{ "paper_id": "O15-1014", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T08:10:13.549012Z" }, "title": "", "authors": [ { "first": "Chia-Yung", "middle": [], "last": "\u5f90\u5bb6\u93de", "suffix": "", "affiliation": {}, "email": "" }, { "first": "", "middle": [], "last": "Hsu", "suffix": "", "affiliation": { "laboratory": "", "institution": "National Central University", "location": {} }, "email": "" }, { "first": "Jia-Ching", "middle": [], "last": "\u3001\u738b\u5bb6\u6176", "suffix": "", "affiliation": {}, "email": "" }, { "first": "", "middle": [], "last": "Wang", "suffix": "", "affiliation": { "laboratory": "", "institution": "National Central University", "location": {} }, "email": "" }, { "first": "Yu", "middle": [], "last": "\u3001\u66f9\u6631", "suffix": "", "affiliation": {}, "email": "" }, { "first": "", "middle": [], "last": "Tsao", "suffix": "", "affiliation": { "laboratory": "\u4e2d\u592e\u7814\u7a76\u9662\u8cc7\u8a0a\u79d1\u6280\u5275\u65b0\u7814\u7a76\u4e2d\u5fc3 Research Center for Information Technology Innovation", "institution": "Academia Sinica", "location": {} }, "email": "" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "\u8fd1\u5e74\u4f86\uff0c\u985e\u795e\u7d93\u7db2\u8def (Neural Network) \u5728\u8a9e\u97f3\u8fa8\u8b58\u4e0a\u7684\u7814\u7a76\u6709\u8457\u8c50\u78a9\u7684\u6210\u679c\uff0c\u6709 \u6548\u5730\u6e1b\u5c11\u74b0\u5883\u4ee5\u53ca\u8a9e\u8005\u8b8a\u7570\u5c0d\u8a9e\u97f3\u8a0a\u865f\u9020\u6210\u7684\u5f71\u97ff\uff0c\u5927\u5e45\u63d0\u5347\u8fa8\u8b58\u7387\uff0c\u4f46\u7cfb\u7d71\u7684\u8a9e\u97f3\u8fa8 \u8b58\u80fd\u529b\u4ecd\u6709\u6539\u5584\u7a7a\u9593\u3002\u672c\u8ad6\u6587\u5373\u63d0\u51fa\u65b0\u7684\u81ea\u52d5\u8a9e\u97f3\u8fa8\u8b58\u7cfb\u7d71\u67b6\u69cb\uff0c\u7d50\u5408 Environment Clustering (EC)\u3001Mixture of Experts \u8207\u985e\u795e\u7d93\u7db2\u8def\u4ee5\u9032\u4e00\u6b65\u63d0\u5347\u7cfb\u7d71\u6548\u80fd\u3002\u6211\u5011\u5c07\u8fa8\u8b58 \u7cfb\u7d71\u5206\u70ba Offline \u8207 Online \u5169\u968e\u6bb5\uff1aOffline \u968e\u6bb5\u4f9d\u64da\u8072\u5b78\u7279\u6027\u5c07\u6574\u500b\u8a13\u7df4\u8cc7\u6599\u96c6\u5206\u5272\u6210 \u591a\u500b\u5b50\u8a13\u7df4\u8cc7\u6599\u96c6\uff0c\u4e26\u5efa\u7acb\u5404\u5b50\u8a13\u7df4\u8cc7\u6599\u96c6\u7684\u985e\u795e\u7d93\u7db2\u8def(\u4ee5\u985e\u795e\u7d93\u5b50\u7db2\u8def\u7a31\u4e4b)\u3002Online \u968e\u6bb5\u5247\u4f7f\u7528 GMM-gate \u4f86\u63a7\u5236\u985e\u795e\u7d93\u5b50\u7db2\u8def\u7684\u8f38\u51fa\u3002\u65b0\u63d0\u51fa\u7684\u7cfb\u7d71\u67b6\u69cb\u4fdd\u7559\u5b50\u8a13\u7df4\u8cc7\u6599 \u96c6\u7684\u8072\u5b78\u7279\u6027\uff0c\u5f37\u5065\u8a9e\u97f3\u8fa8\u8b58\u7cfb\u7d71\u3002\u5be6\u9a57\u4e0a\uff0c\u6211\u5011\u4f7f\u7528 Aurora 2 \u9023\u7e8c\u6578\u5b57\u8a9e\u97f3\u8cc7\u6599\u5eab\uff0c \u4f9d\u64da\u5b57\u932f\u8aa4\u7387(word error rate, WER)\u6bd4\u8f03\u6211\u5011\u63d0\u51fa\u7684\u8a9e\u97f3\u8fa8\u8b58\u7cfb\u7d71\u67b6\u69cb\u8207\u50b3\u7d71\u4ee5\u985e\u795e\u7d93 \u7db2\u8def\u5efa\u7acb\u7684\u8fa8\u8b58\u7cfb\u7d71\uff0c\u5e73\u5747\u5b57\u932f\u8aa4\u7387\u9032\u6b65 5.9% \uff0c\u7531 5.25%\u964d\u4f4e\u81f3 4.94%\u3002", "pdf_parse": { "paper_id": "O15-1014", "_pdf_hash": "", "abstract": [ { "text": "\u8fd1\u5e74\u4f86\uff0c\u985e\u795e\u7d93\u7db2\u8def (Neural Network) \u5728\u8a9e\u97f3\u8fa8\u8b58\u4e0a\u7684\u7814\u7a76\u6709\u8457\u8c50\u78a9\u7684\u6210\u679c\uff0c\u6709 \u6548\u5730\u6e1b\u5c11\u74b0\u5883\u4ee5\u53ca\u8a9e\u8005\u8b8a\u7570\u5c0d\u8a9e\u97f3\u8a0a\u865f\u9020\u6210\u7684\u5f71\u97ff\uff0c\u5927\u5e45\u63d0\u5347\u8fa8\u8b58\u7387\uff0c\u4f46\u7cfb\u7d71\u7684\u8a9e\u97f3\u8fa8 \u8b58\u80fd\u529b\u4ecd\u6709\u6539\u5584\u7a7a\u9593\u3002\u672c\u8ad6\u6587\u5373\u63d0\u51fa\u65b0\u7684\u81ea\u52d5\u8a9e\u97f3\u8fa8\u8b58\u7cfb\u7d71\u67b6\u69cb\uff0c\u7d50\u5408 Environment Clustering (EC)\u3001Mixture of Experts \u8207\u985e\u795e\u7d93\u7db2\u8def\u4ee5\u9032\u4e00\u6b65\u63d0\u5347\u7cfb\u7d71\u6548\u80fd\u3002\u6211\u5011\u5c07\u8fa8\u8b58 \u7cfb\u7d71\u5206\u70ba Offline \u8207 Online \u5169\u968e\u6bb5\uff1aOffline \u968e\u6bb5\u4f9d\u64da\u8072\u5b78\u7279\u6027\u5c07\u6574\u500b\u8a13\u7df4\u8cc7\u6599\u96c6\u5206\u5272\u6210 \u591a\u500b\u5b50\u8a13\u7df4\u8cc7\u6599\u96c6\uff0c\u4e26\u5efa\u7acb\u5404\u5b50\u8a13\u7df4\u8cc7\u6599\u96c6\u7684\u985e\u795e\u7d93\u7db2\u8def(\u4ee5\u985e\u795e\u7d93\u5b50\u7db2\u8def\u7a31\u4e4b)\u3002Online \u968e\u6bb5\u5247\u4f7f\u7528 GMM-gate \u4f86\u63a7\u5236\u985e\u795e\u7d93\u5b50\u7db2\u8def\u7684\u8f38\u51fa\u3002\u65b0\u63d0\u51fa\u7684\u7cfb\u7d71\u67b6\u69cb\u4fdd\u7559\u5b50\u8a13\u7df4\u8cc7\u6599 \u96c6\u7684\u8072\u5b78\u7279\u6027\uff0c\u5f37\u5065\u8a9e\u97f3\u8fa8\u8b58\u7cfb\u7d71\u3002\u5be6\u9a57\u4e0a\uff0c\u6211\u5011\u4f7f\u7528 Aurora 2 \u9023\u7e8c\u6578\u5b57\u8a9e\u97f3\u8cc7\u6599\u5eab\uff0c \u4f9d\u64da\u5b57\u932f\u8aa4\u7387(word error rate, WER)\u6bd4\u8f03\u6211\u5011\u63d0\u51fa\u7684\u8a9e\u97f3\u8fa8\u8b58\u7cfb\u7d71\u67b6\u69cb\u8207\u50b3\u7d71\u4ee5\u985e\u795e\u7d93 \u7db2\u8def\u5efa\u7acb\u7684\u8fa8\u8b58\u7cfb\u7d71\uff0c\u5e73\u5747\u5b57\u932f\u8aa4\u7387\u9032\u6b65 5.9% \uff0c\u7531 5.25%\u964d\u4f4e\u81f3 4.94%\u3002", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Abstract", "sec_num": null } ], "body_text": [ { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "\u9ad8\u65af\u6df7\u548c\u6a21\u578b\u662f\u7528\u4f86\u6a21\u64ec\u8907\u96dc\u8cc7\u6599\u5206\u5e03\u7684\u6a5f\u7387\u6a21\u578b\u3002\u4e00\u500b\u9ad8\u65af\u6df7\u5408\u6a21\u578b\u70ba K \u500b\u9ad8\u65af \u6a5f\u7387\u5bc6\u5ea6\u51fd\u6578\u7684\u52a0\u6b0a\u7e3d\u5408\uff0c\u5982\u5f0f(1)\u3002 \uf0e5 \uf03d \uf053 \uf03d K k k k k x N x p 1 ) , | ( ) | ( \uf06d \uf06c \uf066 ( 1 ) \u5176\u4e2d ) ( ) ( ) ( 2 -1 2 1 k 2 k k 1 e ) 2 ( 1 ) , | ( k k T k x x d x N \uf06d \uf06d \uf070 \uf06d \uf02d \uf053 \uf02d \uf02d \uf053 \uf03d \uf053 ( 2 ) \u53e6\u5916\uff0c K k k k u 1 ) , ( \uf03d \uf0e5 \uf03d \uf066 \u70ba\u5404\u500b\u9ad8\u65af\u7684\u53c3\u6578\uff0c k u \u53ca k \uf0e5 \u5206\u5225\u70ba\u7b2c k \u500b\u9ad8\u65af\u6210\u5206(Component)\u7684 \u5e73\u5747(mean)\u53ca\u5171\u8b8a\u7570\u77e9\u9663(covariance matrix)\u3002 k \uf06c \u70ba\u7b2c k \u500b\u9ad8\u65af\u6210\u5206\u7684\u5148\u9a57\u6a5f\u7387\uff0c\u4e26\u4e14\u6eff \u8db3\uff1a 0 1 1 and k K k k \uf0b3 \uf03d \uf0e5 \uf03d \uf06c \uf06c ( 3 ) \u9ad8\u65af\u6df7\u5408\u6a21\u578b\u7684\u53c3\u6578\uff0c\u53ef\u4ee5\u4f7f\u7528 EM \u6f14\u7b97\u6cd5(Expectation-Maximization algorithm)\uff0c\u7d93\u904e Expectation \u6b65\u9a5f\u53ca Maximization \u6b65\u9a5f\u7684\u758a\u4ee3\u4f86\u9032\u884c\u6a21\u578b\u53c3\u6578\u7684\u4f30\u8a08\u3002 (\u4e09)\u3001\u985e\u795e\u7d93\u7db2\u8def \u985e \u795e \u7d93 \u7db2\u8def (Neural Network \uff0c NN) \u662f \u4e00 \u7a2e \u6a21 \u64ec\u751f \u7269 \u5927\u8166 \u7684\u6a5f \u5668 \u5b78 \u7fd2 (machine learning)\u6a21\u578b\u3002\u69cb\u6210\u4e00\u500b\u985e\u795e\u7d93\u7db2\u8def\u7684\u57fa\u672c\u5143\u7d20\u70ba\u795e\u7d93\u5143(neuron)\uff0c\u5982\u5716\u4e8c\u6240\u793a\u3002\u4e00\u500b\u795e \u7d93\u5143\u7684\u7d50\u69cb\uff0c\u662f\u7531\u591a\u500b\u8f38\u5165\u7d93\u904e\u7dda\u6027\u7d44\u5408\uff0c\u4e26\u7d93\u904e\u6fc0\u767c\u51fd\u6578(activation function)\u5f8c\u7522\u751f \u8f38\u51fa y \u3002 x 1 x 2 x 3 y w 1 w 2 w 3 \u5716\u4e8c\u3001\u795e\u7d93\u5143\u793a\u610f\u5716 \u53ef\u7531\u5f0f(4)\u8868\u793a\uff1a ) ( b w x f y i i i \uf02b \uf03d \uf0e5 ( 4 ) \u5176 \u4e2d \uf07b \uf07d n i x i , ... , 2 , 1 | \uf03d \u70ba\u8f38\u5165\u8cc7\u6599\u3001 \uf07b \uf07d n i w i , ... , 2 , 1 | \uf03d \u70ba\u6b0a\u91cd\u503c (weight) \uff0c\u4ee3 \u8868\u7531\u8cc7\u6599 \u9032\u5165\u795e\u7d93\u5143\u7684\u6b0a\u91cd\uff1bb \u70ba\u504f\u79fb\u91cf(bias)\uff0c\u6700\u5f8c\uff0c ) ( \u2027 f \u70ba\u6fc0\u767c\u51fd\u6578\u3002 \u5e38\u898b\u7684\u6fc0\u767c\u51fd\u6578\u6709\uff1a Sigmoid\uff1a x e x f \uf02d \uf02b \uf03d 1 1 ) ( ( 5 ) Tanh\uff1a x x x x e e e e x f \uf02d \uf02d \uf02b \uf02d \uf03d ) ( ( 6 ) Rectified Linear\uff1a ) , 0 max( ) ( x x f \uf03d", "eq_num": "( 7 )" } ], "section": "", "sec_num": null }, { "text": "\u5716\u4e09\u3001Sigmoid\u3001Tanh \u53ca Rectified Linear ", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "% 100 \uf0b4 \uf02b \uf02b \uf03d N I D S WER ( 9 ) \u5728 \u5b57 \u4e32 \u6bd4 \u5c0d \u4e2d \uff0c \u5169 \u500b \u5b57 \u4e32 \u53ef \u80fd \u6703 \u767c \u751f \u63d2 \u5165 (Insertion) \u3001 \u522a \u9664 (Deletion) \u4ee5 \u53ca \u66ff \u63db (", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "\uf0e5 \uf02d \uf02d i i i T i i w X X ) ( ) ( min arg w t w t ( 10 ) \u5176\u4e2d i X \u70ba\u7b2c i \u7b46\u8cc7\u6599\u7d93\u904e\u5168\u90e8 7 \u500b\u985e\u795e\u7d93\u7db2\u8def\u7684\u8f38\u51fa\u3001 i t \u70ba\u7b2c i \u7b46\u8cc7\u6599\u7d93\u904e\u6b63\u78ba\u985e\u795e\u7d93 \u7db2\u8def\u7684\u8f38\u51fa\u3001 w \u70ba\u6b32\u6c42\u5f97\u7684\u7d44\u5408\u4fc2\u6578\u3002\u5c0d w \u6c42\u89e3\u4e26\u52a0\u5165\u4e00\u822c\u5316\u9805\u53ef\u5beb\u70ba\uff1a ) ( ) ( 1 \uf0e5 \uf0e5 \uf02d \uf02b \uf03d i i T i i i T i X I X X t w \uf064", "eq_num": "(" } ], "section": "", "sec_num": null } ], "back_matter": [], "bib_entries": { "BIBREF0": { "ref_id": "b0", "title": "Maximum a posteriori estimation for multivariate Gaussian mixture observations of Markov chains", "authors": [ { "first": "J", "middle": [], "last": "Gauvain", "suffix": "" }, { "first": "C.-H", "middle": [], "last": "Lee", "suffix": "" } ], "year": 1994, "venue": "IEEE Transactions on Speech and Audio Processing", "volume": "2", "issue": "2", "pages": "291--298", "other_ids": {}, "num": null, "urls": [], "raw_text": "J. Gauvain and C.-H. Lee, \"Maximum a posteriori estimation for multivariate Gaussian mixture observations of Markov chains,\" IEEE Transactions on Speech and Audio Processing, vol. 2, no. 2, pp. 291-298, Apr. 1994.", "links": null }, "BIBREF1": { "ref_id": "b1", "title": "Maximum likelihood linear transformations for HMM-based speech recognition", "authors": [ { "first": "M", "middle": [ "J F" ], "last": "Gales", "suffix": "" } ], "year": 1998, "venue": "Computer Speech and Language", "volume": "12", "issue": "2", "pages": "75--98", "other_ids": {}, "num": null, "urls": [], "raw_text": "M. J. F. Gales, \"Maximum likelihood linear transformations for HMM-based speech recognition,\" Computer Speech and Language, vol. 12, no. 2, pp. 75-98, Apr. 1998.", "links": null }, "BIBREF2": { "ref_id": "b2", "title": "Minimum classification error linear regression for acoustic model adaptation of continuous density HMMs", "authors": [ { "first": "X", "middle": [], "last": "He", "suffix": "" }, { "first": "C", "middle": [], "last": "Wu", "suffix": "" } ], "year": 2003, "venue": "International Conference on Multimedia and Expo", "volume": "1", "issue": "", "pages": "397--400", "other_ids": {}, "num": null, "urls": [], "raw_text": "X. He and C. Wu, \"Minimum classification error linear regression for acoustic model adaptation of continuous density HMMs,\" International Conference on Multimedia and Expo, vol. 1, pp. 397-400, July 2003.", "links": null }, "BIBREF3": { "ref_id": "b3", "title": "Speaker-adaptation for hybrid HMM-ANN continuous speech recognition system", "authors": [ { "first": "J", "middle": [], "last": "Neto", "suffix": "" }, { "first": "L", "middle": [], "last": "Almeida", "suffix": "" }, { "first": "M", "middle": [], "last": "Hochberg", "suffix": "" }, { "first": "C", "middle": [], "last": "Martins", "suffix": "" }, { "first": "L", "middle": [], "last": "Nunes", "suffix": "" }, { "first": "S", "middle": [], "last": "Renals", "suffix": "" }, { "first": "T", "middle": [], "last": "Robinson", "suffix": "" } ], "year": 1995, "venue": "Proceedings of Eurospeech", "volume": "", "issue": "", "pages": "18--21", "other_ids": {}, "num": null, "urls": [], "raw_text": "J. Neto, L. Almeida, M. Hochberg, C. Martins, L. Nunes, S. Renals and T. Robinson, \"Speaker-adaptation for hybrid HMM-ANN continuous speech recognition system,\" In Proceedings of Eurospeech, pp. 18-21, Sep. 1995.", "links": null }, "BIBREF4": { "ref_id": "b4", "title": "Comparison of discriminative input and output transformations for speaker adaptation in the hybrid NN/HMM systems", "authors": [ { "first": "B", "middle": [], "last": "Li", "suffix": "" }, { "first": "K", "middle": [ "C" ], "last": "Sim", "suffix": "" } ], "year": 2010, "venue": "Proceedings of Interspeech", "volume": "", "issue": "", "pages": "526--529", "other_ids": {}, "num": null, "urls": [], "raw_text": "B. Li and K. C. Sim, \"Comparison of discriminative input and output transformations for speaker adaptation in the hybrid NN/HMM systems,\" In Proceedings of Interspeech, pp. 526-529, 2010.", "links": null }, "BIBREF5": { "ref_id": "b5", "title": "On combining DNN and GMM with unsupervised speaker adaptation for robust automatic speech recognition", "authors": [ { "first": "B", "middle": [], "last": "Li", "suffix": "" }, { "first": "K", "middle": [ "C" ], "last": "Sim", "suffix": "" } ], "year": 2014, "venue": "IEEE International Conference on Acoustics, Speech and Signal Processing", "volume": "", "issue": "", "pages": "195--199", "other_ids": {}, "num": null, "urls": [], "raw_text": "B. Li and K. C. Sim, \"On combining DNN and GMM with unsupervised speaker adaptation for robust automatic speech recognition,\" IEEE International Conference on Acoustics, Speech and Signal Processing, pp. 195-199, May 2014.", "links": null }, "BIBREF6": { "ref_id": "b6", "title": "Deep neural network based speech separation for robust speech recognition", "authors": [ { "first": "Y", "middle": [], "last": "Tu", "suffix": "" }, { "first": "J", "middle": [], "last": "Du", "suffix": "" }, { "first": "Y", "middle": [], "last": "Xu", "suffix": "" }, { "first": "L", "middle": [], "last": "Dai", "suffix": "" }, { "first": "C.-H", "middle": [], "last": "Lee", "suffix": "" } ], "year": 2014, "venue": "International Symposium on Chinese Spoken Language Processing", "volume": "", "issue": "", "pages": "532--536", "other_ids": {}, "num": null, "urls": [], "raw_text": "Y. Tu, J. Du, Y. Xu, L. Dai, and C.-H. Lee, \"Deep neural network based speech separation for robust speech recognition,\" in International Symposium on Chinese Spoken Language Processing, pp.532-536, Oct. 2014.", "links": null }, "BIBREF7": { "ref_id": "b7", "title": "Bagging predictors", "authors": [ { "first": "L", "middle": [], "last": "Breiman", "suffix": "" } ], "year": 1996, "venue": "Journal of Machine Learning", "volume": "24", "issue": "2", "pages": "123--140", "other_ids": {}, "num": null, "urls": [], "raw_text": "L. Breiman, \"Bagging predictors,\" Journal of Machine Learning, vol. 24, no. 2, pp. 123-140, Aug. 1996.", "links": null }, "BIBREF8": { "ref_id": "b8", "title": "The strength of weak learnability", "authors": [ { "first": "R", "middle": [ "E" ], "last": "Schapire", "suffix": "" } ], "year": 1990, "venue": "Journal of Machine Learning", "volume": "5", "issue": "2", "pages": "197--227", "other_ids": {}, "num": null, "urls": [], "raw_text": "R. E. Schapire, \"The strength of weak learnability,\" Journal of Machine Learning, vol. 5, no. 2, pp. 197-227, Jun. 1990.", "links": null }, "BIBREF9": { "ref_id": "b9", "title": "Incorporating local information of the acoustic environments to MAP-based feature compensation and acoustic model adaptation", "authors": [ { "first": "Y", "middle": [], "last": "Tsao", "suffix": "" }, { "first": "X", "middle": [], "last": "Lu", "suffix": "" }, { "first": "P", "middle": [], "last": "Dixon", "suffix": "" }, { "first": "T", "middle": [], "last": "Hu", "suffix": "" }, { "first": "S", "middle": [], "last": "Matsuda", "suffix": "" }, { "first": "C", "middle": [], "last": "Hori", "suffix": "" } ], "year": 2014, "venue": "Computer Speech and Language", "volume": "28", "issue": "3", "pages": "709--726", "other_ids": {}, "num": null, "urls": [], "raw_text": "Y. Tsao, X. Lu, P. Dixon, T.