ,( ) X c F \u22c5 \u3001 ( ) ,( ) X u F \u22c5 \u8207 ( ) ,( ) X s F \u22c5 \u5206\u5225\u5b9a\u7fa9\u65bc\u524d\u4e00\u7ae0\u7684(\u5f0f 2-18)\u3001(\u5f0f 2-6) \u8207(\u5f0f 2-", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "\u95dc\u9375\u8a5e\uff1a\u8a9e\u97f3\u8fa8\uf9fc\u3001\u78bc\u7c3f\u3001\u7279\u5fb5\u7d71\u8a08\u503c\u4f30\u6e2c\u6cd5\u3001\u5f37\u5065\u6027\u8a9e\u97f3\u7279\u5fb5\uf96b\uf969", | |
| "sec_num": null | |
| } | |
| ], | |
| "back_matter": [], | |
| "bib_entries": { | |
| "BIBREF1": { | |
| "ref_id": "b1", | |
| "title": "RASTA)\u7b49\uff0c\u4f7f\u5176\u63d0\u5347\u8fa8\uf9fc\u7d50\u679c\u3002", | |
| "authors": [ | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "\u8fa8\uf9fc\u6539\u5584\u7a0b\ufa01\uff0c\u8f03\u5dee\u65bc\u52a0\u6210\u6027\u96dc\u8a0a\u74b0\u5883\uff0c\u56e0\u6b64\u6211\u5011\u671f\u671b\u80fd\u7d50\u5408\u6d88\u9664\u901a\u9053\u6548\u61c9\u7684\u65b9\u6cd5\uff0c\uf9b5 \u5982\u76f8\u5c0d\u983b\u8b5c\u6cd5", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": null, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "\u8fa8\uf9fc\u6539\u5584\u7a0b\ufa01\uff0c\u8f03\u5dee\u65bc\u52a0\u6210\u6027\u96dc\u8a0a\u74b0\u5883\uff0c\u56e0\u6b64\u6211\u5011\u671f\u671b\u80fd\u7d50\u5408\u6d88\u9664\u901a\u9053\u6548\u61c9\u7684\u65b9\u6cd5\uff0c\uf9b5 \u5982\u76f8\u5c0d\u983b\u8b5c\u6cd5(RASTA)\u7b49\uff0c\u4f7f\u5176\u63d0\u5347\u8fa8\uf9fc\u7d50\u679c\u3002", | |
| "links": null | |
| }, | |
| "BIBREF2": { | |
| "ref_id": "b2", | |
| "title": "Cepstral analysis technique for automatic speaker verification", | |
| "authors": [ | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Furui", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1981, | |
| "venue": "IEEE Transactions on Acoustics, Speech and Signal Processing", | |
| "volume": "29", | |
| "issue": "2", | |
| "pages": "254--272", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "S. Furui, ''Cepstral analysis technique for automatic speaker verification'', IEEE Transactions on Acoustics, Speech and Signal Processing, Volume 29, Issue 2, pp. 254-272, 1981", | |
| "links": null | |
| }, | |
| "BIBREF3": { | |
| "ref_id": "b3", | |
| "title": "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": [ | |
| "A" | |
| ], | |
| "last": "Bilmes", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2002, | |
| "venue": "Proc. International Conference on Spoken Language Processing (ICSLP)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "241--244", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "C.-P. Chen, K. Filaliy and J. A. Bilmes, ''Frontend post-processing and backend model enhancement on the Aurora 2.0/3.0 databases'', in Proc. International Conference on Spoken Language Processing (ICSLP), pp. 241-244, 2002", | |
| "links": null | |
| }, | |
| "BIBREF4": { | |
| "ref_id": "b4", | |
| "title": "Investigations on inter-speaker variability in the feature space'", | |
| "authors": [ | |
| { | |
| "first": "R", | |
| "middle": [], | |
| "last": "Haeb-Umbach", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1999, | |
| "venue": "Proc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "397--400", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "R. Haeb-Umbach, ''Investigations on inter-speaker variability in the feature space'', in Proc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 397-400, 1999", | |
| "links": null | |
| }, | |
| "BIBREF5": { | |
| "ref_id": "b5", | |
| "title": "Higher order cepstral moment normalization (HOCMN) for 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 (ICASSP)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "197--200", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "C.-W. Hsu and L.-S. Lee, ''Higher order cepstral moment normalization (HOCMN) for robust speech recognition'', in Proc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 197-200, 2004", | |
| "links": null | |
| }, | |
| "BIBREF6": { | |
| "ref_id": "b6", | |
| "title": "Cepstral gain normalization for noise robust speech recognition'", | |
