{ "paper_id": "O13-1014", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T08:04:07.475604Z" }, "title": "Improved Modulation Spectrum Histogram Equalization for Robust Speech Recognition", "authors": [ { "first": "\u9ad8\u4e88\u771f", "middle": [], "last": "\u9673\u67cf\u7433", "suffix": "", "affiliation": { "laboratory": "", "institution": "National Taiwan Normal University", "location": {} }, "email": "" }, { "first": "Yu -Chen", "middle": [], "last": "Kao", "suffix": "", "affiliation": { "laboratory": "", "institution": "National Taiwan Normal University", "location": {} }, "email": "" }, { "first": "Berlin", "middle": [], "last": "Chen", "suffix": "", "affiliation": { "laboratory": "", "institution": "National Taiwan Normal University", "location": {} }, "email": "berlin@ntnu.edu.tw" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "", "pdf_parse": { "paper_id": "O13-1014", "_pdf_hash": "", "abstract": [], "body_text": [ { "text": "\u76ee\u524d\u7684\u81ea\u52d5\u8a9e\u97f3\u8fa8\u8b58(automatic speech recognition, ASR)\u7cfb\u7d71\uff0c\u5728\u4e0d\u53d7\u5404\u7a2e\u74b0\u5883\u8b8a\u56e0\u5e72\u64fe \u7684\u7406\u60f3\u9304\u97f3\u74b0\u5883\u4e0b\uff0c\u53ef\u4ee5\u5f97\u5230\u76f8\u7576\u512a\u79c0\u7684\u8fa8\u8b58\u6548\u679c\uff1b\u4f46\u5728\u5be6\u52d9\u61c9\u7528\u4e0a\uff0c\u8a9e\u8005\u7684\u5dee\u7570\u3001\u9304 \u97f3\u904e\u7a0b\u7522\u751f\u7684\u566a\u97f3\u3001\u5176\u4ed6\u74b0\u5883\u8072\u97ff\u53ca\u901a\u9053\u6548\u61c9(channel effect)\u7b49\u74b0\u5883\u4e0a\u7684\u8b8a\u56e0\uff0c\u6703\u4f7f\u8a13 \u7df4\u74b0\u5883\u548c\u6e2c\u8a66\u74b0\u5883\u9593\u7522\u751f\u74b0\u5883\u4e0d\u5339\u914d(environmental mismatch)\u7684\u554f\u984c\uff0c\u5728\u672c\u8ad6\u6587\u4e2d\u4e5f\u7a31 \u70ba\u96dc\u8a0a(noise) \u3002 \u96dc \u8a0a \u53ef \u4ee5 \u7c97 \u7565 \u5730 \u5206 \u6210 \u52a0 \u6210 \u6027 \u566a \u97f3 (additive noise) \u53ca\u647a\u7a4d\u6027\u566a\u97f3 (convolutional noise)\uff1a\u52a0\u6210\u6027\u566a\u97f3\u5373\u9664\u4e86\u5be6\u969b\u6240\u9700\u7684\u8a9e\u97f3\u8a0a\u865f\u5916\uff0c\u7cfb\u7d71\u6240\u63a5\u6536\u5230\u7684\u5176\u4ed6 \u8072\u97f3\uff0c\u5176\u5728\u6642\u57df(time domain)\u53ca\u983b\u57df(spectrum domain)\u4e0a\u8207\u539f\u8a9e\u97f3\u8a0a\u865f\u662f\u76f8\u52a0\u7684\u95dc\u4fc2\uff0c \u56e0\u800c\u5f97\u540d\uff1b\u647a\u7a4d\u6027\u566a\u97f3\u53c8\u7a31\u70ba\u901a\u9053\u6548\u61c9(channel effect)\uff0c\u662f\u8a9e\u97f3\u5f9e\u767c\u8072\u5230\u63a5\u6536\u7684\u904e\u7a0b\u4e2d \u7d93\u904e\u7684\u5404\u7a2e\u5be6\u9ad4\u4ecb\u8cea\u53ca\u96fb\u5b50\u8a2d\u5099\u6240\u9020\u6210\u7684\u626d\u66f2\uff0c\u5728\u6642\u57df\u4e0a\u8207\u539f\u8a9e\u97f3\u8a0a\u865f\u70ba\u647a\u7a4d (convolution)\u7684\u95dc\u4fc2\uff0c\u800c\u5728\u983b\u57df\u4e0a\u5247\u8207\u539f\u8a9e\u97f3\u8a0a\u865f\u70ba\u76f8\u4e58\u7684\u95dc\u4fc2\u3002 \u4eba\u8033\u5c0d\u96dc\u8a0a\u6709\u975e\u5e38\u512a\u826f\u7684\u5f37\u5065\u6027(robustness)\uff0c\u9019\u4e9b\u96dc\u8a0a\u5c0d\u4eba\u8033\u7684\u5f71\u97ff\u4e26\u4e0d\u5927\uff1b\u4f46\u5c0d \u65bc\u81ea\u52d5\u8a9e\u97f3\u8fa8\u8b58\u7cfb\u7d71\u800c\u8a00\uff0c\u9019\u6a23\u7684\u4e0d\u5339\u914d\u6703\u4f7f\u8a9e\u97f3\u8fa8\u8b58\u7684\u6b63\u78ba\u7387(recognition accuracy) \u5927\u8f3b\u964d\u4f4e\uff0c\u9700\u8981\u63a1\u7528\u82e5\u5e72\u5f37\u5065\u6027\u8a9e\u97f3\u8fa8\u8b58(robust speech recognition)\u6280\u8853\u6e1b\u5c11\u74b0\u5883\u4e0d\u5339\u914d \u6240\u9020\u6210\u7684\u5f71\u97ff\uff0c\u4f7f\u81ea\u52d5\u8a9e\u97f3\u8fa8\u8b58\u5728\u4e0d\u540c\u7684\u74b0\u5883\u4e0b\u4ecd\u80fd\u4fdd\u6709\u4e00\u5b9a\u7684\u8fa8\u8b58\u6b63\u78ba\u7387\u3002\u5f37\u5065\u6027\u8a9e \u97f3\u8fa8\u8b58\u6280\u8853\u4f9d\u5176\u7279\u6027\u53ef\u4ee5\u5927\u81f4\u5206\u70ba\u4e09\u5927\u985e\u578b [1, 2] ", "cite_spans": [ { "start": 698, "end": 701, "text": "[1,", "ref_id": "BIBREF0" }, { "start": 702, "end": 704, "text": "2]", "ref_id": "BIBREF1" } ], "ref_spans": [], "eq_spans": [], "section": "\u4e00\u3001\u7dd2\u8ad6", "sec_num": null }, { "text": "\u0302P HEQ , -= \u22121 ( ( , -)) = \u2211 ( ( , -)) =0 (4) (\u4e09)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "\u4e00\u3001\u7dd2\u8ad6", "sec_num": null }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "\u5716\u4e8c\u3001ST-PSHE \u6d41\u7a0b\u793a\u610f\u5716 * , -+\u4e4b\u8abf\u8b8a\u983b\u8b5c\u5f37\u5ea6*| , -|+\u7684\u6a5f\u7387\u5206\u4f48\uff0c (\u2022)\u70ba\u6240\u6709\u8a13\u7df4\u8a9e\u6599\u7684\u8abf\u8b8a\u983b\u8b5c\u5f37\u5ea6\u6a5f\u7387\u5206 \u4f48\uff0c\u4e5f\u5c31\u662f\u53c3\u8003\u5206\u4f48\uff0c\u6b64\u65b9\u6cd5\u4e2d\u6b63\u898f\u5316\u5f8c\u7684\u983b\u8b5c\u5f37\u5ea6| , -|\u8207\u539f\u59cb\u983b\u8b5c\u5f37\u5ea6| , -|\u7684\u95dc \u4fc2\u70ba\uff1a | , -| SHE = \u22121 ( (| , -|))", "eq_num": "(6)" } ], "section": "\u4e00\u3001\u7dd2\u8ad6", "sec_num": null }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "\u6642 \u57df t,hp , -= { , - , if = 1 , -\u2212 , \u2212 1- 2 , otherwise (10) t,lp , -= { 0 , if = 1 , -+ , \u2212 1- 2 , otherwise", "eq_num": "(11)" } ], "section": "\u4e00\u3001\u7dd2\u8ad6", "sec_num": null }, { "text": "\u5176\u4e2d , -\u70ba\u8a72\u8a9e\u53e5\u4e2d\u7b2c \u500b\u97f3\u6846\u7b2c \u7dad\u5ea6\u7684\u8a9e\u97f3\u7279\u5fb5\u503c\uff0c = 1\u53ca = 1\u4ee3\u8868\u7b2c\u4e00\u500b\u97f3\u6846\u53ca \u7b2c\u4e00\u500b\u7dad\u5ea6\uff0c\u4f9d\u6b64\u985e\u63a8\uff1b ,hp , -\u3001 ,lp , -\u3001 t,hp , -\u53ca t,lp ", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "\u4e00\u3001\u7dd2\u8ad6", "sec_num": null }, { "text": "Proceedings of the Twenty-Fifth Conference on Computational Linguistics and Speech Processing(ROCLING 2013)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null } ], "back_matter": [], "bib_entries": { "BIBREF0": { "ref_id": "b0", "title": "Environmental robustness", "authors": [ { "first": "J", "middle": [], "last": "Droppo", "suffix": "" }, { "first": "A", "middle": [], "last": "Acero", "suffix": "" } ], "year": 2008, "venue": "Springer handbook of speech processing", "volume": "33", "issue": "", "pages": "653--679", "other_ids": {}, "num": null, "urls": [], "raw_text": "J. Droppo and A. Acero, \"Environmental robustness,\" in Springer handbook of speech processing, 1st ed., J. Benesty, M. M. Sondhi, and Y. Huang, Eds. Springer, 2008, ch. 33, pp. 653-679.", "links": null }, "BIBREF1": { "ref_id": "b1", "title": "Speech recognition in noisy environments: a survey", "authors": [ { "first": "Y", "middle": [], "last": "Gong", "suffix": "" } ], "year": 1995, "venue": "Speech Communication", "volume": "16", "issue": "3", "pages": "261--291", "other_ids": {}, "num": null, "urls": [], "raw_text": "Y. Gong, \"Speech recognition in noisy environments: a survey,\" Speech Communication, vol. 16, no. 3, pp. 261-291, 1995.", "links": null }, "BIBREF2": { "ref_id": "b2", "title": "Maximum likelihood linear regression for speaker adaptation of continuous density hidden markov models", "authors": [ { "first": "C", "middle": [ "J" ], "last": "Leggetter", "suffix": "" }, { "first": "P", "middle": [ "C" ], "last": "Woodland", "suffix": "" } ], "year": 1995, "venue": "Computer Speech & Language", "volume": "9", "issue": "2", "pages": "171--185", "other_ids": {}, "num": null, "urls": [], "raw_text": "C. J. Leggetter and P. C. Woodland, \"Maximum likelihood linear regression for speaker adaptation of continuous density hidden markov models,\" Computer Speech & Language, vol. 9, no. 2, pp. 171-185, 1995.", "links": null }, "BIBREF3": { "ref_id": "b3", "title": "Model based techniques for noise robust speech recognition", "authors": [ { "first": "M", "middle": [ "J" ], "last": "Gales", "suffix": "" } ], "year": 1995, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "M. J. Gales, \"Model based techniques for noise robust speech recognition,\" Ph.D. dissertation, Cambridge University, 1995.", "links": null }, "BIBREF4": { "ref_id": "b4", "title": "A vector taylor series approach for environment-independent speech recognition", "authors": [ { "first": "P", "middle": [], "last": "Moreno", "suffix": "" }, { "first": "B", "middle": [], "last": "Raj", "suffix": "" }, { "first": "R", "middle": [], "last": "Stern", "suffix": "" } ], "year": 1996, "venue": "Proc. IEEE Int. Conf. on Acoustics, Speech, Signal Processing", "volume": "2", "issue": "", "pages": "733--736", "other_ids": {}, "num": null, "urls": [], "raw_text": "P. Moreno, B. Raj, and R. Stern, \"A vector taylor series approach for environment-independent speech recognition,\" in Proc. IEEE Int. Conf. on Acoustics, Speech, Signal Processing, vol. 2, 1996, pp. 733 -736.", "links": null }, "BIBREF5": { "ref_id": "b5", "title": "Suppression of acoustic noise in speech using spectral subtraction", "authors": [ { "first": "S", "middle": [], "last": "Boll", "suffix": "" } ], "year": 1979, "venue": "IEEE Transactions on Acoustics, Speech and Signal Processing", "volume": "27", "issue": "2", "pages": "113--120", "other_ids": {}, "num": null, "urls": [], "raw_text": "S. Boll, \"Suppression of acoustic noise in speech using spectral subtraction,\" IEEE Transactions on Acoustics, Speech and Signal Processing, vol. 27, no. 2, pp. 113-120, 1979.", "links": null }, "BIBREF6": { "ref_id": "b6", "title": "A signal subspace approach for speech enhancement", "authors": [ { "first": "Y", "middle": [], "last": "Ephraim", "suffix": "" }, { "first": "H", "middle": [], "last": "Van Trees", "suffix": "" } ], "year": 1995, "venue": "IEEE Transactions on Speech and Audio Processing", "volume": "3", "issue": "4", "pages": "251--266", "other_ids": {}, "num": null, "urls": [], "raw_text": "Y. Ephraim and H. Van Trees, \"A signal subspace approach for speech enhancement,\" IEEE Transactions on Speech and Audio Processing, vol. 3, no. 4, pp. 251-266, 1995.", "links": null }, "BIBREF7": { "ref_id": "b7", "title": "Enhancement and bandwidth compression of noisy speech", "authors": [ { "first": "J", "middle": [ "S" ], "last": "Lim", "suffix": "" }, { "first": "A", "middle": [ "V" ], "last": "Oppenheim", "suffix": "" } ], "year": 1979, "venue": "Proceedings of the IEEE", "volume": "67", "issue": "12", "pages": "1586--1604", "other_ids": {}, "num": null, "urls": [], "raw_text": "J. S. Lim and A. V. Oppenheim, \"Enhancement and bandwidth compression of noisy speech,\" Proceedings of the IEEE, vol. 67, no. 12, pp. 1586-1604, 1979.", "links": null }, "BIBREF8": { "ref_id": "b8", "title": "Speech enhancement using a minimum mean-square error log-spectral amplitude estimator", "authors": [ { "first": "Y", "middle": [], "last": "Ephraim", "suffix": "" }, { "first": "D", "middle": [], "last": "Malah", "suffix": "" } ], "year": 1985, "venue": "IEEE Transactions on Acoustics, Speech and Signal Processing", "volume": "33", "issue": "2", "pages": "443--445", "other_ids": {}, "num": null, "urls": [], "raw_text": "Y. Ephraim and D. Malah, \"Speech enhancement using a minimum mean-square error log-spectral amplitude estimator,\" IEEE Transactions on Acoustics, Speech and Signal Processing, vol. 33, no. 2, pp. 443-445, 1985.", "links": null }, "BIBREF9": { "ref_id": "b9", "title": "Low distortion speech enhancement", "authors": [ { "first": "I", "middle": [], "last": "Soon", "suffix": "" }, { "first": "S", "middle": [], "last": "Koh", "suffix": "" } ], "year": 2000, "venue": "IEEE Proceedings of Vision, Image and Signal Processing", "volume": "147", "issue": "", "pages": "247--253", "other_ids": {}, "num": null, "urls": [], "raw_text": "I. Soon and S. Koh, \"Low distortion speech enhancement,\" IEEE Proceedings of Vision, Image and Signal Processing, vol. 147, no. 3, pp. 247-253, 2000.", "links": null }, "BIBREF10": { "ref_id": "b10", "title": "Non-linear transformations of the feature space for robust speech recognition", "authors": [ { "first": "A", "middle": [], "last": "De La Torre", "suffix": "" }, { "first": "J", "middle": [ "C" ], "last": "Segura", "suffix": "" }, { "first": "C", "middle": [], "last": "Benitez", "suffix": "" }, { "first": "A", "middle": [ "M" ], "last": "Peinado", "suffix": "" }, { "first": "A", "middle": [ "J" ], "last": "Rubio", "suffix": "" } ], "year": 2002, "venue": "Proc. IEEE Int. Conf. on Acoustics, Speech, Signal Processing", "volume": "", "issue": "", "pages": "401--404", "other_ids": {}, "num": null, "urls": [], "raw_text": "A. de la Torre, J. C. Segura, C. Benitez, A. M. Peinado, and A. J. Rubio, \"Non-linear transformations of the feature space for robust speech recognition,\" in Proc. IEEE Int. Conf. on Acoustics, Speech, Signal Processing, 2002, pp. 401-404.", "links": null }, "BIBREF11": { "ref_id": "b11", "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, vol. 29, no. 2, pp. 254-272, 1981.", "links": null }, "BIBREF12": { "ref_id": "b12", "title": "Cepstral domain segmental feature vector normalization for noise robust speech recognition", "authors": [ { "first": "O", "middle": [], "last": "Viikki", "suffix": "" }, { "first": "K", "middle": [], "last": "Laurila", "suffix": "" } ], "year": 1998, "venue": "Speech Communucation", "volume": "25", "issue": "1-3", "pages": "133--147", "other_ids": {}, "num": null, "urls": [], "raw_text": "O. Viikki and K. Laurila, \"Cepstral domain segmental feature vector normalization for noise robust speech recognition,\" Speech Communucation, vol. 25, no. 1-3, pp. 133- 147, 1998.", "links": null }, "BIBREF13": { "ref_id": "b13", "title": "Large-vocabulary speech recognition under adverse acoustic environments", "authors": [ { "first": "L", "middle": [], "last": "Deng", "suffix": "" }, { "first": "A", "middle": [], "last": "Acero", "suffix": "" }, { "first": "M", "middle": [], "last": "Plumpe", "suffix": "" }, { "first": "X", "middle": [], "last": "Huang", "suffix": "" } ], "year": 2000, "venue": "Proc. Int. Conf. on Spoken Language Processing", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "L. Deng, A. Acero, M. Plumpe, and X. Huang, \"Large-vocabulary speech recognition under adverse acoustic environments,\" in Proc. Int. Conf. on Spoken Language Processing, 2000.", "links": null }, "BIBREF14": { "ref_id": "b14", "title": "An environment compensated maximum likelihood Proceedings of the Twenty-Fifth Conference on Computational Linguistics and Speech Processing (ROCLING 2013) training approach based on stochastic vector mapping", "authors": [ { "first": "J", "middle": [], "last": "Wu", "suffix": "" }, { "first": "Q", "middle": [], "last": "Huo", "suffix": "" }, { "first": "D", "middle": [], "last": "Zhu", "suffix": "" } ], "year": 2005, "venue": "Proc. IEEE Int. Conf. on Acoustics, Speech, Signal Processing", "volume": "1", "issue": "", "pages": "429--432", "other_ids": {}, "num": null, "urls": [], "raw_text": "J. Wu, Q. Huo, and D. Zhu, \"An environment compensated maximum likelihood Proceedings of the Twenty-Fifth Conference on Computational Linguistics and Speech Processing (ROCLING 2013) training approach based on stochastic vector mapping,\" in Proc. IEEE Int. Conf. on Acoustics, Speech, Signal Processing, vol. 1, 2005, pp. 429-432.", "links": null }, "BIBREF15": { "ref_id": "b15", "title": "Higher order cepstral moment normalization for improved robust speech recognition", "authors": [ { "first": "C.-W", "middle": [], "last": "Hsu", "suffix": "" }, { "first": "L.-S", "middle": [], "last": "Lee", "suffix": "" } ], "year": 2009, "venue": "IEEE Transactions on Audio, Speech, and Language Processing", "volume": "17", "issue": "2", "pages": "205--220", "other_ids": {}, "num": null, "urls": [], "raw_text": "C.-W. Hsu and L.-S. Lee, \"Higher order cepstral moment normalization for improved robust speech recognition,\" IEEE Transactions on Audio, Speech, and Language Processing, vol. 17, no. 2, pp. 205-220, 2009.", "links": null }, "BIBREF16": { "ref_id": "b16", "title": "Distribution-based feature compensation for robust speech recognition", "authors": [ { "first": "B", "middle": [], "last": "Chen", "suffix": "" }, { "first": "S.-H", "middle": [], "last": "Lin", "suffix": "" } ], "year": 2011, "venue": "Recent Advances in Robust Speech Recognition Technology", "volume": "", "issue": "", "pages": "155--168", "other_ids": {}, "num": null, "urls": [], "raw_text": "B. Chen and S.-H. Lin, \"Distribution-based feature compensation for robust speech recognition,\" in Recent Advances in Robust Speech Recognition Technology. Bentham Science Publishers, 2011, ch. 10, pp. 155-168.", "links": null }, "BIBREF17": { "ref_id": "b17", "title": "Histogram equalization of speech representation for robust speech recognition", "authors": [ { "first": "A", "middle": [], "last": "De La Torre", "suffix": "" }, { "first": "A", "middle": [ "M" ], "last": "Peinado", "suffix": "" }, { "first": "J", "middle": [ "C" ], "last": "Segura", "suffix": "" }, { "first": "J", "middle": [ "L" ], "last": "Perez-Cordoba", "suffix": "" }, { "first": "M", "middle": [ "C" ], "last": "Benitez", "suffix": "" }, { "first": "A", "middle": [ "J" ], "last": "Rubio", "suffix": "" } ], "year": 2005, "venue": "IEEE Transactions on Speech and Audio Processing", "volume": "13", "issue": "3", "pages": "355--366", "other_ids": {}, "num": null, "urls": [], "raw_text": "A. de la Torre, A. M. Peinado, J. C. Segura, J. L. Perez-Cordoba, M. C. Benitez, and A. J. Rubio, \"Histogram equalization of speech representation for robust speech recognition,\" IEEE Transactions on Speech and Audio Processing, vol. 13, no. 3, pp. 355-366, 2005.", "links": null }, "BIBREF18": { "ref_id": "b18", "title": "A nonlinear unsupervised adaptation technique for speech recognition", "authors": [ { "first": "D", "middle": [ "P" ], "last": "Ibm", "suffix": "" }, { "first": "S", "middle": [], "last": "Dharanipragada", "suffix": "" }, { "first": "M", "middle": [], "last": "Padmanabhan", "suffix": "" } ], "year": 2000, "venue": "Proc. Int. Conf. on Spoken Language Processing", "volume": "", "issue": "", "pages": "556--559", "other_ids": {}, "num": null, "urls": [], "raw_text": "D. P. Ibm, S. Dharanipragada, and M. Padmanabhan, \"A nonlinear unsupervised adaptation technique for speech recognition,\" in Proc. Int. Conf. on Spoken Language Processing, 2000, pp. 556-559.", "links": null }, "BIBREF19": { "ref_id": "b19", "title": "Robust speech recognition using spatial-temporal feature distribution characteristics", "authors": [ { "first": "B", "middle": [], "last": "Chen", "suffix": "" }, { "first": "W.-H", "middle": [], "last": "Chen", "suffix": "" }, { "first": "S.-H", "middle": [], "last": "Lin", "suffix": "" }, { "first": "W.-Y", "middle": [], "last": "Chu", "suffix": "" } ], "year": 2011, "venue": "Pattern Recognition Letter", "volume": "32", "issue": "7", "pages": "919--926", "other_ids": {}, "num": null, "urls": [], "raw_text": "B. Chen, W.-H. Chen, S.-H. Lin, and W.-Y. Chu, \"Robust speech recognition using spatial-temporal feature distribution characteristics,\" Pattern Recognition Letter, vol. 32, no. 7, pp. 919-926, 2011.", "links": null }, "BIBREF20": { "ref_id": "b20", "title": "Filtering on the temporal probability sequence in histogram equalization for robust speech recognition", "authors": [ { "first": "S.-S", "middle": [], "last": "Wang", "suffix": "" }, { "first": "Y", "middle": [], "last": "Tsao", "suffix": "" }, { "first": "J.-W", "middle": [], "last": "Hung", "suffix": "" } ], "year": 2013, "venue": "Proc. IEEE Int. Conf. on Acoustics, Speech, Signal Processing", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "S.