-y. Hu, S. Matsuda, and C. Hori, \"Incorporating local information of the acoustic environments to MAP-based feature compensation and acoustic model adaptation,\" Computer Speech and Language, vol. 28, no. 3, pp. 709-726, May 2014.", "links": null }, "BIBREF10": { "ref_id": "b10", "title": "Adaptive mixtures of local experts", "authors": [ { "first": "R", "middle": [ "A" ], "last": "Jacobs", "suffix": "" }, { "first": "M", "middle": [ "I" ], "last": "Jordan", "suffix": "" }, { "first": "S", "middle": [ "J" ], "last": "Nowlan", "suffix": "" }, { "first": "G", "middle": [ "E" ], "last": "Hinton", "suffix": "" } ], "year": 1991, "venue": "Neural Computation", "volume": "3", "issue": "1", "pages": "79--87", "other_ids": {}, "num": null, "urls": [], "raw_text": "R. A. Jacobs, M. I. Jordan, S. J. Nowlan and G. E. Hinton, \"Adaptive mixtures of local experts,\" Neural Computation, vol. 3, no. 1, pp. 79-87, Spring 1991.", "links": null }, "BIBREF11": { "ref_id": "b11", "title": "Learning deep architectures for AI", "authors": [ { "first": "Y", "middle": [], "last": "Bengio", "suffix": "" } ], "year": 2009, "venue": "Foundation and Trends in Machine Learning", "volume": "2", "issue": "", "pages": "1--127", "other_ids": {}, "num": null, "urls": [], "raw_text": "Y. Bengio, \"Learning deep architectures for AI,\" Foundation and Trends in Machine Learning, vol. 2, pp. 1-127, 2009.", "links": null }, "BIBREF12": { "ref_id": "b12", "title": "Universal background model based speech recognition", "authors": [ { "first": "D", "middle": [], "last": "Povey", "suffix": "" }, { "first": "S", "middle": [ "M" ], "last": "Chu", "suffix": "" }, { "first": "B", "middle": [], "last": "Varadarajan", "suffix": "" } ], "year": 2008, "venue": "IEEE International Conference on Acoustics, Speech, and Signal Processing", "volume": "", "issue": "", "pages": "4561--4564", "other_ids": {}, "num": null, "urls": [], "raw_text": "D. Povey, S. M. Chu, B. Varadarajan, \"Universal background model based speech recognition,\" IEEE International Conference on Acoustics, Speech, and Signal Processing, pp. 4561-4564, Mar. 2008.", "links": null }, "BIBREF14": { "ref_id": "b14", "title": "The Kaldi speech recognition toolkit", "authors": [ { "first": "Y", "middle": [], "last": "Motlicek", "suffix": "" }, { "first": "P", "middle": [], "last": "Qian", "suffix": "" }, { "first": "J", "middle": [], "last": "Schwarz", "suffix": "" }, { "first": "G", "middle": [], "last": "Silovsky", "suffix": "" }, { "first": "K", "middle": [], "last": "Stemmer", "suffix": "" }, { "first": "", "middle": [], "last": "Vesely", "suffix": "" } ], "year": 2011, "venue": "IEEE Workshop on Automatic Speech Recognition and Understanding", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Motlicek, Y. Qian, P. Schwarz, J. Silovsky, G. Stemmer, and K. Vesely, \"The Kaldi speech recognition toolkit,\" IEEE Workshop on Automatic Speech Recognition and Understanding, Dec. 2011.", "links": null }, "BIBREF15": { "ref_id": "b15", "title": "The Aurora experimental framework for the performance evaluation of speech recognition systems under noisy conditions", "authors": [ { "first": "D", "middle": [], "last": "Pearce", "suffix": "" }, { "first": "H.-G", "middle": [], "last": "Hirsch", "suffix": "" } ], "year": 2000, "venue": "ASR2000 Automatic Speech Recognition: Challenges for the new Millenium ISCA Tutorial and Research Workshop", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "D. Pearce and H.-G. Hirsch, \"The Aurora experimental framework for the performance evaluation of speech recognition systems under noisy conditions,\" in ASR2000 Automatic Speech Recognition: Challenges for the new Millenium ISCA Tutorial and Research Workshop, Sep. 2000.", "links": null }, "BIBREF16": { "ref_id": "b16", "title": "Dropout: a simple way to prevent neural networks from overfitting", "authors": [ { "first": "N", "middle": [], "last": "Srivastava", "suffix": "" }, { "first": "G", "middle": [], "last": "Hinton", "suffix": "" }, { "first": "A", "middle": [], "last": "Krizhevsky", "suffix": "" }, { "first": "I", "middle": [], "last": "Sutskever", "suffix": "" }, { "first": "R", "middle": [], "last": "Salakhutdinov", "suffix": "" } ], "year": 2014, "venue": "Journal of Machine Learning Research", "volume": "15", "issue": "", "pages": "1929--1958", "other_ids": {}, "num": null, "urls": [], "raw_text": "N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever and R. Salakhutdinov, \"Dropout: a simple way to prevent neural networks from overfitting,\" Journal of Machine Learning Research, vol. 15, pp. 1929-1958, Jan. 2014.", "links": null }, "BIBREF17": { "ref_id": "b17", "title": "A spectral masking approach to noise-robust speech recognition using deep neural networks", "authors": [ { "first": "B", "middle": [], "last": "Li", "suffix": "" }, { "first": "K", "middle": [ "C" ], "last": "Sim", "suffix": "" } ], "year": 2014, "venue": "IEEE Transactions on Audio, Speech and Language Processing", "volume": "22", "issue": "", "pages": "1296--1305", "other_ids": {}, "num": null, "urls": [], "raw_text": "B. Li and K. C. Sim, \"A spectral masking approach to noise-robust speech recognition using deep neural networks,\" IEEE Transactions on Audio, Speech and Language Processing, vol. 22, pp. 1296-1305, Aug. 2014.", "links": null } }, "ref_entries": { "TABREF0": { "num": null, "content": "