| "authors": [ | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Yoshizawa", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "N", | |
| "middle": [], | |
| "last": "Hayasaka", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "W", | |
| "middle": [], | |
| "last": "Naoya", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Miyanaga", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2004, | |
| "venue": "Proc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "209--212", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "S. Yoshizawa, N. Hayasaka, W. Naoya, and Y. Miyanaga. ''Cepstral gain normalization for noise robust speech recognition'', in Proc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 209-212, 2004", | |
| "links": null | |
| }, | |
| "BIBREF7": { | |
| "ref_id": "b7", | |
| "title": "Quantile based histogram equalization for noise robust speech recognition", | |
| "authors": [ | |
| { | |
| "first": "F", | |
| "middle": [], | |
| "last": "Hilger", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Ney", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2001, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "1135--1138", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "F. Hilger and H. Ney, ''Quantile based histogram equalization for noise robust speech recognition'', European Conference on Speech Communication and Technology (Eurospeech), pp. 1135-1138, 2001", | |
| "links": null | |
| }, | |
| "BIBREF8": { | |
| "ref_id": "b8", | |
| "title": "Cepstral statistics compensation and normalization using online pseudo stereo codebooks for robust speech recognition in additive noise environments", | |
| "authors": [ | |
| { | |
| "first": "J.-W", | |
| "middle": [], | |
| "last": "Hung", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2008, | |
| "venue": "IEICE Trans. Information and Systems", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "296--311", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "J.-W. Hung, ''Cepstral statistics compensation and normalization using online pseudo stereo codebooks for robust speech recognition in additive noise environments'', IEICE Trans. Information and Systems, pp. 296-311, 2008", | |
| "links": null | |
| }, | |
| "BIBREF9": { | |
| "ref_id": "b9", | |
| "title": "Feature statistics compensation for robust speech recognition in additive noise environments", | |
| "authors": [ | |
| { | |
| "first": "T.-H", | |
| "middle": [], | |
| "last": "Hsieh", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2007, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "T.-H. Hsieh, \"Feature statistics compensation for robust speech recognition in additive noise environments\", M.S. thesis, National Chi Nan University, Taiwan, 2007", | |
| "links": null | |
| }, | |
| "BIBREF10": { | |
| "ref_id": "b10", | |
| "title": "Study of cepstral statistics normalization techniques for robust speech recognition in additive noise environments", | |
| "authors": [ | |
| { | |
| "first": "K.-C", | |
| "middle": [], | |
| "last": "Wu", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2008, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "K.-C. Wu, \"Study of cepstral statistics normalization techniques for robust speech recognition in additive noise environments\", M.S. thesis, National Chi Nan University, Taiwan, 2008", | |
| "links": null | |
| }, | |
| "BIBREF11": { | |
| "ref_id": "b11", | |
| "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": "Proc. 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\", in Proc. of ISCA IIWR ASR2000, Paris, France, 2000", | |
| "links": null | |
| } | |
| }, | |
| "ref_entries": { | |
| "FIGREF0": { | |
| "type_str": "figure", | |
| "uris": null, | |
| "num": null, | |