-S. Wang, Y. Tsao, and J.-W. Hung, \"Filtering on the temporal probability sequence in histogram equalization for robust speech recognition,\" in Proc. IEEE Int. Conf. on Acoustics, Speech, Signal Processing, 2013.", "links": null }, "BIBREF21": { "ref_id": "b21", "title": "Exploring joint equalization of spatial-temporal contextual statistics of speech features for robust speech recognition", "authors": [ { "first": "H.-J", "middle": [], "last": "Hsieh", "suffix": "" }, { "first": "J.-W", "middle": [], "last": "Hung", "suffix": "" }, { "first": "B", "middle": [], "last": "Chen", "suffix": "" } ], "year": 2012, "venue": "Proc. Annu. Conf. of the Int. Speech Communication Association", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "H.-J. Hsieh, J.-W. Hung, and B. Chen, \"Exploring joint equalization of spatial-temporal contextual statistics of speech features for robust speech recognition,\" in Proc. Annu. Conf. of the Int. Speech Communication Association, 2012.", "links": null }, "BIBREF22": { "ref_id": "b22", "title": "Sub-band level histogram equalization for robust speech recognition", "authors": [ { "first": "V", "middle": [], "last": "Joshi", "suffix": "" }, { "first": "R", "middle": [], "last": "Biligi", "suffix": "" }, { "first": "U", "middle": [ "S" ], "last": "", "suffix": "" }, { "first": "L", "middle": [], "last": "Garcia", "suffix": "" }, { "first": "C", "middle": [], "last": "Benitez", "suffix": "" } ], "year": 2011, "venue": "Proc. Annu. Conf. of the Int. Speech Communication Association", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "V. Joshi, R. Biligi, U. S., L. Garcia, and C. Benitez, \"Sub-band level histogram equalization for robust speech recognition,\" in Proc. Annu. Conf. of the Int. Speech Communication Association, 2011.", "links": null }, "BIBREF23": { "ref_id": "b23", "title": "Speech enhancement based on physiological and psychoacoustical models of modulation perception and binaural interaction", "authors": [ { "first": "B", "middle": [], "last": "Kollmeier", "suffix": "" }, { "first": "R", "middle": [], "last": "Koch", "suffix": "" } ], "year": 1994, "venue": "The Journal of the Acoustical Society of America", "volume": "95", "issue": "3", "pages": "1593--1602", "other_ids": {}, "num": null, "urls": [], "raw_text": "B. Kollmeier and R. Koch, \"Speech enhancement based on physiological and psychoacoustical models of modulation perception and binaural interaction,\" The Journal of the Acoustical Society of America, vol. 95, no. 3, pp. 1593-1602, 1994.", "links": null }, "BIBREF24": { "ref_id": "b24", "title": "Sub-band modulation spectrum compensation for robust speech recognition", "authors": [ { "first": "W.-H", "middle": [], "last": "Tu", "suffix": "" }, { "first": "S.-Y", "middle": [], "last": "Huang", "suffix": "" }, { "first": "J.-W", "middle": [], "last": "Hung", "suffix": "" } ], "year": 2009, "venue": "Proc. IEEE Workshop on Automatic Speech Recognition Understanding", "volume": "", "issue": "", "pages": "261--265", "other_ids": {}, "num": null, "urls": [], "raw_text": "W.-H. Tu, S.-Y. Huang, and J.-W. Hung, \"Sub-band modulation spectrum compensation for robust speech recognition,\" in Proc. IEEE Workshop on Automatic Speech Recognition Understanding, 2009, pp. 261-265.", "links": null }, "BIBREF25": { "ref_id": "b25", "title": "Modulation spectrum equalization for robust speech recognition", "authors": [ { "first": "L.-C", "middle": [], "last": "Sun", "suffix": "" }, { "first": "C.-W", "middle": [], "last": "Hsu", "suffix": "" }, { "first": "L.-S", "middle": [], "last": "Lee", "suffix": "" } ], "year": 2007, "venue": "IEEE Workshop on Automatic Speech Recognition Understanding", "volume": "", "issue": "", "pages": "81--86", "other_ids": {}, "num": null, "urls": [], "raw_text": "L.-C. Sun, C.-W. Hsu, and L.-S. Lee, \"Modulation spectrum equalization for robust speech recognition,\" in IEEE Workshop on Automatic Speech Recognition Understanding, 2007, pp. 81-86.", "links": null }, "BIBREF26": { "ref_id": "b26", "title": "Improved modulation spectrum enhancement methods for robust speech recognition", "authors": [ { "first": "J.-W", "middle": [], "last": "Hung", "suffix": "" }, { "first": "W.-H", "middle": [], "last": "Tu", "suffix": "" }, { "first": "C.-C", "middle": [], "last": "Lai", "suffix": "" } ], "year": 2012, "venue": "Signal Processing", "volume": "92", "issue": "11", "pages": "2791--2814", "other_ids": {}, "num": null, "urls": [], "raw_text": "J.-W. Hung, W.-H. Tu, and C.-C. Lai, \"Improved modulation spectrum enhancement methods for robust speech recognition,\" Signal Processing, vol. 92, no. 11, pp. 2791- 2814, 2012.", "links": null }, "BIBREF27": { "ref_id": "b27", "title": "Normalization of the speech modulation spectra for robust speech recognition", "authors": [ { "first": "X", "middle": [], "last": "Xiao", "suffix": "" }, { "first": "E", "middle": [ "S" ], "last": "Chng", "suffix": "" }, { "first": "H", "middle": [], "last": "Li", "suffix": "" } ], "year": 2008, "venue": "IEEE Transactions on Audio, Speech, and Language Processing", "volume": "16", "issue": "8", "pages": "1662--1674", "other_ids": {}, "num": null, "urls": [], "raw_text": "X. Xiao, E. S. Chng, and H. Li, \"Normalization of the speech modulation spectra for robust speech recognition,\" IEEE Transactions on Audio, Speech, and Language Processing, vol. 16, no. 8, pp. 1662-1674, 2008.", "links": null }, "BIBREF28": { "ref_id": "b28", "title": "Leveraging distributional characteristics of modulation spectra for robust speech recognition", "authors": [ { "first": "Y. -C", "middle": [], "last": "Kao", "suffix": "" }, { "first": "B", "middle": [], "last": "Chen", "suffix": "" } ], "year": 2012, "venue": "Proc. Int. Conf. on Information Science", "volume": "", "issue": "", "pages": "120--125", "other_ids": {}, "num": null, "urls": [], "raw_text": "Y. -C. Kao and B. Chen, \"Leveraging distributional characteristics of modulation spectra for robust speech recognition,\" in Proc. Int. Conf. on Information Science, Signal Processing and their Applications, 2012, pp. 120-125.", "links": null }, "BIBREF29": { "ref_id": "b29", "title": "Modulation spectrum equalization for improved robust speech recognition", "authors": [ { "first": "L.-C", "middle": [], "last": "Sun", "suffix": "" }, { "first": "L.