Training Stage
FeatureModel
ExtractionTraining
Training Speech
AcousticLanguage
ModelModel
Testing Stage
\u795e\u7d93\u7db2\u8def\u7522\u751f\u5206\u5225\u70ba\u76ee\u6a19\u8a0a\u865f\u8207\u5e72\u64fe\u8a0a\u865f\u7684\u5169\u500b\u8f38\u51fa\uff0c\u4f7f\u7528\u5206\u96e2\u7684\u7d50\u679c\u9032\u884c\u8fa8\u8b58[7]\u7b49 \u7b49\u8a31\u591a\u65b9\u5f0f\u3002\u5728\u9019\u4e9b\u76f8\u95dc\u7814\u7a76\u4e2d\uff0c\u90fd\u662f\u4f7f\u7528\u540c\u4e00\u500b\u985e\u795e\u7d93\u7db2\u8def\u4f86\u8655\u7406\u6240\u6709\u74b0\u5883\u7684\u60c5\u6cc1\u3002 Output Sentence Feature Extraction Decode
\u5728\u6574\u9ad4\u5b78\u7fd2(ensemble learning)\u7684\u76f8\u95dc\u7814\u7a76\u4e2d\uff0c\u6709\u4f7f\u7528 bagging[8]\u6216\u662f boosting[9]\u7b49\u7b49\u65b9
\u5f0f\uff0c\u9019\u88e1\u6211\u5011\u4f7f\u7528\u57fa\u65bc Environment Clustering (EC)[10]\u53ca Mixture of Experts[11]\u7684\u67b6\u69cb Testing Speech
\u4f86\u8a13\u7df4\u591a\u500b\u985e\u795e\u7d93\u7db2\u8def\uff0c\u4e26\u5728\u6700\u5f8c\u9078\u64c7\u4e00\u500b\u9069\u7576\u7684\u985e\u795e\u7d93\u7db2\u8def\u9032\u884c\u8f38\u51fa\u3002 \u5728\u63a5\u4e0b\u4f86\u7684\u5167\u5bb9\uff0c\u7b2c\u4e8c\u7ae0\u5c07\u4ecb\u7d39\u6574\u500b\u8a9e\u97f3\u8fa8\u8b58\u7684\u4e3b\u8981\u6d41\u7a0b\u4ee5\u53ca\u4e00\u4e9b\u76f8\u95dc\u7684\u7814\u7a76\u65b9 \u5716\u4e00\u3001\u8a9e\u8005\u8fa8\u8b58\u6d41\u7a0b\u5716
\u6cd5\u3002\u7b2c\u4e09\u7ae0\u5c07\u4ecb\u7d39\u672c\u7bc7\u8ad6\u6587\u7684\u7cfb\u7d71\u67b6\u69cb\u3002\u7b2c\u56db\u7ae0\u70ba\u5be6\u9a57\u7684\u90e8\u5206\uff0c\u6b64\u7ae0\u7bc0\u5305\u542b\u4ecb\u7d39\u5be6\u9a57\u8a9e
\u6599\u8207\u5be6\u9a57\u8a2d\u5b9a\u3001baseline \u7cfb\u7d71\u4ee5\u53ca\u672c\u8ad6\u6587\u7cfb\u7d71\u7684 Word Error Rate (WER)\u3002\u7b2c\u4e94\u7ae0\u70ba\u6b64\u7814 \u7a76\u7684\u7d50\u8ad6\u3002 (\u4e8c)\u3001\u9ad8\u65af\u6df7\u5408\u6a21\u578b(Gaussian Mixture Model, GMM)
\u4e8c\u3001\u8a9e\u97f3\u8fa8\u8b58\u6d41\u7a0b\u53ca\u76f8\u95dc\u7814\u7a76\u65b9\u6cd5\u4ecb\u7d39
\u5728\u6b64\u7ae0\u7bc0\u4e2d\u6211\u5011\u5c07\u7c21\u55ae\u4ecb\u7d39\u57fa\u672c\u7684\u8a9e\u97f3\u8fa8\u8b58\u6d41\u7a0b\uff0c\u53ca\u8fa8\u8b58\u4e2d\u6240\u4f7f\u7528\u7684\u9ad8\u65af\u6df7\u5408\u6a21\u578b
(Gaussian Mixture Model\uff0cGMM) \u8207\u985e\u795e\u7d93\u7db2\u8def(Neural Network)\u3002
(\u4e00)\u3001\u8a9e\u97f3\u8fa8\u8b58\u6d41\u7a0b
\u5716\u4e00\u70ba\u4e00\u500b\u57fa\u672c\u7684\u8a9e\u97f3\u8fa8\u8b58\u6d41\u7a0b\uff0c\u53ef\u5206\u70ba\u8a13\u7df4\u53ca\u6e2c\u8a66\u968e\u6bb5\u3002\u9996\u5148\u64f7\u53d6\u8a9e\u97f3\u8a0a\u865f\u7684\u7279
\u5fb5(feature extraction)\uff0c\u5982\u6885\u723e\u5012\u983b\u8b5c\u4fc2\u6578(Mel-Frequency Cepstral Coefficients, MFCC)\uff1b
\u63a5\u8457\u5229\u7528\u64f7\u53d6\u7684\u8a9e\u97f3\u7279\u5fb5\u5728\u8a13\u7df4\u968e\u6bb5\u8a13\u7df4\u6a21\u578b(model training)\uff0c\u6216\u5728\u6e2c\u8a66\u968e\u6bb5\u89e3\u78bc
(decode) \u70ba \u6587 \u5b57 \u3002 \u8a13 \u7df4 \u968e \u6bb5 \u5c07 \u7522 \u751f \u8072 \u5b78 \u6a21 \u578b (acoustic model) \u53ca\u8a9e\u8a00\u6a21\u578b(language
model)\uff0c\u4e26\u4f9b\u7d66\u6e2c\u8a66\u968e\u6bb5\u89e3\u78bc\u4f7f\u7528\u3002\u6b64\u5916\uff0c\u76ee\u524d\u8a13\u7df4\u8072\u5b78\u6a21\u578b\u7684\u65b9\u5f0f\u4e3b\u8981\u70ba GMM \u8207\u985e
\u795e\u7d93\u7db2\u8def\uff0c\u5c07\u65bc\u4e0b\u4e00\u7bc0\u4ecb\u7d39\u3002
", "html": null, "type_str": "table", "text": "The 2015 Conference on Computational Linguistics and Speech Processing ROCLING 2015, pp. 136-147 \uf0d3 The Association for Computational Linguistics and Chinese Language Processing \u4e00\u3001\u7c21\u4ecb \u96d6\u7136\u8a9e\u97f3\u8fa8\u8b58\u7cfb\u7d71\u5728\u5b89\u975c\u74b0\u5883\u4e0b\u53ef\u4ee5\u9054\u5230\u4e0d\u932f\u7684\u8fa8\u8b58\u7387\uff0c\u4f46\u662f\u5728\u5be6\u969b\u61c9\u7528\u4e0a\uff0c\u7531\u65bc \u74b0\u5883\u566a\u97f3(environment noise) \u7522 \u751f \u7684 \u52a0 \u6210 \u6027 \u96dc \u8a0a (additive noise) \u53ca \u901a \u9053 \u5931 \u771f (channel distortion)\u7522\u751f\u7684\u5377\u7a4d\u6027\u96dc\u8a0a(convolutive noise)\u7b49\u60c5\u6cc1\uff0c\u9020\u6210\u8a13\u7df4\u53ca\u6e2c\u8a66\u8a9e\u6599\u7684\u74b0\u5883\u4e0d\u5339 \u914d\u554f\u984c\uff0c\u9650\u5236\u8a9e\u97f3\u8fa8\u8b58\u7cfb\u7d71\u7684\u6548\u80fd\u3002 \u6b32\u89e3\u6c7a\u4e0a\u8ff0\u7684\u4e0d\u5339\u914d\u554f\u984c\uff0c\u5728\u6a21\u578b\u7a7a\u9593(model space)\u7684\u8655\u7406\u4e2d\u6709\u8a31\u591a\u6a21\u578b\u8abf\u9069 (model adaptation)\u7684\u65b9\u5f0f\uff0c\u4f8b\u5982\u6700\u5927\u5f8c\u9a57\u6a5f\u7387\u4f30\u8a08(maximum a posteriori estimation)[1]\u3001 \u6700\u5927\u4f3c\u7136\u7dda\u6027\u8ff4\u6b78(maximum likelihood linear regression)[2]\u3001\u6700\u5c0f\u5206\u985e\u932f\u8aa4\u7dda\u6027\u56de\u6b78 (minimum classification error linear regression)[3]\u7b49\u7b49\u3002 \u5728\u5f37\u5065\u6027\u8a9e\u97f3\u8fa8\u8b58\u4e0a\u5df2\u7d93\u6709\u8a31\u591a\u4f7f\u7528\u7814\u7a76\u4f7f\u7528\u985e\u795e\u7d93\u7db2\u8def\uff0c\u4f8b\u5982\uff0c\u5728\u74b0\u5883\u4e0d\u5339\u914d\u7684 \u60c5\u6cc1\u4e0b\u4f7f\u7528\u7dda\u6027\u8f49\u63db\u5f37\u5065\u6a21\u578b[4][5]\uff1b\u7d50\u5408 GMM-HMM \u8207 DNN-HMM \u9032\u884c\u8f38\u51fa[6]\uff1b\u985e" }, "TABREF1": { "num": null, "content": "