| "text": "ed reco g n iti on accu r acy (% ) \u4f75\u5408\u5f0f\u65b9\u6cd5\u5e7e\u4e4e\u7686\u6bd4\u6574\uf906\u5f0f\u3001\u5206\u6bb5\u5f0f\u8207\u78bc\u7c3f\u5f0f\u65b9\u6cd5\uf92d\u7684\u597d\uff0c\uf9b5\u5982\uff1aCU-HEQ (90.42%) \u8207 CS-HEQ (90.76%)\u512a\u65bc U-HEQ (87.62%)\u3001S-HEQ (85.9%)\u8207 C-HEQ (85.54%)\u3002\u9019 \u7d50\u679c\u8b49\u5be6\uf9ba\u6574\uf906\u5f0f\u3001\u5206\u6bb5\u5f0f\u8207\u78bc\u7c3f\u5f0f\u65b9\u6cd5\uff0c\u7368\u81ea\u6240\u63d0\u5347\u7684\u8fa8\uf9fc\uf961\u8f03\u5c0f\uff0c\u4f46\u7d50\u5408\uf9ba\u78bc \u7c3f\u5f0f\u8207\u6574\uf906\u5f0f(\u6216\u5206\u6bb5\u5f0f)\u7684\u4f75\u5408\u5f0f\u65b9\u6cd5\uff0c\u5c07\u4f7f\u8fa8\uf9fc\uf961\u5927\u5e45\u5730\u63d0\u5347\u3002 (3) \u5728\u4f75\u5408\u5f0f\u65b9\u6cd5\u4e2d\uff0c\u6211\u5011\u5c07 \u03b1 \u90fd\u8a2d\u70ba 0.5\uff0c\u9019\u8868\u793a\u8457\u78bc\u7c3f\u8207\u6574\u6bb5\u8a9e\uf906\u7684\u7d71\u8a08\u8cc7\u8a0a\u4e4b\u4f7f\u7528 \u6bd4\uf9b5\u76f8\u7b49\uff0c\u800c\u6c92\u6709\u4efb\u4f55\u504f\u5dee\u3002\u96d6\u7136 0.5 \u03b1 = \u672a\u5fc5\u662f\u6700\u4f73\u7684\u8a2d\u5b9a\uf96b\uf969\uff0c\u4f46\u81f3\u5c11\u4ee3\u8868\uf9ba\u6211 \u5011\u7121\u9808\u7cbe\u5fae\u5730\u6311\u9078\u6b64\uf96b\uf969\u503c\uff0c\uf965\u80fd\u5f97\u5230\u660e\u986f\u7684\u4f75\u5408\u6548\u76ca\u3002 (4) \u5728\uf978\u7a2e\u4f75\u5408\u5f0f\u65b9\u6cd5\u4e2d\uff0c\u78bc\u5b57\uf969\u76ee R=16 \u6240\u5c0d\u61c9\u7684\u8fa8\uf9fc\u7d50\u679c\u660e\u986f\u6bd4 R=256 \u6240\u5c0d\u61c9\u7684\u8fa8 \uf9fc\u7d50\u679c\u6709\u660e\u986f\u7684\u6539\u5584\uff0c\u6b64\u73fe\u8c61\u5728 CMS\u3001CMVN\u3001HOCMN\u3001CGN \u8207 HEQ \u7686\u662f\u5982\u6b64\uff0c \u6b64\u53ef\u80fd\u539f\u56e0\u5728\u65bc\u78bc\u7c3f\u8cc7\u8a0a\u8207\u6574\uf906\u8a9e\u97f3\u8cc7\u8a0a(\u6211\u5011\u5e0c\u671b\uf9dd\u7528\u4f75\u5408\u5f0f\u7279\u5fb5\u7d71\u8a08\u4f30\u6e2c\u6cd5\uff0c\u904b\u7528\u5728\u5176\u4ed6\u7684\u5012\u983b\u8b5c\u7d71\u8a08\u6b63\u898f\u5316\u6280\u8853\u4e0a\uff0c\uf9b5 \u5982\uff1a\u5012\u983b\u8b5c\u5f62\uf9fa\u6b63\u898f\u5316\u6cd5(CSN)\u6216\u5176\u4ed6\u968e\u5c64\u7684 HOCMN \u7b49\u6280\u8853\uff0c\u4ee5\u89c0\u5bdf\u5176\u6548\u80fd\u3002" | |
| }, | |
| "TABREF0": { | |
| "content": "<table><tr><td>\u5316\u6cd5(HOCMN)[4]\u3001\u5012\u983b\u8b5c\u589e\u76ca\u6b63\u898f\u5316\u6cd5(CGN)[5]\u4ee5\u53ca\u5012\u983b\u8b5c\u7d71\u8a08\u5716\u7b49\u5316\u6cd5(HEQ)[6]</td></tr><tr><td>\u7b49\uff0c\u9019\u4e9b\u6280\u8853\u6240\u9700\u4f7f\u7528\u7684\u7279\u5fb5\u7d71\u8a08\u76f8\u95dc\u8cc7\u8a0a\uff0c\uf9b5\u5982\uff1a\u5e73\u5747\u503c\u3001\u8b8a\uf962\uf969\u3001\u9ad8\u968e\u52d5\u5dee\u6216\u662f\u6a5f</td></tr><tr><td>\uf961\u5206\u4f48\u7b49\uff0c\u53ef\u7531\uf967\u540c\u7684\u65b9\u6cd5\u4f30\u6e2c\uff0c\u800c\u6709\uf967\u540c\u7684\u6548\u679c\u3002\u5728\u672c\u7ae0\u4e2d\uff0c\u6211\u5011\u5c07\u4ecb\u7d39\u904e\u53bb\u5b78\u8005\u6240</td></tr><tr><td>\u4e00\u3001\u7dd2\uf941</td></tr><tr><td>\u4e00\u8a9e\u97f3\u8fa8\uf9fc\u7cfb\u7d71\uff0c\u7576\u5176\u61c9\u7528\u65bc\u771f\u5be6\u74b0\u5883\u6642\uff0c\u5e38\u56e0\u74b0\u5883\u4e2d\u8af8\u591a\u7121\u6cd5\u9810\u671f\u7684\u8b8a\uf962\u6027</td></tr><tr><td>(variation)\uff0c\u800c\u4f7f\u5176\u8fa8\uf9fc\u6548\u80fd\u53d7\u5230\u660e\u986f\u5f71\u97ff\uff0c\u70ba\uf9ba\ufa09\u4f4e\u8af8\u591a\u7684\u8b8a\uf962\u6027\u6240\u767c\u5c55\u7684\u5404\u7a2e\u6280\u8853\uff0c</td></tr><tr><td>\u4e00\u822c\u800c\u8a00\u7d71\u7a31\u70ba\u5f37\u5065\u6027\u6280\u8853(robustness techniques)\uff0c\u800c\u672c\uf941\u6587\u4e2d\uff0c\u6211\u5011\u5247\u662f\u4e3b\u8981\u8457\u91cd\u65bc</td></tr><tr><td>\u767c\u5c55\ufa09\u4f4e\u74b0\u5883\u4e4b\u96dc\u8a0a\u5e72\u64fe\u6216\u901a\u9053\u6548\u61c9\u7684\u5f37\u5065\u6027\u6280\u8853\u3002</td></tr><tr><td>\u5728\u8af8\u591a\ufa09\u4f4e\uf93f\u97f3\u74b0\u5883\u4e4b\u96dc\u8a0a\u5e72\u64fe\u7684\u5f37\u5065\u6027\u6f14\u7b97\u6cd5\u4e2d\uff0c\u6709\u4e00\u5927\uf9d0\u7684\u65b9\u6cd5\u662f\u5c07\u8a13\uf996\u8207\u6e2c</td></tr><tr><td>\u8a66\u74b0\u5883\u4e0b\u7684\u8a9e\u97f3\u7279\u5fb5\u5176\u6642\u9593\u5e8f\uf99c\u7d71\u8a08\u7279\u6027\u52a0\u4ee5\u6b63\u898f\u5316(normalization)\uff0c\u4ee5\ufa09\u4f4e\u8a13\uf996\u8207\u6e2c</td></tr><tr><td>\u8a66\u74b0\u5883\u4e4b\u9593\u7684\uf967\u5339\u914d\uff0c\u9054\u5230\u63d0\u6607\u8fa8\uf9fc\uf961\u7684\u76ee\u7684\u3002\u5728\u9019\u4e9b\u6f14\u7b97\u6cd5\u4e2d\uff0c\u9996\u8981\u6b65\u9a5f\u901a\u5e38\u662f\u4f30\u6e2c</td></tr><tr><td>\u8a9e\u97f3\u7279\u5fb5\u7684\u7d71\u8a08\u503c\u76f8\u95dc\u8cc7\u8a0a\uff0c\uf9b5\u5982\u5728 