-S", "middle": [], "last": "Lee", "suffix": "" } ], "year": 2012, "venue": "IEEE Transactions on Audio, Speech, and Language Processing", "volume": "20", "issue": "3", "pages": "828--843", "other_ids": {}, "num": null, "urls": [], "raw_text": "L.-C. Sun and L.-S. Lee, \"Modulation spectrum equalization for improved robust speech recognition,\" IEEE Transactions on Audio, Speech, and Language Processing, vol. 20, no. 3, pp. 828-843, 2012.", "links": null }, "BIBREF30": { "ref_id": "b30", "title": "Image Processing: Principles and Applications", "authors": [ { "first": "T", "middle": [], "last": "Acharya", "suffix": "" }, { "first": "A", "middle": [], "last": "Ray", "suffix": "" } ], "year": 2005, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "T. Acharya and A. Ray, Image Processing: Principles and Applications. Wiley, 2005.", "links": null }, "BIBREF31": { "ref_id": "b31", "title": "Quantile based histogram equalization for noise robust large vocabulary speech recognition", "authors": [ { "first": "F", "middle": [], "last": "Hilger", "suffix": "" }, { "first": "H", "middle": [], "last": "Ney", "suffix": "" } ], "year": 2006, "venue": "IEEE Transactions on Audio, Speech, and Language Processing", "volume": "14", "issue": "3", "pages": "845--854", "other_ids": {}, "num": null, "urls": [], "raw_text": "F. Hilger and H. Ney, \"Quantile based histogram equalization for noise robust large vocabulary speech recognition,\" IEEE Transactions on Audio, Speech, and Language Processing, vol. 14, no. 3, pp. 845-854, 2006.", "links": null }, "BIBREF32": { "ref_id": "b32", "title": "Exploring the use of speech features and their corresponding distribution characteristics for robust speech recognition", "authors": [ { "first": "S.-H", "middle": [], "last": "Lin", "suffix": "" }, { "first": "B", "middle": [], "last": "Chen", "suffix": "" }, { "first": "Y.-M", "middle": [], "last": "Yeh", "suffix": "" } ], "year": 2009, "venue": "IEEE Transactions on Audio, Speech, and Language Processing", "volume": "17", "issue": "1", "pages": "84--94", "other_ids": {}, "num": null, "urls": [], "raw_text": "S.-H. Lin, B. Chen, and Y.-M. Yeh, \"Exploring the use of speech features and their corresponding distribution characteristics for robust speech recognition,\" IEEE Transactions on Audio, Speech, and Language Processing, vol. 17, no. 1, pp. 84-94, 2009.", "links": null }, "BIBREF33": { "ref_id": "b33", "title": "Nonlinear and linear transformations of speech features to compensate for channel and noise effects", "authors": [ { "first": "S", "middle": [], "last": "Prasad", "suffix": "" }, { "first": "S", "middle": [ "A" ], "last": "Zahorian", "suffix": "" } ], "year": 2005, "venue": "Proc. European Conf. on Speech Communication and Technology", "volume": "", "issue": "", "pages": "969--972", "other_ids": {}, "num": null, "urls": [], "raw_text": "S. Prasad and S. A. Zahorian, \"Nonlinear and linear transformations of speech features to compensate for channel and noise effects,\" in Proc. European Conf. on Speech Communication and Technology, 2005, pp. 969-972.", "links": null }, "BIBREF34": { "ref_id": "b34", "title": "Cepstral domain segmental nonlinear feature transformations for robust speech recognition", "authors": [ { "first": "J", "middle": [ "C" ], "last": "Segura", "suffix": "" }, { "first": "C", "middle": [], "last": "Benitez", "suffix": "" }, { "first": "A", "middle": [], "last": "De La Torre", "suffix": "" }, { "first": "A", "middle": [ "J" ], "last": "Rubio", "suffix": "" }, { "first": "J", "middle": [], "last": "Ramirez", "suffix": "" } ], "year": 2004, "venue": "IEEE Signal Processing Letters", "volume": "11", "issue": "5", "pages": "517--520", "other_ids": {}, "num": null, "urls": [], "raw_text": "J. C. Segura, C. Benitez, A. de la Torre, A. J. Rubio, and J. Ramirez, \"Cepstral domain segmental nonlinear feature transformations for robust speech recognition,\" IEEE Signal Processing Letters, vol. 11, no. 5, pp. 517-520, 2004.", "links": null }, "BIBREF35": { "ref_id": "b35", "title": "Rasta processing of speech", "authors": [ { "first": "H", "middle": [], "last": "Hermansky", "suffix": "" }, { "first": "N", "middle": [], "last": "Morgan", "suffix": "" } ], "year": 1994, "venue": "IEEE Transactions on Speech and Audio Processing", "volume": "2", "issue": "4", "pages": "578--589", "other_ids": {}, "num": null, "urls": [], "raw_text": "H. Hermansky and N. Morgan, \"Rasta processing of speech,\" IEEE Transactions on Speech and Audio Processing, vol. 2, no. 4, pp. 578-589, 1994.", "links": null }, "BIBREF36": { "ref_id": "b36", "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": "Filali", "suffix": "" }, { "first": "J", "middle": [ "A" ], "last": "Bilmes", "suffix": "" } ], "year": 2002, "venue": "Proc. Annu. Conf. of the Int. Speech Communication Association", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "C.-P. Chen, K. Filali, and J. A. Bilmes, \"Frontend post-processing and backend model enhancement on the aurora 2.0/3.0 databases,\" in Proc. Annu. Conf. of the Int. Speech Communication Association, 2002.", "links": null }, "BIBREF37": { "ref_id": "b37", "title": "Communication in the presence of noise", "authors": [ { "first": "C", "middle": [ "E" ], "last": "Shannon", "suffix": "" } ], "year": 1949, "venue": "Proceedings of the IRE", "volume": "37", "issue": "1", "pages": "10--21", "other_ids": {}, "num": null, "urls": [], "raw_text": "C. E. Shannon, \"Communication in the presence of noise,\" Proceedings of the IRE, vol. 37, no. 1, pp. 10-21, 1949.", "links": null }, "BIBREF38": { "ref_id": "b38", "title": "On the importance of various modulation frequencies for speech recognition", "authors": [ { "first": "N", "middle": [], "last": "Kanedera", "suffix": "" }, { "first": "T", "middle": [], "last": "Arai", "suffix": "" }, { "first": "H", "middle": [], "last": "Hermansky", "suffix": "" }, { "first": "M", "middle": [], "last": "Pavel", "suffix": "" } ], "year": 1997, "venue": "Proc. European Conf. on Speech Communication and Technology", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "N. Kanedera, T. Arai, H. Hermansky, and M. Pavel, \"On the importance of various modulation frequencies for speech recognition,\" in Proc. European Conf. on Speech Communication and Technology, 1997.", "links": null }, "BIBREF39": { "ref_id": "b39", "title": "On the origins of speech intelligibility in the real world", "authors": [ { "first": "S", "middle": [], "last": "Greenberg", "suffix": "" } ], "year": 