\u6b64\u5916\uff0c\u4e00\u500b\u5b8c\u6574\u7684\u985e\u795e\u7d93\u7db2\u8def\u70ba\u591a\u500b\u795e\u7d93\u5143\u67b6\u69cb\u800c\u6210\uff0c\u5982\u5716\u4e94\u70ba\u96d9\u96b1\u85cf\u5c64( hidden layer) \u7684\u985e\u795e\u7d93\u7db2\u8def\uff0c\u7e3d\u5171\u7531\u4e94\u500b\u795e\u7d93\u5143\u7d44\u6210(\u7b2c\u4e00\u5c64\u6709\u4e09\u500b\u795e\u7d93\u5143\u7bc0\u9ede\uff0c\u7b2c\u4e8c\u5c64\u5247\u70ba\u5169\u500b\u795e \u7684\u8072\u5b78\u7279\u6027\u5206\u985e\u8a13\u7df4\u8cc7\u6599\u593e\u70ba\u516d\u500b\u5b50\u8a13\u7df4\u8cc7\u6599\u96c6\uff0c\u4e26\u5206\u5225\u5c0d ALL NN \u9032\u884c\u5012\u50b3\u905e\u8abf\u6574\u7db2\u8def \u56db\u3001\u5be6\u9a57\u8207\u7d50\u679c HMM observation probability \u53c3\u6578\uff0c\u5f97\u5230\u516d\u500b\u985e\u795e\u7d93\u5b50\u7db2\u8def male NN \u3001 female NN \u3001 FH NN \u3001 FL NN \u3001 MH NN \u4ee5\u53ca NN \u3002 ML \u7d93\u5143\u7bc0\u9ede)\u3002\u8cc7\u6599\u8f38\u5165\u81f3\u7b2c\u4e00\u5c64\u7684\u795e\u7d93\u5143\u7684\uff0c\u800c\u7b2c\u4e8c\u5c64\u7684\u8f38\u5165\u5247\u70ba\u7b2c\u4e00\u5c64\u7684\u8f38\u51fa\u3002\u5176\u4e2d \u7684\u53c3\u6578\uf07b \uf07d n i w i , ... , 2 , 1 | \uf03d \u8207b \u53ef\u7531\u5012\u50b3\u905e(back propagation)\u8a13\u7df4\u800c\u5f97\uff1b\u8a73\u7d30\u7684\u7db2\u8def\u8a13\u7df4\u6d41 \u7a0b\u53ef\u53c3\u8003[12]\u3002 Control Gating function GMM \u3001 female GMM \u3001 FH GMM \u3001 FL GMM \u3001 \u5728\u672c\u7bc0\uff0c\u6211\u5011\u5c07\u4ecb\u7d39\u5be6\u9a57\u7684\u8a2d\u5b9a\u3001\u4e26\u5206\u6790\u6bd4\u8f03\u50b3\u7d71\u5229\u7528\u985e\u795e\u7d93\u7db2\u8def\u6a21\u578b\u7684\u8fa8\u8b58\u7cfb\u7d71 \u53e6\u5916\uff0cGMM \u6a21\u578b\u7684\u8a13\u7df4\uff0c\u9996\u5148\u4ee5\u8a13\u7df4\u8cc7\u6599\u96c6\u4f9d\u64da\u5f0f(1)male \u4ee5\u53ca\u672c\u6587\u63d0\u51fa\u7684\u5f37\u5065\u6027\u8a9e\u97f3\u8fa8\u8b58\u7cfb\u7d71\u7684\u7d50\u679c\u3002
MH GMM \u8207ML GMM \u3002
x 1 (\u4e00)\u3001\u5be6\u9a57\u8a9e\u97f3\u8cc7\u6599\u8207\u5be6\u9a57\u8a2d\u5b9a
x 2 (\u4e8c)\u3001Online \u7cfb\u7d71\u5efa\u69cb \u8a9e\u97f3\u8fa8\u8b58\u7684\u5be6\u9a57\uff0c\u6211\u5011\u4f7f\u7528 Kaldi \u9019\u5957\u7528\u65bc\u8a9e\u97f3\u8fa8\u8b58\u7684\u958b\u653e\u539f\u59cb\u78bc\u5de5\u5177[14] \uff0c\u4e26\u505a y 1 NN 1 NN 2 NN 7 \u2027\u2027\u2027 \u70ba\u6211\u5011\u7684 baseline NN-HMM \u7cfb\u7d71\uff1b\u4e26\u4ee5 Aurora 2 \u8cc7\u6599\u5eab[15] \u505a\u70ba\u672c\u5be6\u9a57\u7684\u8a9e\u6599\u5eab\u3002 \u524d\u4e00\u5c0f\u7bc0\u5f97\u5230\u7684\u6574\u9ad4\u6a21\u578b\u53ca\u516d\u500b\u5b50\u96c6\u6a21\u578b\uff0c\u5c07\u63d0\u4f9b\u7d66 online \u968e\u6bb5\u4f7f\u7528\u3002\u5982\u5716\u4e03\u6240 Aurora 2 \u70ba\u4e00\u500b\u82f1\u6587\u9023\u7e8c\u6578\u5b57\u8a9e\u97f3\u7684\u8cc7\u6599\u5eab\uff0c\u5305\u542b\u516b\u7a2e\u4e0d\u540c\u7684\u52a0\u6210\u6027\u96dc\u8a0a\u74b0\u5883(Subway, \u793a\uff0c\u5728 online \u968e\u6bb5\u6642\uff0c\u6211\u5011\u5c07\u6574\u53e5\u6e2c\u8a66\u8cc7\u6599\u5229\u7528\u5f0f(1)\uff0c\u5206\u5225\u8a08\u7b97\u5404\u5b50\u96c6 GMM \u6a21\u578b\u7684 \u4e03\u500b\u5e73\u5747\u5f8c\u9a57\u6a5f\u7387\uff0c\u5f97\u5230\u4e03\u500b\u5e73\u5747\u5f8c\u9a57\u6a5f\u7387 7 2 1 , ... , , p p p Babble, Car, Exhibition, Airport, Street, Train Station, Restaurant)\u3001\u5169\u7a2e\u4e0d\u540c\u7684\u901a\u9053\u96dc\u8a0a \uff0c\u4e26\u6c7a\u5b9a\u5176\u4e2d\u7684\u6700\u5927\u503c\u8207\u76f8\u5c0d (G712 and MIRS) \u8207\u4e03\u7a2e\u4e0d\u540c\u7684 SNR (clean, 20 dB, 15 dB, 10 dB, 5 dB, 0 dB, -5 dB)\u3002\u8a9e
y 2 \u6599\u5eab\u4e2d\uff0c\u542b\u6709\u96dc\u8a0a\u7684\u8a9e\u97f3\u70ba\u4eba\u5de5\u6dfb\u52a0\u4e0d\u540c\u7684\u96dc\u8a0a\u74b0\u5883\u8207 SNR \u5230\u4e7e\u6de8\u8a9e\u97f3\u4e0a\u3002\u53e6\u5916\uff0cAurora \u61c9\u7684\u7b2c i \u500b\u5b50\u96c6\uff0c\u5176\u4e2d i \u70ba\uff1a
x 3 \u4e09\u3001\u672c\u8ad6\u6587\u7cfb\u7d71\u67b6\u69cb \u540c\u4e00\u53e5\u8a9e\u97f3\u8a0a\u865f\u5728\u4e0d\u540c\u7684\u8a9e\u8005\u3001\u74b0\u5883\u7b49\u7b49\u60c5\u6cc1\u8868\u73fe\u7684\u8072\u5b78\u7279\u6027\u4e0d\u76e1\u76f8\u540c\uff0c\u56e0\u6b64\u53ef\u4f9d Layer 1 Layer 2 \u5716\u56db\u3001\u96d9\u96b1\u85cf\u5c64\u985e\u795e\u7d93\u7db2\u8def\u793a\u610f\u5716 \u64da\u4e0d\u540c\u7684\u8072\u5b78\u5206\u985e\u65b9\u5f0f\uff0c\u4f8b\u5982\u6027\u5225\u3001\u8a0a\u566a\u6bd4(signal-to-noise ratio\uff0cSNR)\u7b49\u7b49\uff0c\u5c07\u4e00\u4efd\u8a13 \u7df4\u8a9e\u6599\u5eab\u5206\u5272\u6210\u6578\u7a2e\u4e0d\u540c\u7684\u5b50\u96c6\uff0c\u4e26\u4ee5\u985e\u795e\u7d93\u7db2\u8def\u8207 GMM \u6a21\u578b\u5316\u6bcf\u4e00\u500b\u5b50\u96c6\u6240\u4ee3\u8868\u7684 \u8072\u5b78\u7279\u6027\u3002\u6e2c\u8a66\u6642\uff0c\u9996\u5148\u4ee5 GMM \u6a21\u578b\u6c7a\u5b9a\u6e2c\u8a66\u8a9e\u6599\u7684\u985e\u5225\uff0c\u518d\u4f9d\u64da\u5176\u7d50\u679c\uff0c\u9078\u64c7\u76f8\u5c0d \u61c9\u7684\u985e\u795e\u7d93\u7db2\u8def\u6a21\u578b\uff0c\u6700\u5f8c\u5f97\u5230\u8f03\u5177\u4ee3\u8868\u6027\u7684\u8a9e\u97f3\u7279\u6027\u8f38\u51fa\uff0c\u9032\u800c\u589e\u9032\u8fa8\u8b58\u6548\u679c\u3002 \u6211\u5011\u5c07\u7cfb\u7d71\u5206\u6210 online \u8207 offline \u968e\u6bb5\uff0c\u5716 \u4e94\u5247\u70ba\u4e00 online \u7684\u6d41\u7a0b\u5716\u3002offline \u968e\u6bb5\u4f9d \u64da\u4e0d\u540c\u8072\u5b78\u7279\u6027\u7684\u8cc7\u6599\u96c6\uff0c\u5404\u5225\u8a13\u7df4\u51fa\u5c0d\u61c9\u7684\u985e\u795e\u7d93\u7db2\u8def\uf07b \uf07d n 2 1 NN , ... , NN , NN (\u4ee5\u985e\u795e \u5716\u4e94\u3001\u672c\u8ad6\u6587\u7cfb\u7d71\u67b6\u69cb\u793a\u610f\u5716 (\u4e00)\u3001Offline \u7cfb\u7d71\u5efa\u69cb \u5728 offline \u7cfb\u7d71\u4e2d\uff0c\u6211\u5011\u5c07\u8a13\u7df4\u8cc7\u6599\u96c6\u4f9d\u64da\u6027\u5225\u4ee5\u53ca\u8a0a\u566a\u6bd4\u5206\u6210\u516d\u500b\u5b50\u8a13\u7df4\u8cc7\u6599\u96c6\uff1a \u7537\u6027\u3001\u5973\u6027\u3001\u7537\u6027\u9ad8 SNR\u3001\u7537\u6027\u4f4e SNR\u3001\u5973\u6027\u9ad8 SNR \u4ee5\u53ca\u5973\u6027\u4f4e SNR\uff1b\u5982\u5716\u516d\u6240\u793a\uff1a ALL (NN ALL ) Data 1 Data 2 Data 7 \u2027\u2027\u2027 k k p i 7 ,..., 2 , 1 max arg \uf03d ( 8 ) 2 \u8a9e\u6599\u5eab\u5305\u542b\u8a13\u7df4\u8207\u6e2c\u8a66\u7684\u8a9e\u6599\u96c6\uff1a\u8a13\u7df4\u8a9e\u6599\u5eab\u5305\u542b clean-\u8207 multi-condition \u5169\u7a2e\u8a13\u7df4 \uf03d \u8a9e\u6599\u5eab\uff0c\u672c\u5be6\u9a57\u4f7f\u7528 multi-condition \u8a13\u7df4\u8a9e\u6599\u5eab\u3002\u8a72\u8a9e\u6599\u5eab\u5305\u542b\u56db\u7a2e\u566a\u97f3\u985e\u578b (Subway, \u6700\u5f8c\uff0c\u518d\u7531\u7b2ci \u500b\u5b50\u96c6\u5c0d\u61c9\u7684\u985e\u795e\u7d93\u7db2\u8def\u7684\u8f38\u51fa\u4f5c\u70ba HMM \u7684\u89c0\u6e2c\u6a5f\u7387\u3002 NN 1 NN 2 NN 7 \u2027\u2027\u2027 Control GMM 1 GMM 2 \u2027 \u2027 \u2027 GMM 7 HMM observation probability y 1 y 2 y 7 p 1 p 2 p 7 Babble, Car, Exhibition) (\u4e8c)\u3001\u8a55\u4f30\u65b9\u6cd5
\u7d93\u5b50\u7db2\u8def\u7a31\u4e4b)\uff0c\u4e26\u4f9b\u7d66 online \u968e\u6bb5\u4f7f\u7528\u3002\u53e6\u5916\uff0c\u5728 online \u968e\u6bb5\u4f7f\u7528\u4e00\u500b gating function \u5be6\u9a57\u7d50\u679c\u7684\u8a55\u4f30\u65b9\u9762\uff0c\u6211\u5011\u4f7f\u7528\u5b57\u932f\u8aa4\u7387(Word Error Rate, WER)\u4f86\u8a55\u4f30\u5be6\u9a57\u7d50\u679c\uff0c \u4f86\u9078\u64c7\u985e\u795e\u7d93\u5b50\u7db2\u8def\u7684\u8f38\u51fa\uff0c\u4e26\u5f97\u5230\u6700\u5f8c\u7684\u8fa8\u8b58\u7d50\u679c\u3002\u6700 \u5f8c\uff0c\u6211\u5011\u9078\u64c7 GMM \u505a\u70ba gating function\uff0c\u4ee5 GMM-gate \u7a31\u4e4b\u3002 Male Female High Low High Low \u6211\u5011\u63d0\u51fa\u57fa\u65bc\u74b0\u5883\u7fa4\u96c6(Environment Clustering\uff0cEC)[10]\u4ee5\u53ca mixture of local Male Male Female Female Data \u5176\u8a08\u7b97\u65b9\u5f0f\u5982\u4e0b\u5f0f\uff1a
experts[11]\u7684\u591a\u985e\u795e\u7d93\u5b50\u7db2\u8def\u8a13\u7df4\u53ca\u7d50\u5408\u5404\u5b50\u7db2\u8def\u8f38\u51fa\u4e4b\u67b6\u69cb\uff0c\u4e0b\u4e00\u7bc0\u5c07\u4ecb\u7d39 offline \u7684 \u7cfb\u7d71\u5efa\u69cb\u6d41\u7a0b\u53ca online \u7684\u6e2c\u8a66\u6d41\u7a0b\u3002 SNR SNR SNR SNR \u5716\u4e03\u3001Online \u968e\u6bb5\u67b6\u69cb\u5716
(NN male )(NN female )(NN MH )(NN ML )(NN FH )(NN FL )
\u5716\u516d\u3001EC \u6a39\u67b6\u69cb
\u5176\u4e2d\uff0c\u985e\u795e\u7d93\u5b50\u7db2\u8def\u7684\u8a13\u7df4\uff0c\u9996\u5148\u4ee5\u8a13\u7df4\u8cc7\u6599\u96c6\u8a13\u7df4\u51fa global \u7684ALL NN \uff0c\u63a5\u8457\u4f9d\u64da\u4e0d\u540c
", "html": null, "type_str": "table", "text": "\uff0c\u4f7f\u7528 EM \u6f14\u7b97\u6cd5\u8a13\u7df4 UBM[13] \u6a21\u578b\uff0c ALL GMM \uff0c\u63a5\u8457\u5c0d\u6bcf\u4e00\u7a2e\u5b50\u8a13\u7df4\u8cc7\u6599\u96c6\u4ee5 MAP(Maximum a Posteriori) estimation \u8abf\u9069(adaptation)\u51fa\u516d\u7a2e\u5b50\u96c6 GMM \u6a21\u578b\uff1a \u8207\u4e94\u7a2e SNR (clean, 20 dB, 15 dB, 10 dB, 5 dB)\uff0c\u4e00\u5171\u6709 8440 \u53e5\uff0c \u7e3d\u9577\u5ea6\u7d04\u70ba\u56db\u500b\u5c0f\u6642\uff1b\u6e2c\u8a66\u8a9e\u6599\u96c6\u5247\u5206\u6210\u4e09\u500b\u5b50\u96c6 Set A\u3001Set B \u53ca Set C\uff0c\u5404\u6e2c\u8a66\u5b50\u96c6 \u4e2d\u7686\u6709\u4e0d\u540c\u7684 SNR \u74b0\u5883\uff0c\u5f9e 20 dB \u81f3-5 dB \u8207 clean\u3002Set A \u5305\u542b\u8207\u8a13\u7df4\u8a9e\u6599\u76f8\u540c\u7684\u56db\u7a2e \u566a\u97f3\uff0cSet B \u5247\u70ba\u5305\u542b Restaurant, Street, Airport \u8207 Train Station \u7684\u74b0\u5883\u96dc\u8a0a\uff0cSet C \u70ba\u5169 \u7a2e\u566a\u97f3 (Subway, Street) \u52a0\u4e0a\u901a\u9053\u5931\u771f\u3002 \u6211\u5011\u4f7f\u7528\u6b50\u6d32\u96fb\u4fe1\u6a19\u6e96\u5316\u5354\u6703 (European Telecommunications Standards Institute, ETSI) \u6240\u63d0\u51fa\u7528\u65bc\u9032\u884c\u5206\u6563\u5f0f\u8a9e\u97f3\u8fa8\u8b58\u7684 AFE (Advanced Front-End)\uff0c\u505a\u70ba\u5be6\u9a57\u7528\u7684\u7279 \u5fb5\u3002\u97f3\u6846\u9577\u5ea6\u70ba 25 \u6beb\u79d2\uff0c\u97f3\u6846\u79fb\u52d5\u9577\u5ea6\u70ba 10 \u6beb\u79d2\u3002\u795e\u7d93\u7db2\u8def\u7684\u8a13\u7df4\u4f7f\u7528 13 \u7dad AFE \u52a0 \u4e0a\u5176\u4e00\u968e\u53ca\u4e8c\u968e\u52d5\u614b\u7279\u5fb5\uff0c\u4e26\u524d\u5f8c\u4e32\u63a5 5 \u500b\u97f3\u6846\uff0c\u8f38\u5165\u5411\u91cf\u5171 429 \u7dad\u3002HMM \u6211\u5011\u5b9a\u7fa9 \u975c\u97f3\u70ba 3 \u500b\u72c0\u614b\uff0c\u6578\u5b57\u7684\u8072\u97f3\u70ba 16 \u500b\u72c0\u614b\uff0c\u5171\u6709 179 \u500b\u72c0\u614b\u3002 \u5728\u5be6\u9a57\u4e2d\uff0c\u985e\u795e\u7d93\u5b50\u7db2\u8def\u6211\u5011\u4f7f\u7528\uff11\u5c64\u96b1\u85cf\u5c64\uff0c\u4e00\u5c64\u6709 2560 \u500b\u795e\u7d93\u5143\u3002\u8a13\u7df4\u4f7f\u7528 dropout[16]\u4ee5\u907f\u514d overfitting\u3002\u6b64\u5916\uff0cdropout rate \u70ba 0.8\uff1b\u8a73\u7d30\u7684\u5be6\u9a57\u8a2d\u5b9a\u53ef\u53c3\u8003[17]\u3002" }, "TABREF3": { "num": null, "content": "
\u6599\u5e73\u5747\u7684\u6548\u679c\u3002
\u8868\u4e09\u3001\u7dda\u6027\u7d44\u5408\u6cd5\u8207 baseline \u6bd4\u8f03
Set ASet BSet CAvg.