CMVN \u6cd5\u4e2d\u6240\u9700\u4f30\u6e2c\u7684\u7d71\u8a08\u503c\u70ba\u5e73\u5747\u503c(mean)\u8207</td></tr><tr><td>\u8b8a\uf962\uf969(variance)\uff0c\u800c\u5728 HEQ \u6cd5\u4e2d\u5fc5\u9700\u4f30\u6e2c\u51fa\u7279\u5fb5\u6642\u9593\u5e8f\uf99c\u7684\u6a5f\uf961\u5206\u4f48(probability</td></tr><tr><td>distribution)\u3002\u9019\u4e9b\u7d71\u8a08\u4f30\u6e2c\u503c\u7684\u7cbe\u78ba\ufa01\uff0c\u76f4\u63a5\u5f71\u97ff\u5230\u5176\u5c0d\u61c9\u4e4b\u6b63\u898f\u5316\u6f14\u7b97\u6cd5\u7684\u6548\u80fd\u3002</td></tr><tr><td>\u5728\u904e\u53bb\u95dc\u65bc\u4e0a\u8ff0\u7279\u5fb5\u7d71\u8a08\u6b63\u898f\u5316\u6cd5\u7684\u6587\u737b\u4e2d\uff0c\u6839\u64da\uf967\u540c\u7684\u6a23\u672c\uf92d\u6e90\uff0c\u5927\u81f4\u4e0a\u6709\u4e09\u7a2e</td></tr><tr><td>\u7d71\u8a08\u503c\u4f30\u6e2c\u6cd5\uff0c\u5206\u5225\u70ba\u6574\uf906\u5f0f\u3001\u7247\u6bb5\u5f0f\u8207\u78bc\u7c3f\u5f0f\u7684\u4f30\u6e2c\u6cd5\uff0c\u9867\u540d\u601d\u7fa9\uff0c\u7b2c\u4e00\u7a2e\u76f4\u63a5\u4f7f\u7528</td></tr><tr><td>\uf9ba\u6574\uf906\u7684\u8a9e\u97f3\u7279\u5fb5\uf92d\u4f30\u6e2c\u7d71\u8a08\u503c\uff0c\u7b2c\u4e8c\u7a2e\u5247\u4f7f\u7528\uf9ba\u90e8\u5206(\u7247\u6bb5)\u7684\u8a9e\u97f3\u7279\u5fb5\uff0c\u800c\u7b2c\u4e09\u7a2e\u5247</td></tr><tr><td>\u9593\u63a5\u900f\u904e\u8a9e\u97f3\u7279\u5fb5\u5efa\uf9f7\u7684\u78bc\u7c3f[7]\uf92d\u4f5c\u7d71\u8a08\u503c\u4e4b\u4f30\u6e2c\u3002\u6211\u5011\u767c\u73fe\u9019\u4e09\u7a2e\u65b9\u6cd5\u5404\u6709\u5176\u512a\u7f3a</td></tr><tr><td>\u9ede\uff0c\u56e0\u6b64\u5728\u672c\uf941\u6587\u4e2d\uff0c\u6211\u5011\u6240\u63d0\u51fa\u7684\u65b0\u7d71\u8a08\u4f30\u6e2c\u6280\u8853\uff0c\u9069\u7576\u5730\u4f75\u5408\u78bc\u7c3f\u8207\u6574\uf906\u6216\u7247\u6bb5\u7684</td></tr><tr><td>\u7279\u5fb5\u8cc7\u8a0a\uff0c\u5e0c\u671b\u5f97\u5230\uf901\u7cbe\u6e96\u7684\u8a9e\u97f3\u7279\u5fb5\u7d71\u8a08\u503c\uff0c\u9032\u800c\u4f7f\u5404\u7a2e\u7279\u5fb5\u7d71\u8a08\u6b63\u898f\u5316\u6cd5\uff0c\u5728\u53d7\u96dc</td></tr><tr><td>\u8a0a\u5e72\u64fe\u7684\u74b0\u5883\u4e2d\u80fd\u5920\uf901\u6709\u6548\u5730\u63d0\u6607\u8a9e\u97f3\u7279\u5fb5\u7684\u5f37\u5065\u6027\uff0c\u4ee5\u6539\u5584\u8fa8\u8a8d\u7cbe\u78ba\ufa01\u3002</td></tr><tr><td>\u672c\uf941\u6587\u5176\u9918\u7684\u7ae0\u7bc0\u6982\u8981\u5982\u4e0b\uff1a\u5728\u7b2c\u4e8c\u7ae0\uff0c\u6211\u5011\u5c07\u7c21\u8981\u4ecb\u7d39\u904e\u53bb\u4e09\u7a2e\u7279\u5fb5\u7d71\u8a08\u503c\u4f30\u6e2c</td></tr><tr><td>\u6cd5\u4e4b\u6b65\u9a5f\u53ca\u5176\u53ef\u80fd\u7684\u512a\u7f3a\u9ede\u3002\u7b2c\u4e09\u7ae0\u5247\u4ecb\u7d39\u6211\u5011\u65b0\u63d0\u51fa\u7684\uf978\u7a2e\u4f75\u5408\u5f0f(hybrid-based)\u7684\u7d71</td></tr><tr><td>\u8a08\u503c\u4f30\u6e2c\u6cd5\uff0c\u53ca\u5176\u5982\u4f55\u904b\u7528\u65bc\u5404\u7a2e\u7279\u5fb5\u7d71\u8a08\u6b63\u898f\u5316\u6cd5\u4e2d\u3002\u5728\u7b2c\u56db\u7ae0\u4e2d\uff0c\u6211\u5011\u4ecb\u7d39\u8a9e\u97f3\u8fa8</td></tr><tr><td>\uf9fc\u5be6\u9a57\u4e4b\u8a9e\u97f3\u8cc7\uf9be\u5eab\u3001\u53ca\u65b0\u63d0\u51fa\u7684\uf978\u7a2e\u7d71\u8a08\u4f30\u6e2c\u6cd5\u5728\u5404\u7a2e\u7279\u5fb5\u7d71\u8a08\u6b63\u898f\u5316\u6cd5\u7684\u8a9e\u97f3\u8fa8\uf9fc</td></tr><tr><td>\u7d50\u679c\u53ca\u5176\u76f8\u95dc\u8a0e\uf941\u3002\u6700\u5f8c\uff0c\u7b2c\u4e94\u7ae0\u70ba\u4e00\u7c21\u8981\u7d50\uf941\u53ca\u672a\uf92d\u7814\u7a76\u4e4b\u5c55\u671b\u3002</td></tr><tr><td>\u4e8c\u3001\u6574\uf906\u5f0f\u3001\u5206\u6bb5\u5f0f\u8207\u78bc\u7c3f\u5f0f\u7279\u5fb5\u7d71\u8a08\u503c\u4f30\u6e2c\u6cd5</td></tr><tr><td>\u6211\u5011\u5728\u672c\uf941\u6587\u4e2d\u6240\u8a0e\uf941\u7684\u4e94\u7a2e\u8457\u540d\u5f37\u5065\u6027\u8a9e\u97f3\u7279\u5fb5\u6b63\u898f\u5316\u6280\u8853\uff0c\u5206\u5225\u70ba\u5012\u983b\u8b5c\u5e73\u5747</td></tr><tr><td>\u6d88\u53bb\u6cd5(CMS)[1]\u3001\u5012\u983b\u8b5c\u5e73\u5747\u503c\u8207\u8b8a\uf962\uf969\u6b63\u898f\u5316\u6cd5(CMVN)[2,3]\u3001\u9ad8\u968e\u5012\u983b\u8b5c\u52d5\u5dee\u6b63\u898f</td></tr></table>", | |
| "type_str": "table", | |
| "html": null, | |
| "text": "Keywords: speech recognition, codebook, feature statistics estimate, robust speech features", | |