1997, "venue": "Proc. ESCA-NATO Tutorial and Research Workshop on Robust Speech Recognition for Unknown Communication Channels", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "S. Greenberg, \"On the origins of speech intelligibility in the real world,\" in Proc. ESCA-NATO Tutorial and Research Workshop on Robust Speech Recognition for Unknown Communication Channels, 1997.", "links": null }, "BIBREF40": { "ref_id": "b40", "title": "Comparison of parametric representations for monosyllabic word recognition in continuously spoken sentences", "authors": [ { "first": "S", "middle": [], "last": "Davis", "suffix": "" }, { "first": "P", "middle": [], "last": "Mermelstein", "suffix": "" } ], "year": 1980, "venue": "IEEE Transactions on Acoustics, Speech and Signal Processing", "volume": "28", "issue": "4", "pages": "357--366", "other_ids": {}, "num": null, "urls": [], "raw_text": "S. Davis and P. Mermelstein, \"Comparison of parametric representations for monosyllabic word recognition in continuously spoken sentences,\" IEEE Transactions on Acoustics, Speech and Signal Processing, vol. 28, no. 4, pp. 357-366, 1980.", "links": null }, "BIBREF41": { "ref_id": "b41", "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", "middle": [ "G" ], "last": "Hirsch", "suffix": "" }, { "first": "D", "middle": [], "last": "Gmbh", "suffix": "" } ], "year": 2000, "venue": "Proc. ISCA Workshop on ASR", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "D. Pearce, H. G. Hirsch, and D. Gmbh, \"The aurora experimental framework for the performance evaluation of speech recognition systems under noisy conditions,\" in Proc. ISCA Workshop on ASR, 2000.", "links": null }, "BIBREF43": { "ref_id": "b43", "title": "Evaluation of a noise-robust dsr front-end on aurora databases", "authors": [ { "first": "D", "middle": [], "last": "Macho", "suffix": "" }, { "first": "L", "middle": [], "last": "Mauuary", "suffix": "" }, { "first": "B", "middle": [], "last": "Noe", "suffix": "" }, { "first": "Y", "middle": [ "M" ], "last": "Cheng", "suffix": "" }, { "first": "D", "middle": [], "last": "Ealey", "suffix": "" }, { "first": "D", "middle": [], "last": "Jouvet", "suffix": "" }, { "first": "H", "middle": [], "last": "Kelleher", "suffix": "" }, { "first": "D", "middle": [], "last": "Pearce", "suffix": "" }, { "first": "F", "middle": [], "last": "Saadoun", "suffix": "" } ], "year": 2002, "venue": "Proc. Annu. Conf. of the Int. Speech Communication Association", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "D. Macho, L. Mauuary, B. Noe, Y. M. Cheng, D. Ealey, D. Jouvet, H. Kelleher, D. Pearce, and F. Saadoun, \"Evaluation of a noise-robust dsr front-end on aurora databases,\" in Proc. Annu. Conf. of the Int. Speech Communication Association, 2002.", "links": null }, "BIBREF44": { "ref_id": "b44", "title": "Experimental framework for the performance evaluation of speech recognition front-ends on a large vocabulary task", "authors": [ { "first": "H", "middle": [ "G" ], "last": "Hirsch", "suffix": "" } ], "year": 2002, "venue": "ETSI STQ-Aurora DSR Working Group", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "H. G. Hirsch, \"Experimental framework for the performance evaluation of speech recognition front-ends on a large vocabulary task,\" ETSI STQ-Aurora DSR Working Group, Tech. Rep. AU/384/02, 2002.", "links": null }, "BIBREF45": { "ref_id": "b45", "title": "MATBN: A Mandarin Chinese Broadcast News Corpus", "authors": [ { "first": "H.-M", "middle": [], "last": "Wang", "suffix": "" }, { "first": "B", "middle": [], "last": "Chen", "suffix": "" }, { "first": "J.-W", "middle": [], "last": "Kuo", "suffix": "" }, { "first": "S.-S", "middle": [], "last": "Cheng", "suffix": "" } ], "year": 2005, "venue": "International Journal of Computational Linguistics and Chinese Language Processing", "volume": "10", "issue": "2", "pages": "219--236", "other_ids": {}, "num": null, "urls": [], "raw_text": "H.-M. Wang, B. Chen, J.-W. Kuo, and S.-S. Cheng, \"MATBN: A Mandarin Chinese Broadcast News Corpus,\" International Journal of Computational Linguistics and Chinese Language Processing, vol. 10, no. 2, pp. 219-236, 2005.", "links": null } }, "ref_entries": { "FIGREF0": { "text": "\u4e03\u3001\u8a8c\u8b1d \u672c\u8ad6\u6587\u4e4b\u7814\u7a76\u627f\u8499\u6559\u80b2\u90e8-\u570b\u7acb\u81fa\u7063\u5e2b\u7bc4\u5927\u5b78\u9081\u5411\u9802\u5c16\u5927\u5b78\u8a08\u756b(102J1A0800)\u8207\u884c \u653f\u9662\u570b\u5bb6\u79d1\u5b78\u59d4\u54e1\u6703\u7814\u7a76\u8a08\u756b(NSC 101-2221-E-003-024-MY3, NSC 101-2511-S-003-057-MY3, NSC 101-2511-S-003-047-MY3 \u548c NSC 102-2221-E-003-014-) \u4e4b\u7d93\u8cbb\u652f\u6301\uff0c\u8b39\u6b64\u81f4\u8b1d\u3002 \u53c3\u8003\u6587\u737b", "num": null, "uris": null, "type_str": "figure" }, "TABREF0": { "type_str": "table", "html": null, "num": null, "text": "\u4ee5\u8072\u5b78\u6a21\u578b\u70ba\u57fa\u790e\u4e4b\u5f37\u5065\u6027\u6280\u8853(model-based techniques)\uff1a\u85c9\u7531\u4fee\u6539\u5df2\u8a13\u7df4\u4e4b\u8072 \u5b78\u6a21\u578b(acoustic model)\u7684\u6a21\u578b\u53c3\u6578\uff0c\u4f7f\u8072\u5b78\u6a21\u578b\u80fd\u5920\u9069\u61c9\u8207\u8a13\u7df4\u6642\u4e0d\u540c\u7684\u74b0\u5883\uff0c \u8a9e\u6599\u5eab\u4e2d\u4e0d\u540c\u8a0a\u566a\u6bd4\u8a9e\u53e5 MFCC \u7279\u5fb5 c1 \u53c3\u6578\u4e4b\u8abf\u8b8a\u983b\u8b5c\u7684\u5dee\u7570 Proceedings of the Twenty-Fifth Conference on Computational Linguistics and Speech Processing (ROCLING 2013)", "content": "
Proceedings of the Twenty-Fifth Conference on Computational Linguistics and Speech Processing (ROCLING 2013)
\u53e6\u5916\uff0c\u8fd1\u5e74\u4f86\u4ea6\u6709\u4e00\u4e9b\u7814\u7a76\u986f\u793a\uff0c\u74b0\u5883\u4e2d\u7684\u5e72\u64fe\u56e0\u7d20\u4e0d\u53ea\u6703\u6539\u8b8a\u8a9e\u97f3\u7279\u5fb5\u7684\u5206\u4f48\u7279
\u6027\uff0c\u4e5f\u6703\u4f7f\u8a9e\u97f3\u7279\u5fb5\u7684\u6642\u57df\u7d50\u69cb(temporal structure)\u7522\u751f\u626d\u66f2\u3002\u8abf\u8b8a\u983b\u8b5c(modulation
spectrum)[24]\u70ba\u4e00\u6709\u6548\u63cf\u7e6a\u6574\u500b\u8a9e\u53e5\u8a9e\u97f3\u7279\u5fb5\u4e4b\u6642\u57df\u7d50\u69cb\u7684\u5a92\u4ecb\uff0c\u76f8\u8f03\u65bc\u4e00\u822c\u7684\u8a9e\u97f3\u7279
\u5fb5\u80fd\u5448\u73fe\u51fa\u66f4\u5ee3\u6cdb\u7684\u8a9e\u97f3\u8b8a\u5316\u7279\u6027\u3002\u800c\u8abf\u8b8a\u983b\u8b5c\u6b63\u898f\u5316\u7684\u7814\u7a76\uff0c\u4fbf\u8a66\u5716\u5c07\u4e0a\u8ff0\u8a9e\u97f3\u7279\u5fb5