Baseline4.655.835.285.25
Linear Combination4.785.815.485.33
\u5728\u9032\u884c\u8a9e\u97f3\u8fa8\u8b58\u7684\u5be6\u9a57\u524d\uff0c\u6211\u5011\u9996\u5148\u6e2c\u8a66\u4f7f\u7528 GMM \u4f86\u9032\u884c\u6a21\u578b\u9078\u64c7\u7684\u80fd\u529b\u3002\u5728\u8868
\u56db\u7684\u5be6\u9a57\u4e2d\uff0c\u5206\u5225\u70ba GMM componentsTest Error Rate
GMM647.8
GMM1287.3
\u8868\u4e94\u3001\u672c\u6587\u65b9\u6cd5\u8207 baseline \u6bd4\u8f03
Set ASet BSet CAvg.
Baseline4.655.835.285.25
Proposed method4.395.415.104.94
11 )
\u5247\u6211\u5011\u53ef\u4ee5\u4f7f\u7528 w \u7dda\u6027\u7d44\u5408\u6574\u9ad4\u6a21\u578b\u53ca\u5b50\u96c6\u6a21\u578b\u7684\u8f38\u51fa\uff0c\u5176\u8fa8\u8b58\u7d50\u679c\u5982\u8868\u4e09\u3002\u5f9e\u7d50\u679c\u53ef
\u4ee5\u770b\u51fa\u5176\u6548\u679c\u660e\u986f\u4f4e\u65bc baseline \u7cfb\u7d71\uff0c\u6211\u5011\u63a8\u6e2c\u539f\u56e0\u70ba\u5c0d\u65bc\u6bcf\u7b46\u6e2c\u8a66\u8cc7\u6599\u90fd\u4f7f\u7528\u540c\u4e00\u7d44
\u52a0\u6b0a\u503c\u9032\u884c\u7d44\u5408\uff0c\u6c92\u6709\u8003\u616e\u5230\u6bcf\u7b46\u6e2c\u8a66\u8cc7\u6599\u7684\u7368\u7279\u6027\uff0c\u6574\u500b\u7cfb\u7d71\u53ea\u6703\u5f97\u5230\u5c0d\u65bc\u5404\u985e\u578b\u8cc7
", "html": null, "type_str": "table", "text": "GMM \u5206\u5225\u6709 64 \u500b\u8207 128 \u500b\u9ad8\u65af\u6210\u5206\u7684\u6027\u5225\u8fa8\u8b58\u932f\u8aa4\u7387\u3002\u7531\u7d50\u679c \u53ef\u4ee5\u770b\u51fa\u4f7f\u7528 128 \u500b\u9ad8\u65af\u6210\u5206\u7684\u932f\u8aa4\u7387\u8f03\u4f4e\uff0c\u800c\u4e14\u4e5f\u6709\u8457\u4e0d\u932f\u7684\u8fa8\u8b58\u7387\uff0c\u56e0\u6b64\u5728\u5f8c\u9762\u7684 \u5be6\u9a57\u4e2d\uff0c\u6211\u5011\u4f7f\u7528 128 \u500b\u9ad8\u65af\u6210\u5206\u7684 GMM \u4f86\u9032\u884c\u985e\u795e\u7d93\u7db2\u8def\u7684\u9078\u64c7\u3002 \u8868\u4e94\u6bd4\u8f03\u672c\u6587\u63d0\u51fa\u7684\u5f37\u5065\u6027\u8a9e\u97f3\u8fa8\u8b58\u7cfb\u7d71\u8207 baseline \u8fa8\u8b58\u7cfb\u7d71\u7684\u7cfb\u7d71\u8fa8\u8b58\u6548\u80fd\uff0c\u5728 \u4e09\u500b\u6e2c\u8a66\u5b50\u96c6\u4e2d\u3002\u53ef\u4ee5\u770b\u51fa\u5728\u4e09\u500b\u6e2c\u8a66\u5b50\u96c6\u7684\u90e8\u5206\uff0c\u672c\u6587\u63d0\u51fa\u7684\u8fa8\u8b58\u7cfb\u7d71\uff0c\u8a5e\u932f\u8aa4\u7387\u76f8 \u8f03\u65bc baseline \u90fd\u6709\u660e\u986f\u7684\u4e0b\u964d\uff0c\u5e73\u5747\u7684\u8a5e\u932f\u8aa4\u7387\u5247\u964d\u4f4e\u4e86 5.9% (\u5f9e 5.25 \u5230 4.94)\uff0c\u6211\u5011 \u76f8\u4fe1\u6b64\u8fa8\u8b58\u7d50\u679c\u652f\u6301\u4f9d\u64da\u8072\u5b78\u7d50\u6027\u5207\u5272\u8a13\u7df4\u8a9e\u6599\u5eab\uff0c\u4e26\u5728\u6e2c\u8a66\u4e2d\u9078\u64c7\u8f03\u4f73\u7684\u8072\u5b78\u6a21\u578b\u505a \u70ba\u8f38\u51fa\uff0c\u5373\u80fd\u9069\u7576\u7684\u63d0\u5347\u8a9e\u97f3\u8fa8\u8b58\u7cfb\u7d71\u7684\u6548\u80fd\u4e26\u5f37\u5065\u8a9e\u97f3\u8fa8\u8b58\u7cfb\u7d71\u3002 \u4e94\u3001\u7d50\u8ad6 \u5728\u6b64\u7bc7\u8ad6\u6587\u4e2d\uff0c\u6211\u5011\u63d0\u51fa\u57fa\u65bc EC \u53ca Mixture of Experts \u7684\u67b6\u69cb\u4f86\u8a13\u7df4\u795e\u7d93\u7db2\u8def\uff1b\u4f9d \u64da\u8a13\u7df4\u8a9e\u6599\u4e0d\u540c\u7684\u8072\u5b78\u7279\u6027\uff0c\u5207\u5272\u4e26\u4ee5\u985e\u795e\u7d93\u7db2\u8def\u8207 GMM \u6a21\u578b\u5316\u4e0d\u540c\u7684\u8072\u5b78\u6a21\u578b\uff1b\u5728 \u6e2c\u8a66\u6642\uff0c\u5c07 \u6e2c\u8a66\u8a9e\u6599\u7d93\u7531 GMM-gate \u5f97\u5230\u5c0d\u6bcf\u500b\u8072\u5b78\u6a21\u578b\u7684\u5f8c\u9a57\u6a5f\u7387\uff0c\u9078\u64c7\u6700\u4f73\u7684\u8072\u5b78 \u6a21\u578b\u505a\u70ba\u8fa8\u8b58\u7cfb\u7d71\u7684\u57fa\u790e\u3002\u5be6\u9a57\u4e0a\uff0c\u6211\u5011\u4ee5 Aurora 2 \u505a\u70ba\u5be6\u9a57\u7684\u8a9e\u6599\u5eab\uff0c\u5c07\u8a13\u7df4\u8a9e\u6599\u4f9d \u64da\u6027\u5225\u4ee5\u53ca SNR \u7684\u65b9\u5f0f\u5207\u5272\u8a13\u7df4\u8a9e\u6599\uff0c\u4e26\u6bd4\u8f03\u4e86\u50b3\u7d71\u4f7f\u7528 DNN-HMM \u67b6\u69cb\u8207\u672c\u6587\u63d0\u51fa \u7684\u5f37\u5065\u6027\u8a9e\u97f3\u8fa8\u8b58\u7cfb\u7d71\u3002\u6211\u5011\u63d0\u51fa\u7684\u8a9e\u97f3\u8fa8\u8b58\u7cfb\u7d71\u80fd\u63d0\u5347\u50b3\u7d71\u7684\u8a9e\u97f3\u8fa8\u8b58\u7cfb\u7d71\u9054 5.9%\u3002\u672a\u4f86\u6211\u5011\u5c07\u63a2\u8a0e\u4e0d\u540c\u7684\u8072\u5b78\u7279\u6027\u6a21\u578b\u8207\u4e0d\u540c\u7684 gate function\uff0c\u4e26\u5617\u8a66\u5728\u5927\u8a5e\u5f59\u8a9e \u6599\u5eab\u4e2d\u3002 \u8868\u56db\u3001GMM \u6027\u5225\u8fa8\u8b58\u4e4b\u7d50\u679c" } } } }