| "num": null | |
| }, | |
| "TABREF5": { | |
| "content": "<table><tr><td>2\u3001\u5404\u7a2e\u5012\u983b\u8b5c\u5e73\u5747\u503c\u8207\u8b8a\uf962\uf969\u6b63\u898f\u5316\u6cd5\u4e4b\u5be6\u9a57\u7d50\u679c \u9020\u6210 Set C \u4e4b\u8fa8\uf9fc\u7d50\u679c\uf967\u76e1\uf9e4\u60f3\u3002 5\u3001\u5404\u7a2e\u7d71\u8a08\u5716\u7b49\u5316\u6cd5\u4e4b\u5be6\u9a57\u7d50\u679c</td></tr><tr><td>\u8868\u4e8c\u5448\u73fe\uf9ba\u5404\u7a2e\u5012\u983b\u8b5c\u5e73\u5747\u503c\u8207\u8b8a\uf962\uf969\u6b63\u898f\u5316\u6cd5(CMVN)\u7684\u8fa8\uf9fc\uf961\uff0c\u7531\u8868\u4e8c\u4e2d\uff0c\u9996 \u8868\u4e94\uf99c\u51fa\uf9ba\u5404\u7a2e\uf967\u540c\u578b\u614b\u7684\u7d71\u8a08\u5716\u7b49\u5316\u6cd5(HEQ)\u4e4b\u8fa8\uf9fc\uf961\uff0c\u5176\u7d50\u679c\u8207\u8868\u4e09\u8207\u8868\u56db\uf9d0</td></tr><tr><td>12)\u3002 \u5148\uff0c\u6211\u5011\u53ef\u4ee5\u770b\u51fa\u4e09\u7a2e\u50b3\u7d71\u7684\u7279\u5fb5\u7d71\u8a08\u503c\u4f30\u6e2c\u6cd5\u904b\u7528\u65bc CMVN \u6642\uff0c\u5176\u8fa8\uf9fc\u7d50\u679c\u76f8\u8f03\u65bc Method Set A Set B Set C Average RR \u4f3c\uff0c\u6211\u5011\u660e\u986f\u770b\u51fa\uff0c\u96d6\u7136\u78bc\u7c3f\u5f0f\u7684 HEQ(C-HEQ)\u6548\u679c\uf967\u76e1\uf9e4\u60f3\uff0c\u7136\u800c\u7576\u6211\u5011\u628a\u78bc\u7c3f\u8207\u6574 (\u4e8c) \u5404\u7a2e\u5f37\u5065\u6027\u8a9e\u97f3\u7279\u5fb5\u6b63\u898f\u5316\u6280\u8853\u4e4b\u5be6\u9a57\u7d50\u679c\u8207\u8a0e\uf941 \u57fa\u790e\u5be6\u9a57\u800c\u8a00\uff0c\u5728 U-CMVN\u3001S-CMVN \u8207 C-CMVN \u4e2d\uff0c\u4ee5 C-CMVN (R=256)\u6240\u5f97\u7684\u7e3d Baseline 71.92 68.22 77.61 71.58 -\uf906\u7279\u5fb5\u5408\u4f75\uff0c\u6216\u628a\u78bc\u7c3f\u8207\u7247\u6bb5\u7279\u5fb5\u5408\u4f75\uff0c\u6240\u5206\u5225\u5c0d\u61c9\u7684 CU-HEQ \u8207 CS-HEQ\uff0c\u5176\u5e36\uf92d \u672c\u7ae0\u5c07\u4ecb\u7d39\u6211\u5011\u6240\u63d0\u7684\u4e94\u7a2e\u5f37\u5065\u6027\u8a9e\u97f3\u7279\u5fb5\u6b63\u898f\u5316\u6280\u8853\u4e4b\u8fa8\uf9fc\u5be6\u9a57\u7d50\u679c(20dB\u3001 15dB\u300110dB\u30015dB \u8207 0dB \u4e94\u7a2e\u8a0a\u96dc\u6bd4\u4e0b\u7684\u8fa8\uf9fc\uf961\u5e73\u5747)\uff0c\u5206\u5225\u70ba CMS\u3001CMVN\u3001HOCMN\u3001 CGN \u4ee5\u53ca HEQ\uff0c\u800c\u9019\u4e9b\u65b9\u6cd5\u7684\u7279\u5fb5\u7d71\u8a08\u76f8\u95dc\u8cc7\u8a0a\uff0c\u5206\u5225\u4f7f\u7528\u4e09\u7a2e\u7279\u5fb5\u7d71\u8a08\u503c\u4f30\u6e2c\u6cd5(\u6574 \uf906\u5f0f\u3001\u5206\u6bb5\u5f0f\u8207\u78bc\u7c3f\u5f0f)\u4ee5\u53ca\uf978\u7a2e\u4f75\u5408\u5f0f(hybrid-based)\u7279\u5fb5\u7d71\u8a08\u503c\u4f30\u6e2c\u6cd5(\u78bc\u7c3f/\u6574\uf906\u5f0f\u8207 \u78bc\u7c3f/\u5206\u6bb5\u5f0f)\uf92d\u6c42\u5f97\uff0c\u7136\u5f8c\u9032\u4e00\u6b65\u6bd4\u8f03\u3001\u8a0e\uf941\u8207\u5206\u6790\u3002 \u5728\u672c\uf941\u6587\u4e2d\uff0c\u6240\u6709\u7247\u6bb5\u5f0f\u7d71\u8a08\u4f30\u6e2c\u6cd5\u5be6\u9a57\u4e2d\u7684\u7247\u6bb5\u9577\ufa01 2 1 L + \u90fd\u8a2d\u70ba 101(\u9664\uf9ba HOCMN \u4e4b\u5916\uff0c\u5b83\u7684\u7247\u6bb5\u9577\ufa01 2 1 L + \u8a2d\u70ba 87) \uff1b\u6240\u6709\u78bc\u7c3f\u5f0f\u8207\u4f75\u5408\u5f0f\u5be6\u9a57\u4e2d\u7684\u78bc\u5b57\uf969\u76ee R \u7d71\u4e00\u8a2d\u5b9a\u70ba 16 \u6216 256\uff0c\u800c\u4f75\u5408\u5f0f\u5be6\u9a57\u4e2d\u7684 \u03b1 \uff0c\u6211\u5011\u56fa\u5b9a\u8a2d\u70ba 0.5\uff0c\u4f7f\u4f75\u5408\u4e4b\u96d9\u65b9\u7d71\u8a08\u8cc7 \u8a0a\u6240\u4f54\u7684\u6bd4\uf9b5\u76f8\u7b49\u3002 \u5e73\u5747\u8fa8\uf9fc\uf961\u6700\u9ad8\uff0c\u6bd4\u57fa\u790e\u5be6\u9a57\u7d50\u679c\u63d0\u5347\uf9ba 15.14%\uff0c\u800c\u76f8\u5c0d\u932f\u8aa4\uf961\ufa09\u4f4e\uf961\u9054\u5230\uf9ba 53.27%\u3002 \u63a5\u4e0b\uf92d\uff0c\u672c\uf941\u6587\u63d0\u51fa\uf978\u7a2e\u4f75\u5408\u5f0f\u7279\u5fb5\u7d71\u8a08\u503c\u4f30\u6e2c\u6cd5\u904b\u7528\u5728 CMVN \u4e0a\uff0c\u540c\u6a23\u5176\u8fa8\uf9fc \u7d50\u679c\u76f8\u8f03\u65bc\u57fa\u790e\u5be6\u9a57\u4e5f\u6709\u660e\u986f\u7684\u9032\u6b65\u3002\u7531\u8868\u4e8c\u6240\u793a\uff0c\u5728 CU-CMVN \u8207 CS-CMVN \u4e2d\uff0c C-HOCMN (R=16) 84.86 83.40 85.68 84.44 \u7684 \u8fa8 \uf9fc \uf961 \u63d0 \u6607 \u7a0b \ufa01 \u7686 \u5341 \u5206 \u986f \u8457 \uff0c \u660e \u986f \u8d85 \u8d8a \uf9ba \u6574 \uf906 \u5f0f HEQ(U-HEQ) \u8207 \u7247 \u6bb5 \u5f0f 45.25 C-HOCMN (R=256) 86.30 86.34 83.53 85.76 HEQ(S-HEQ)\uff0c\u800c\u5176\u4e2d\u53c8\u4ee5 CS-HEQ \u8868\u73fe\u6700\u4f73\u3002\u5e73\u5747\u8fa8\uf9fc\uf961\u9ad8\u9054 90.76%\uff0c\u76f8\u5c0d\u932f\u8aa4\ufa09 49.89 U-HOCMN 87.43 88.54 87.52 87.89 \u4f4e\uf961\u70ba 67.49%\u3002 57.39 \u4ee5 CS-CMVN (R=16)\u7684\u7e3d\u5e73\u5747\u8fa8\uf9fc\uf961\u6700\u9ad8\uff0c\u6bd4\u57fa\u790e\u5be6\u9a57\u7d50\u679c\u63d0\u5347\uf9ba 17.38%\uff0c\u800c\u76f8\u5c0d\u932f CU-HOCMN (R=16) 87.55 88.85 85.57 87.67* 56.62 method Set A Set B Set C Average RR \u8aa4\uf961\ufa09\u4f4e\uf961\u9ad8\u9054 61.15%\u3002\u56e0\u6b64\uf978\u7a2e\u4f75\u5408\u5f0f CMVN \u5728\u8a9e\u97f3\u7279\u5fb5\u5f37\u5065\u6027\u65b9\u9762\uff0c\u8ddf\u524d\u4e00\u7bc0\u7684 CU-HOCMN (R=256) 86.66 88.15 85.12 86.95 54.08 Baseline 71.92 68.22 77.61 71.58 -\u4f75\u5408\u5f0f CMS \u4e00\u6a23\uff0c\u7686\u512a\u65bc U-CMVN\u3001S-CMVN \u8207 C-CMVN\u3002 Method Set A Set B Set C Average S-HOCMN 85.60 86.63 86.16 86.12 51.16 C-HEQ (R=16) 80.15 80.41 76.03 79.43 27.62 RR Baseline 71.92 68.22 77.61 71.58 CS-HOCMN (R=16) 89.17 90.26 87.96 89.36 62.56 C-HEQ (R=256) 86.23 85.77 83.71 85.54 49.12 -CS-HOCMN (R=256) 87.44 88.78 86.22 87.73 56.83 U-HEQ 86.95 88.39 87.40 87.62 56.44 1\u3001\u5404\u7a2e\u5012\u983b\u8b5c\u5e73\u5747\u6d88\u53bb\u6cd5\u4e4b\u5be6\u9a57\u7d50\u679c C-CMVN (R=16) 85.75 85.5 83.78 85.26 48.14 \u8868\u4e09\u3001\u5404\u7a2e HOCMN \u7684\u6574\u9ad4\u5e73\u5747\u8fa8\uf9fc\uf961\u8207\u76f8\u5c0d\u932f\u8aa4\ufa09\u4f4e\uf961(relative error rate reduction, RR) CU-HEQ (R=16) 90.21 91.16 89.37 90.42 66.29 \u5f9e\u8868\u4e00\u4e2d\uff0c\u6211\u5011\u5f97\u77e5\u5404\u7a2e CMS \u7684\u7e3d\u5e73\u5747\u8fa8\uf9fc\uf961\u60c5\u5f62\u3002\u6211\u5011\u5148\u63a2\u8a0e\u4e09\u7a2e\u50b3\u7d71\u7684\u7279\u5fb5 C-CMVN (R=256) 87.16 87.44 84.39 86.72 53.27 \u4e4b\u6bd4\u8f03 CU-HEQ (R=256) 88.76 89.68 87.85 88.95 61.12 \u7d71\u8a08\u503c\u4f30\u6e2c\u6cd5\u4f5c\u7528\u5728 CMS \u4e0a\u7684\u6548\u679c\uff0c\u767c\u73fe U-CMS\u3001S-CMS \u8207 C-CMS \u4e2d\uff0c\u4ee5 C-CMS (R=256)\u6240\u5f97\u7684\u6574\u9ad4\u5e73\u5747\u8fa8\uf9fc\uf961\u6700\u5927\uff0c\u6bd4\u57fa\u790e\u5be6\u9a57\u7d50\u679c\u63d0\u5347\uf9ba 10.75%\uff0c\u800c\u76f8\u5c0d\u932f\u8aa4\uf961\u8870 \u6e1b\uf961(RR)\u70ba 37.83%\u3002 \u63a5\u8457\uff0c\u6211\u5011\u53ef\u770b\u51fa\u672c\uf941\u6587\u65b0\u63d0\u51fa\u7684\uf978\u7a2e\u4f75\u5408\u5f0f\u7279\u5fb5\u7d71\u8a08\u503c\u4f30\u6e2c\u6cd5\u904b\u7528\u5728 CMS \u4e0a\u4e4b \u6548\u679c\uff0c\u5982\u4ee5\u4e0b\u5e7e\u9ede\u6240\u8ff0\uff1a (1) CU-CMS \u8207 CS-CMS \u7686\u660e\u986f\u512a\u65bc\u50b3\u7d71\u4e4b U-CMS\u3001S-CMS \u8207 C-CMS\u3002 (3)\u5728\u5404\u7a2e\uf967\u540c\u578b\u614b\u7684 CMS \u4e2d\uff0c\u4ee5 CS-CMS (R=16)\u7684\u7e3d\u5e73\u5747\u8fa8\uf9fc\u7d50\u679c\u6700\u4f73\uff0c\u800c\u5176\u5b83 CMS \u7684\u7e3d\u5e73\u5747\u8fa8\uf9fc\u7d50\u679c\u7684\u512a\uf99d\u9806\u5e8f\uff0c\u4f9d\u5e8f\u70ba\uff1aCS-CMS (R=256)\u3001CU-CMS (R=16)\u3001CU-CMS \u65b0\u63d0\u51fa\u4e4b\uf978\u7a2e\u4f75\u5408\u5f0f CMS \u5728\u63d0\u6607\u8a9e\u97f3\u7279\u5fb5\u5f37\u5065\u6027\u4e0a\uff0c\u6bd4 U-CMS\u3001S-CMS \u8207 C-CMS \u9084 \u8981\uf92d\u7684\u512a\u8d8a\u3002 \u8868\u4e09\u5448\u73fe\uf9ba\u5404\u7a2e\u9ad8\u968e\u5012\u983b\u8b5c\u52d5\u5dee\u6b63\u898f\u5316\u6cd5(HOCMN)\u7684\u8fa8\uf9fc\uf961\uff0c\u5728\u9019\uf9e8\uff0c\u4e2d\u592e\u52d5\u5dee (R=256)\uf976\u4f4e\u65bc U-CGN \u4e4b\u5916\uff0c\u7686\u512a\u65bc U-CGN\u3001S-CGN \u8207 C-CGN\u3002 3\u3001\u5404\u7a2e\u9ad8\u968e\u5012\u983b\u8b5c\u52d5\u5dee\u6b63\u898f\u5316\u6cd5\u4e4b\u5be6\u9a57\u7d50\u679c 17.88%\uff0c\u800c\u76f8\u5c0d\u932f\u8aa4\ufa09\u4f4e\uf961\u9ad8\u9054\uf9ba 62.91%\uff0c\u6211\u5011\u8b49\u5be6\uf9ba\u5728\u4f75\u5408\u5f0f CGN \u4e2d\uff0c\u9664\uf9ba CU-CGN (R=256)\u3001C-CMS (R=256)\u3001C-CMS(R=16)\u3001U-CMS \u4ee5\u53ca S-CMS\u3002\u56e0\u6b64\uff0c\u6211\u5011\u9a57\u8b49\uf9ba\u6240 \u5341\u5206\u986f\u8457\u7684\u63d0\u6607\uff0c\u5176\u4e2d\u4ee5 CS-CGN(R=16)\u7684\u7e3d\u5e73\u5747\u8fa8\uf9fc\uf961\u6700\u9ad8\uff0c\u6bd4\u57fa\u790e\u5be6\u9a57\u63d0\u5347\uf9ba \u4e4b\u6bd4\u8f03 \u7136\u800c\uff0c\u7576\uf978\u7a2e\u4f75\u5408\u5f0f\u7279\u5fb5\u7d71\u8a08\u503c\u4f30\u6e2c\u6cd5\uff0c\u5206\u5225\u4f7f\u7528\u5728 CGN \u6642\uff0c\u5176\u8fa8\uf9fc\u7d50\u679c\u90fd\u80fd\u6709 \u8868\u4e8c\u3001\u5404\u7a2e CMVN \u7684\u6574\u9ad4\u5e73\u5747\u8fa8\uf9fc\uf961\u8207\u76f8\u5c0d\u932f\u8aa4\ufa09\u4f4e\uf961(relative error rate reduction, RR) \u5e72\u64fe\u3002 14.61%\uff0c\u800c\u76f8\u5c0d\u932f\u8aa4\uf961\u8870\u6e1b\uf961\u9ad8\u9054 51.41%\u3002 CS-CMVN (R=256) 88.18 89.09 86.73 88.25 58.66 \u7121\u6cd5\u8f03\u7cbe\u78ba\u5730\u4f30\u6e2c CGN \u6240\u9700\u7528\u5230\u7684\u52d5\u614b\u7bc4\u570d\u503c\uff0c\u4ee5\u53ca\u5176\u672a\u8003\u616e\u5230 Set C \u7684\u647a\u7a4d\u6027\u96dc\u8a0a (2) CS-CMS(R=16)\u8868\u73fe\u512a\u65bc CU-CMS(R=16)\uff0c\u76f8\u5c0d\u65bc\u57fa\u790e\u5be6\u9a57\u7d50\u679c\uff0c\u5728\u8fa8\uf9fc\uf961\u4e0a\u63d0\u5347\uf9ba U-CMVN 85.03 85.56 85.61 85.36 48.49 CU-CMVN (R=16) 87.87 88.67 86.14 87.84 57.21 CU-CMVN (R=256) 87.25 88.06 85.78 87.28 55.24 S-CMVN 83.99 84.85 84.78 84.49 CS-CMVN (R=16) 88.98 89.82 87.19 88.96 \u70ba 57.46%\uff0c\u76f8\u5c0d\u800c\u8a00\uff0cC-CGN \u8868\u73fe\u8f03\u5dee\uff0c\u6b64\u53ef\u80fd\u539f\u56e0\u8ddf\u4e0a\u4e00\u7bc0\u6240\u8ff0\uf9d0\u4f3c\uff0c\u5373\u78bc\u7c3f\u53ef\u80fd 61.15 \u6240\u5f97\u7684\u7e3d\u5e73\u5747\u8fa8\uf9fc\uf961\u6700\u9ad8\uff0c\u8207\u57fa\u790e\u5be6\u9a57\u7d50\u679c\u76f8\u6bd4\uff0c\u63d0\u5347\uf9ba 16.33%\uff0c\u800c\u76f8\u5c0d\u932f\u8aa4\ufa09\u4f4e\uf961 45.43 S-HEQ 85.10 86.82 85.64 85.90 50.39 4\u3001\u5404\u7a2e\u5012\u983b\u8b5c\u589e\u76ca\u6b63\u898f\u5316\u6cd5\u4e4b\u5be6\u9a57\u7d50\u679c CS-HEQ (R=16) 90.57 91.54 89.57 90.76 67.49 \u8868\u56db\uf99c\u51fa\uf9ba\u5404\u7a2e\u5012\u983b\u8b5c\u589e\u76ca\u6b63\u898f\u5316\u6cd5(CGN)\u7684\u8fa8\uf9fc\uf961\uff0c\uf974\u8207\u8868\u4e09\u76f8\u6bd4\u8f03\uff0c\u6211\u5011\u767c\u73fe \u5b83\u5011\u7684\u7d50\u679c\u5341\u5206\uf9d0\u4f3c\uff0c\u4e09\u7a2e\u50b3\u7d71\u4f30\u6e2c\u6cd5\u6240\u5c0d\u61c9\u7684 U-CGN\u3001S-CGN \u8207 C-CGN \u4e2d\uff0c\u4ee5 U-CGN CS-HEQ (R=256) 89.11 90.21 88.35 89.40 62.70</td></tr><tr><td>\u56db\u3001\u5be6\u9a57\u74b0\u5883\u8a2d\u5b9a\u8207\u5404\u7a2e\u5f37\u5065\u6027\u8a9e\u97f3\u7279\u5fb5\u6b63\u898f\u5316\u6280\u8853\u4e4b\u5be6\u9a57\u7d50\u679c\u8207\u8a0e\uf941 (\u4e00) \u5be6\u9a57\u74b0\u5883\u8a2d\u5b9a \u8cc7\uf9be\u5eab\u6709\uf978\u7a2e\uf967\u540c\u7684\u8a13\uf996\u74b0\u5883\uff1a\u4e7e\u6de8\u74b0\u5883(clean-condition)\u8207\u591a\u91cd\u74b0\u5883(multi-condition)\u4ee5 \u53ca\u4e09\u7a2e\uf967\u540c\u7684\u6e2c\u8a66\u96c6\u5408\uff1aA \u7d44(\u5730\u4e0b\u9435\u3001\u4eba\u8072\u3001\u6c7d\uf902\u548c\u5c55\u89bd\u9928\u96dc\u8a0a)\u3001B \u7d44(\u9910\u5ef3\u3001\u8857\u9053\u3001 \u6a5f\u5834\u548c\u706b\uf902\u7ad9\u96dc\u8a0a)\u8207 C \u7d44(\u5730\u4e0b\u9435\u3001\u8857\u9053\u96dc\u8a0a\u5916\u52a0 MIRS \u901a\u9053\u6548\u61c9)\u96dc\u8a0a\u8a9e\u97f3\u96c6\u5408\u3002\u4e7e \u6de8\u74b0\u5883\u4ee3\u8868\u6c92\u6709\u4efb\u4f55\u96dc\u8a0a\u7684\u8a9e\u97f3\u74b0\u5883\uff0c\u800c\u591a\u91cd\u74b0\u5883\u5247\u4ee3\u8868\u9069\u7576\u52a0\u5165\u5404\u7a2e\u9644\u52a0\u96dc\u8a0a\u7684\u8a9e\u97f3 \u5728\u9019\uf9e8\uff0c\u57fa\u790e\u5be6\u9a57(baseline experiment)\u5c07\u63a1\u7528\u672a\u8655\uf9e4\u7684\u6885\u723e\u5012\u983b\u8b5c\u7279\u5fb5\u4fc2\uf969(MFCC) \u4f5c\u70ba\u8a13\uf996\u8ddf\u6e2c\u8a66\uff0c\u6240\u4f7f\u7528\u7684MFCC\u7279\u5fb5\uf96b\uf969\u70ba13\u7dad(c 0~c12 )\uff0c\u518d\u52a0\u4e0a\u5176\u4e00\u968e\u548c\u4e8c\u968e\u5dee\uf97e\uff0c \u4e4b\u6bd4\u8f03 (convolutional noise)\uff0c\u800c\u78bc\u7c3f(codebook)\u4e2d\u53ea\u8003\u616e\u5230\u52a0\u6210\u6027\u96dc\u8a0a(additive noise)\uff0c\u6240\u4ee5 \u8072\u5b78\u6a21\u578b\u70ba\u7531\u5de6\u5411\u53f3(left-to-right)\u4e4b\u96b1\u85cf\u5f0f\u99ac\u53ef\u592b\u6a21\u578b(hidden Markov model, HMM) \u8868\u4e00\u3001\u5404\u7a2e CMS \u7684\u6574\u9ad4\u5e73\u5747\u8fa8\uf9fc\uf961\u8207\u76f8\u5c0d\u932f\u8aa4\ufa09\u4f4e\uf961(relative error rate reduction, RR) \u7e3d\u5e73\u5747\u8fa8\uf9fc\uf961\uf967\u53ca U-HOCMN\uff0c\u6b64\u73fe\u8c61\u53ef\u6b78\u56e0\u65bc Set C \u4e2d\u7684\u8a9e\u97f3\u5305\u542b\uf9ba\u647a\u7a4d\u6027\u96dc\u8a0a \u7e3d\u5171\u670939\u7dad\u7279\u5fb5\uf96b\uf969\u4f5c\u70ba\u6700\u7d42\u4f7f\u7528\u4e4b\u7279\u5fb5\uf96b\uf969\u5411\uf97e\u3002 CS-CMS (R=16) 85.38 87.43 85.31 86.19 CS-CMS (R=256) 84.63 86.98 84.42 85.53 \u73fe\u7684\u7d50\u679c\u3002\uf901\u9032\u4e00\u6b65\u89c0\u5bdf\uff0c\u53ef\u770b\u51fa CU-HOCMN \u5728 Set C \u7684\u8fa8\uf9fc\uf961\u76f8\u5c0d\u8f03\u4f4e\uff0c\u9020\u6210 49.09 \u5f0f HOCMN \u7686\u4f4e\u65bc U-HOCMN\uff0c\uf967\u540c\u65bc\u524d\u9762\u5e7e\u7bc0\u6240\u8ff0\u7684\u4f75\u5408\u5f0f CMS \u8207 CMVN \u6240\u5448 51.41 (3) \u5728\u5404\u7a2e\u578b\u614b\u7684 HOCMN \u4e2d\uff0c\u4ee5 CS-HOCMN (R=16)\u7684\u7e3d\u5e73\u5747\u8fa8\uf9fc\uf961\u6700\u4f73\uff0c\u5176\u5b83\u4f75\u5408 \u96c6\u5408\u52a0\u4ee5\u8fa8\uf9fc\u3002 S-CMS 77.28 80.66 77.63 78.70 25.05 62.56%\u3002 \u74b0\u5883\u3002\u672c\uf941\u6587\u7684\u5be6\u9a57\u53ea\u63a1\u7528\u4e7e\u6de8\u74b0\u5883\u7684\u8a9e\u97f3\u7279\u5fb5\u4f5c\u8072\u5b78\u6a21\u578b\u7684\u8a13\uf996\uff0c\u4e26\u5c0d\u4e09\u7d44\u96dc\u8a0a\u8a9e\u97f3 C-CMS (R=256) 81.62 81.58 85.25 82.33 U-CMS 79.35 82.46 79.91 80.71 CU-CMS (R=16) 83.28 84.92 84.28 84.14 CU-CMS (R=256) 82.13 84.29 83.06 83.18 \u6bd4\u57fa\u790e\u5be6\u9a57\u7d50\u679c\u63d0\u5347\uf9ba 16.09%\u8207 17.78%\uff0c\u76f8\u5c0d\u932f\u8aa4\ufa09\u4f4e\uf961\u8870\u6e1b\u5206\u5225\u9ad8\u9054 56.62%\u8207 40.82 \u76f8\u8f03\u65bc\u57fa\u790e\u5be6\u9a57\u7d50\u679c\u4ecd\u6709\u660e\u986f\u6539\u5584\uff1aCU-HOCMN(R=16)\u8207 CS-HOCMN(R=16)\u5206\u5225 44.19 (2) \u672c\uf941\u6587\u63d0\u51fa\uf978\u7a2e\u4f75\u5408\u5f0f\u7279\u5fb5\u7d71\u8a08\u503c\u4f30\u6e2c\u6cd5\uff0c\u7576\u5176\u904b\u7528\u5728 HOCMN \u4e0a\u6642\uff0c\u5176\u8fa8\uf9fc\u7d50\u679c 32.13 \u8f03\u70ba\u6e96\u78ba\uff0c\u4f46\u7121\u6cd5\u6709\u6548\u4f30\u6e2c\u8f03\u9ad8\u968e\u7684\u52d5\u5dee\u503c\u3002 37.83 \uf967\uf9e4\u60f3\uff0c\u5176\u53ef\u80fd\u539f\u56e0\u70ba\uff0c\u78bc\u7c3f\u5c0d\u65bc\u8f03\u4f4e\u968e\u7684\u52d5\u5dee\u503c(\uf9b5\u5982\u5e73\u5747\u503c\u8207\u8b8a\uf962\uf969)\u4e4b\u4f30\u6e2c \u6240\u767c\ufa08\u7684 AURORA2 \u8a9e\u97f3\u8cc7\uf9be\u5eab[10]\uff0c\u5176\u5167\u5bb9\u662f\u7531\uf99a\u7e8c\u7684\u82f1\u6587\uf969\u5b57\u5b57\uf905\u6240\u69cb\u6210\u3002\u6b64\u8a9e\u97f3 C-CMS (R=16) 80.83 79.29 86.13 81.27 34.10 \u6b64\u73fe\u8c61\u8ddf\u524d\uf978\u7bc0\u6240\u5448\u73fe\u7684\u7d50\u679c\u4e26\uf967\u76f8\u540c\uff0c\u56e0\u70ba\u78bc\u7c3f\u5f0f\u7684 HOCMN(C-HOCMN)\u8868\u73fe\u4e26 \u672c\uf941\u6587\u63a1\u7528\u6b50\u6d32\u96fb\u4fe1\u6a19\u6e96\u5354\u6703(European Telecommunication Standard Institute, ETSI) method Set A Set B Set C Average baseline 71.92 68.22 77.61 71.58 \uf9fc\uf961\u6700\u5927\uff0c\u8207\u57fa\u790e\u5be6\u9a57\u76f8\u6bd4\uff0c\u63d0\u5347\uf9ba 16.31%\uff0c\u800c\u76f8\u5c0d\u932f\u8aa4\ufa09\u4f4e\uf961\u9054\u5230\uf9ba 57.39%\u3002 -(1) \u5728\u50b3\u7d71\u4e4b U-HOCMN\u3001S-HOCMN \u8207 C-HOCMN \u4e2d\uff0c\u4ee5 U-HOCMN \u6240\u5f97\u7684\u7e3d\u5e73\u5747\u8fa8 RR \u7684\u968e\uf969J \u7686\u8a2d\u70ba 100\u3002\u5f9e\u8868\u4e09\u53ef\u770b\u51fa\u4ee5\u4e0b\u5e7e\u9ede\u73fe\u8c61\uff1a method Set A Set B Set C Average RR</td></tr></table>", | |
| "type_str": "table", | |
| "html": null, | |
| "text": "\u7684\u5f62\u5f0f\uff0c\u662f\u4f7f\u7528\u96b1\u85cf\u5f0f\u99ac\u53ef\u592b\u6a21\u578b\u8a13\uf996\u8edf\u9ad4 HTK[11]\u8a13\uf996\u6240\u5f97\uff0c\u5176\u4e2d\u5305\u542b 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\u5305\u542b 16 \u500b\uf9fa\u614b\uff0c \u800c\u6bcf\u500b\uf9fa\u614b\u5247\u5305\u542b 20 \u500b\u9ad8\u65af\u5bc6\ufa01\u6df7\u5408\u3002 \u8868\u4e94\u3001\u5404\u7a2e HEQ \u7684\u6574\u9ad4\u5e73\u5747\u8fa8\uf9fc\uf961\u8207\u76f8\u5c0d\u932f\u8aa4\ufa09\u4f4e\uf961(relative error rate reduction, RR)", | |
| "num": null | |
| } | |
| } | |
| } | |
| } |