\u5206\u4f48\u7279\u6027\u6b63\u898f\u5316\u7684\u6982\u5ff5\uff0c\u61c9\u7528\u5728\u8a9e\u97f3\u7279\u5fb5\u7684\u8abf\u8b8a\u983b\u8b5c\u4e0a\u3002\u4e0d\u540c\u65bc\u5728\u6642\u57df\u4e0a\u8a9e\u97f3\u7279\u5fb5\u6b63\u898f \u5f9e\u800c\u6e1b\u5c11\u74b0\u5883\u4e0d\u5339\u914d\u9020\u6210\u7684\u554f\u984c\u3002\u4f8b\u5982\u7d93\u5178\u7684\u6700\u5927\u76f8\u4f3c\u5ea6\u7dda\u6027\u56de\u6b78\u6cd5(maximum \u5316\u7684\u6280\u8853\uff0c\u8abf\u8b8a\u983b\u8b5c\u6b63\u898f\u5316\u6280\u8853\u8003\u616e\u4e86\u8a9e\u53e5\u7684\u6574\u9ad4\u8b8a\u5316\u60c5\u5f62\uff0c\u8207\u8a9e\u97f3\u7279\u5fb5\u6b63\u898f\u5316\u6280\u8853\u63a1 likelihood linear regression, MLLR)[3]\u3001\u5e73\u884c\u6a21\u578b\u7d50\u5408\u6cd5(parallel model combination, \u7528\u4e0d\u540c\u7684\u89d2\u5ea6\u5207\u5165\u74b0\u5883\u5e72\u64fe\u7684\u554f\u984c\u3002\u985e\u4f3c\u65bc\u8a9e\u97f3\u7279\u5fb5\u6b63\u898f\u5316\u7684\u7814\u7a76\u9014\u5f91\uff0c\u8abf\u8b8a\u983b\u8b5c\u5e73\u5747 PMC)[4]\u3001\u57fa\u65bc\u5411\u91cf\u6cf0\u52d2\u5c55\u958b\u5f0f(vector Taylor series)\u7684\u6a21\u578b\u8abf\u9069[5]\u7b49\u3002\u6b64\u985e\u65b9\u6cd5 \u503c\u6b63\u898f\u5316\u6cd5 (spectral mean normalization, SMN) \u53ca\u8abf\u8b8a\u983b\u8b5c\u5e73\u5747\u503c\u8b8a\u7570\u6578\u6b63\u898f\u5316\u6cd5 \u901a\u5e38\u80fd\u5c0d\u5f37\u5065\u6027\u6709\u76f8\u7576\u4e0d\u932f\u7684\u6539\u5584\uff0c\u4f46\u6240\u9700\u8981\u7684\u8abf\u9069\u8a9e\u6599\u8f03\u591a\uff0c\u904b\u7b97\u8907\u96dc\u5ea6\u4e5f\u8f03 (spectral mean and variance normalization, SMVN)[25]\u3001\u8abf\u8b8a\u983b\u8b5c\u7d71\u8a08\u5716\u7b49\u5316\u6cd5(spectral \u9ad8[1]\u3002 histogram equalization, SHE)[26]\u7b49\u65b9\u6cd5\u90fd\u5c6c\u65bc\u6b64\u4e00\u7814\u7a76\u9818\u57df\u7684\u6210\u679c\u3002\u53e6\u5916\uff0c\u4e5f\u6709\u4e00\u4e9b\u7814 2. \u8a9e\u97f3\u5f37\u5316(speech enhancement)\uff1a\u5f37\u5316\u6240\u63a5\u6536\u5230\u7684\u8a9e\u97f3\u8a0a\u865f\uff0c\u4f7f\u8a72\u8a9e\u97f3\u8a0a\u865f\u6240\u53d7\u5230 \u7a76\u6839\u64da\u8abf\u8b8a\u983b\u8b5c\u7684\u7279\u6027\u767c\u5c55\u65b0\u7684\u6b63\u898f\u5316\u65b9\u6cd5\uff0c\u4f8b\u5982\u8abf\u8b8a\u983b\u8b5c\u53d6\u4ee3\u6cd5(modulation spectrum \u7684\u74b0\u5883\u56e0\u7d20\u5e72\u64fe\u6e1b\u5c11\u6216\u6d88\u5931\uff0c\u5f9e\u800c\u6a21\u64ec\u5728\u7406\u60f3\u9304\u97f3\u74b0\u5883\u4e0b\u6240\u53d6\u5f97\u7684\u8a9e\u97f3\u8a0a\u865f\uff0c\u85c9 replacement, MSR)[27]\u3001\u57fa\u65bc\u6ffe\u6ce2\u5668\u8a2d\u8a08\u7684\u6642\u57df\u5e8f\u5217\u7d50\u69cb\u6b63\u898f\u5316\u6cd5(temporal structure \u4ee5\u964d\u4f4e\u96dc\u8a0a\u7684\u5f71\u97ff\u3002\u4f8b\u5982\u7d93\u5178\u7684\u983b\u8b5c\u6d88\u53bb\u6cd5(spectral subtraction, SS)[6]\u3001\u8a0a\u865f\u5b50 normalization, TSN)[28]\u3001\u4ee5\u53ca\u6b63\u898f\u5316\u9ad8\u4f4e\u983b\u6bd4\u4f8b\u7684\u5f37\u5ea6\u983b\u8b5c\u6bd4\u4f8b\u6b63\u898f\u5316\u6cd5(magnitude \u7a7a\u9593\u6cd5(signal subspace approach)[7]\u3001\u7dad\u7d0d\u6ffe\u6ce2\u5668(Wiener filtering)[8]\u3001\u6216\u662f\u57fa\u65bc ratio equalization, MRE)[26]\u7b49\u3002\u5176\u4e2d SHE \u6240\u63a1\u7528\u7684\u6982\u5ff5\u8207\u4f5c\u7528\u65bc\u7279\u5fb5\u4e0a\u7684 HEQ \u985e\u4f3c\uff0c \u7d71\u8a08\u4f30\u6e2c\u5b50\u7684\u8a9e\u97f3\u5f37\u5316\u6280\u8853[9]\u7b49\u3002\u9019\u4e00\u985e\u7684\u65b9\u6cd5\u7d93\u5e38\u662f\u91dd\u5c0d\u4eba\u8033\u7684\u7279\u6027\u8a2d\u8a08\uff0c \u4f46 HEQ \u662f\u76f4\u63a5\u8abf\u6574\u7279\u5fb5\u7684\u6578\u503c\uff0cSHE \u8abf\u6574\u7684\u5247\u662f\u7279\u5fb5\u8b8a\u5316\u7684\u8da8\u52e2\u8207\u898f\u5f8b\uff0c\u6b64\u5169\u7a2e\u8abf\u6574 \u4f46\u5176\u5f15\u5165\u7684\u975e\u7dda\u6027\u626d\u66f2\u6709\u6642\u6703\u5c0d\u81ea\u52d5\u8a9e\u97f3\u8fa8\u8b58\u7cfb\u7d71\u6709\u8ca0\u9762\u7684\u5f71\u97ff[10]\u3002 \u6a19\u7684\u662f\u4e0d\u540c\u7684\uff0c\u56e0\u6b64\u5177\u6709\u9ad8\u5ea6\u7684\u4e92\u88dc\u6027[29,30]\u3002 \u5716\u4e00\u3001Aurora-2
3. \u5f37\u5065\u6027\u8a9e\u97f3\u7279\u5fb5\u64f7\u53d6(robust speech feature extraction)\uff1a\u85c9\u7531\u6539\u8b8a\u8a9e\u97f3\u7279\u5fb5\u64f7\u53d6\u7684 \u904e\u7a0b\uff0c\u627e\u51fa\u8f03\u4e0d\u6703\u56e0\u74b0\u5883\u4e0d\u5339\u914d\u800c\u6539\u8b8a\u5176\u7279\u6027\u7684\u8a9e\u97f3\u7279\u5fb5\u53c3\u6578\u3002\u5176\u4e2d\u6709\u4e00\u90e8\u4efd\u7684 \u65b9\u6cd5\u5e0c\u671b\u627e\u5230\u4e00\u7a2e\u901a\u7528\u7684\u7279\u5fb5\u8868\u793a\u6cd5\uff0c\u4f7f\u4e7e\u6de8\u7684\u8a9e\u97f3\u548c\u53d7\u96dc\u8a0a\u5e72\u64fe\u7684\u8a9e\u97f3\u80fd\u8868\u73fe \u51fa\u985e\u4f3c\u7684\u7279\u6027[11-13]\uff1b\u800c\u53e6\u4e00\u4e9b\u65b9\u6cd5\u5247\u662f\u8a66\u8457\u904b\u7528\u5404\u7a2e\u88dc\u511f\u7684\u65b9\u5f0f\uff0c\u5c07\u8a9e\u97f3\u7279\u5fb5 \u6709\u9451\u65bc\u6b64\uff0c\u672c\u8ad6\u6587\u5ef6\u7e8c\u4ee5\u5206\u983b\u5e36\u7684\u65b9\u5f0f\u5f15\u5165\u6587\u8108\u8cc7\u8a0a\u4e4b\u7814\u7a76\uff0c\u63d0\u51fa\u5c07\u5176\u6982\u5ff5\u61c9\u7528\u5728 \u7279\u5fb5\u9593\u56e0\u70ba\u96dc\u8a0a\u800c\u7522\u751f\u7684\u52d5\u614b\u7bc4\u570d(dynamic range)\u5dee\u7570\uff0c\u4f7f\u5f97\u96dc\u8a0a\u5c0d\u8a9e\u97f3\u7279\u5fb5\u7684\u5f71\u97ff\u66f4 \u8abf\u8b8a\u983b\u8b5c\u7d71\u8a08\u5716\u7b49\u5316\u6cd5\u4e2d\u7684 \u300c\u57fa\u65bc\u7a7a\u9593\u57df-\u6642\u57df\u6587\u8108\u7d71\u8a08\u8cc7\u8a0a\u7684\u8abf\u8b8a\u983b\u8b5c\u7d71\u8a08\u5716\u7b49\u5316\u6cd5\u300d \u70ba\u7e2e\u5c0f\u3002\u5728\u9019\u4e9b\u57fa\u790e\u4e4b\u4e0b\uff0c\u4e5f\u6709\u5b78\u8005\u63d0\u51fa\u6b63\u898f\u5316\u8a9e\u97f3\u7279\u5fb5\u7684\u7b2c\u4e09\u968e\u52d5\u5dee\u6216\u4efb\u610f\u968e\u6578\u7684\u52d5 (ST-PSHE)\u3002\u5229\u7528\u7c21\u55ae\u7684\u9ad8\u901a(high-pass)\u53ca\u4f4e\u901a(low-pass)\u6ffe\u6ce2\u5668\u53d6\u5f97\u9ad8\u983b\u53ca\u4f4e\u983b\u7684\u6587\u8108 \u5dee\u7684\u6280\u8853[16]\u3002 \u8cc7\u8a0a\uff0c\u91dd\u5c0d\u9019\u4e9b\u6587\u8108\u8cc7\u8a0a\u9032\u884c\u8abf\u8b8a\u983b\u8b5c\u7d71\u8a08\u5716\u7b49\u5316\u6cd5\uff0c\u518d\u5c07\u6b63\u898f\u5316\u5f8c\u7684\u9ad8\u4f4e\u983b\u6210\u4efd\u7d50\u5408 \u7576\u4e2d\u53d7\u5230\u7684\u5e72\u64fe\u9084\u539f\u6210\u672a\u53d7\u5e72\u64fe\u524d\u7684\u6a23\u5b50[14,15]\u3002\u672c\u8ad6\u6587\u7684\u4e3b\u8981\u7684\u8a0e\u8ad6\u90fd\u96c6\u4e2d\u5728 \u6210\u70ba\u65b0\u7684\u8a9e\u97f3\u7279\u5fb5\uff0c\u85c9\u6b64\u6539\u5584\u50b3\u7d71\u7d71\u8a08\u5716\u7b49\u5316\u6cd5\u4e2d\u7684\u9650\u5236\uff0c\u53c8 \u540c\u6642\u80fd\u8abf\u6574\u8a9e\u53e5\u7684\u6642\u57df\u7d50 (\u4e8c)\u7d71\u8a08\u5716\u7b49\u5316\u6cd5 \u5f37\u5065\u6027\u8a9e\u97f3\u7279\u5fb5\u64f7\u53d6\u4e2d\u3002 \u5728\u5f37\u5065\u6027\u8a9e\u97f3\u7279\u5fb5\u64f7\u53d6\u7684\u7814\u7a76\u4e2d\uff0c\u5176\u4e2d\u4e00\u500b\u91cd\u8981\u7684\u7814\u7a76\u9818\u57df\u7a31\u70ba\u8a9e\u97f3\u7279\u5fb5\u6b63\u898f\u5316 (feature normalization)\u3002\u9019\u500b\u9818\u57df\u7684\u7814\u7a76\u4e3b\u5f35\u5c07\u8a9e\u97f3\u7279\u5fb5\u5e8f\u5217\u4e2d\u7684\u67d0\u4e9b\u7279\u6027\u8b8a\u70ba\u4e00\u81f4\uff0c \u4f7f\u9019\u7a2e\u65b0\u7684\u8a9e\u97f3\u7279\u5fb5\u8868\u793a\u6cd5\u80fd\u8f03\u4e0d\u53d7\u96dc\u8a0a\u7684\u5f71\u97ff\u3002\u5176\u4e2d\uff0c\u672c\u8ad6\u6587\u8a0e\u8ad6\u7684\u4e3b\u8981\u70ba\u57fa\u65bc\u7d71\u8a08 \u69cb\u8cc7\u8a0a\uff0c\u4e5f\u5c31\u662f\u7279\u5fb5\u8b8a\u5316\u7684\u898f\u5f8b\u3002\u5728\u7b2c\u4e8c\u7ae0\u53ca\u7b2c\u4e09\u7ae0\u4e2d\uff0c\u6211\u5011\u5c07\u5148\u7c21\u8981\u4ecb\u7d39\u8a9e\u97f3\u7279\u5fb5\u6b63 \u7d71\u8a08\u5716\u7b49\u5316\u6cd5\u70ba\u5f71\u50cf\u8655\u7406\u9818\u57df\u5e38\u7528\u7684\u6f14\u7b97\u6cd5\uff0c\u7528\u4ee5\u8abf\u6574\u5982\u660e\u5ea6\u3001\u8272\u5f69\u5e73\u8861\u7b49\u5f71\u50cf\u53c3\u6578 \u898f\u5316\u7684\u65b9\u6cd5\u53ca\u57fa\u65bc\u8abf\u8b8a\u983b\u8b5c\u7684\u6b63\u898f\u5316\u65b9\u6cd5\uff1b\u7b2c\u56db\u7ae0\u5247\u8a73\u7d30\u8aaa\u660e\u672c\u8ad6\u6587\u6240\u63d0\u51fa\u4e4b\u6539\u826f\u5f0f\u67b6 [31]\uff1b\u800c\u5728\u81ea\u52d5\u8a9e\u97f3\u8fa8\u8b58\u7684\u9818\u57df\uff0c\u4e5f\u6709\u5b78\u8005\u63d0\u51fa\u5229\u7528\u7d71\u8a08\u5716\u7b49\u5316\u6cd5\u4f86\u88dc\u511f\u96dc\u8a0a\u5728\u8a9e\u97f3\u7279 \u69cb\uff1b\u63a5\u8457\uff0c\u5be6\u9a57\u7684\u8a2d\u5b9a\u3001\u7d50\u679c\u8207\u5206\u6790\u5c07\u5728\u7b2c\u4e94\u7ae0\u4e2d\u5448\u73fe\uff0c\u800c\u7b2c\u516d\u7ae0\u5247\u70ba\u7d50\u8ad6\u8207\u672a\u4f86\u53ef\u80fd \u5fb5\u4e0a\u9020\u6210\u7684\u5931\u771f\uff0c\u8a31\u591a\u7814\u7a76\u4e5f\u8b49\u660e\u4e86\u5b83\u7684\u6709\u6548\u6027[18,32-35]\u3002\u524d\u4e00\u7bc0\u6240\u4ecb\u7d39\u7684 CMS \u8207 \u7684\u7814\u7a76\u65b9\u5411\u3002 CMVN\uff0c\u4e43\u81f3\u65bc\u66f4\u9ad8\u968e\u52d5\u5dee\u7684\u6b63\u898f\u5316\u65b9\u6cd5\uff0c\u5747\u662f\u4ee5\u7dda\u6027(linear)\u7684\u65b9\u5f0f\u88dc\u511f\u96dc\u8a0a\u5c0d\u8a9e\u97f3\u7279 \u5206\u4f48\u7684\u8a9e\u97f3\u7279\u5fb5\u6b63\u898f\u5316(distribution-based feature normalization)\uff0c\u4ea6\u5373\u5c07\u540c\u4e00\u7dad\u5ea6\u7684\u8a9e\u97f3 \u7279\u5fb5\u5e8f\u5217\u8996\u70ba\u96a8\u6a5f\u8b8a\u6578(random variable)\u7684\u4e00\u7d44\u6a23\u672c(sample)\uff0c\u5229\u7528\u9019\u4e9b\u6a23\u672c\u4f30\u8a08\u8a72\u96a8\u6a5f \u5fb5\u7684\u5e72\u64fe\uff0c\u4f46\u5c0d\u65bc\u975e\u7dda\u6027\u7684\u626d\u66f2\u88dc\u511f\u6548\u679c\u6709\u9650\uff0c\u7d71\u8a08\u5716\u7b49\u5316\u6cd5\u5247\u5f4c\u88dc\u4e86\u52d5\u5dee\u6b63\u898f\u5316\u6cd5\u7684 \u4e8c\u3001\u8a9e\u97f3\u7279\u5fb5\u6b63\u898f\u5316\u6280\u8853 \u6b64\u4e00\u7f3a\u5931\u3002\u76f8\u8f03\u65bc\u52d5\u5dee\u6b63\u898f\u5316\u6cd5\uff0c\u7d71\u8a08\u5716\u7b49\u5316\u6cd5\u4e0d\u5c0d\u52d5\u5dee\u9032\u884c\u6b63\u898f\u5316\uff0c\u800c\u662f\u5229\u7528\u4e00\u975e\u7dda \u8b8a\u6578\u7684\u7d71\u8a08\u91cf\uff0c\u64da\u6b64\u5c0d\u7279\u5fb5\u5e8f\u5217\u7684\u5206\u4f48\u9032\u884c\u7dda\u6027\u6216\u975e\u7dda\u6027\u7684\u8f49\u63db\u3002\u4f8b\u5982\u57fa\u65bc\u52d5\u5dee\u6b63\u898f\u5316 (moment normalization)\u7684\u5012\u983b\u8b5c\u5e73\u5747\u503c\u6e1b\u53bb\u6cd5(cepstral mean subtraction, CMS)[12]\u3001\u5012\u983b \u6027(non-linear)\u7684\u8f49\u63db\uff0c\u5c07\u6240\u6709\u8a9e\u97f3\u7279\u5fb5\u7684\u7d71\u8a08\u5206\u4f48\u76f4\u63a5\u8b8a\u5f97\u8207\u672a\u53d7\u96dc\u8a0a\u5e72\u64fe\u6642\u7684\u7d71\u8a08\u5206 (\u4e00)\u52d5\u5dee\u6b63\u898f\u5316\u6cd5 \u4f48\u4e00\u81f4\uff0c\u4e26\u4e14\u7121\u9700\u5c0d\u8a72\u7d71\u8a08\u5206\u4f48\u64c1\u6709\u5148\u9a57\u77e5\u8b58(prior knowledge)\uff0c\u5373\u53ef\u6709\u6548\u5730\u6539\u5584\u96dc\u8a0a \u8b5c\u5e73\u5747\u503c\u8b8a\u7570\u6578\u6b63\u898f\u5316\u6cd5(cepstral mean and variance normalization, CMVN)[13]\u3001\u9ad8\u968e\u5012 \u52d5\u5dee\u6b63\u898f\u5316(moment normalization)\u7684\u6280\u8853\uff0c\u4e3b\u8981\u900f\u904e\u6b63\u898f\u5316\u6bcf\u4e00\u500b\u8a9e\u53e5(utterance)\u4e2d\u5404\u7dad \u8a9e\u97f3\u7684\u8fa8\u8b58\u6b63\u78ba\u7387\u3002 \u983b\u8b5c\u52d5\u5dee\u6b63\u898f\u5316\u6cd5(higher order cepstral moment normalization, HOCMN)[16]\uff0c\u4ee5\u53ca\u53ef\u4ee5 \u6d88\u9664\u66f4\u591a\u975e\u7dda\u6027\u74b0\u5883\u56e0\u7d20\u5f71\u97ff\u7684\u7d71\u8a08\u5716\u7b49\u5316\u6cd5(histogram equalization, HEQ)[11]\u7b49\u90fd\u662f \u6b64\u4e00\u7814\u7a76\u65b9\u5411\u7684\u6210\u54e1\u3002\u6b64\u985e\u7684\u6280\u8853\u5927\u591a\u5177\u6709\u76f4\u89c0\u3001\u5feb\u901f\u4e14\u6709\u6548\u7684\u7279\u6027\uff0c\u662f\u5f37\u5065\u6027\u8a9e\u97f3\u7279 \u5fb5\u64f7\u53d6\u7684\u9818\u57df\u4e0d\u53ef\u7f3a\u5c11\u7684\u4e00\u74b0\u3002 \u8a31\u591a\u904e\u53bb\u7814\u7a76[17-19]\u90fd\u8aaa\u660e\u4e86\u7d71\u8a08\u5716\u7b49\u5316\u6cd5\u80fd\u5920\u6709\u6548\u5730\u88dc\u511f\u975e\u7dda\u6027\u7684\u96dc\u8a0a\u5e72\u64fe\uff0c\u800c \u5c0d\u8fa8\u8b58\u7684\u6b63\u78ba\u7387\u6709\u986f\u8457\u7684\u63d0\u5347\uff0c\u4f46\u7d71\u8a08\u5716\u7b49\u5316\u6cd5\u4ecd\u7136\u6709\u4e00\u4e9b\u4e0d\u76e1\u6b63\u78ba\u7684\u5047\u8a2d\u3002\u4f8b\u5982\u5176\u5047 \u8a2d\u8a9e\u97f3\u7279\u5fb5\u4e2d\u5404\u7dad\u5ea6\u9593\u5f7c\u6b64\u7368\u7acb\uff0c\u56e0\u800c\u53ef\u4ee5\u5c0d\u500b\u5225\u7dad\u5ea6\u5206\u5225\u9032\u884c\u6b63\u898f\u5316\uff0c\u4f46\u5e38\u898b\u7684\u904b\u7528 \u5ea6\u7279\u5fb5\u7d71\u8a08\u5206\u4f48\u7684\u52d5\u5dee\uff0c\u4f86\u6e1b\u5c11\u96dc\u8a0a\u5c0d\u8a9e\u97f3\u7279\u5fb5\u7684\u5f71\u97ff\u3002\u4f8b\u5982\u5012\u983b\u8b5c\u5e73\u5747\u503c\u6e1b\u53bb\u6cd5 \u7d71\u8a08\u5716\u7b49\u5316\u6cd5\u4e3b\u8981\u7684\u505a\u6cd5\uff0c\u662f\u5c07\u76ee\u524d\u8a9e\u53e5\u4e2d\u7279\u5fb5\u5206\u4f48\u7684\u7d2f\u7a4d\u5bc6\u5ea6\u51fd\u6578(cumulative [12](\u4e0b\u7a31 CMS)\u5e0c\u671b\u85c9\u7531\u5c07\u6bcf\u4e00\u500b\u8a9e\u53e5\u7684\u7b2c\u4e00\u968e\u52d5\u5dee(first-order moment)\uff0c\u4e5f\u5c31\u662f\u671f\u671b\u503c distribution function, CDF)\uff0c\u5c0d\u61c9\u81f3\u7531\u8a13\u7df4\u8a9e\u6599\u6240\u7d71\u8a08\u51fa\u4f86\u7684\u53c3\u8003\u5206\u5e03\uff0c\u85c9\u6b64\u5c07\u6574\u53e5\u8a71\u7684 \u6e1b\u53bb\uff0c\u4f86\u6e1b\u5c11\u96dc\u8a0a\u7684\u5f71\u97ff\uff1b\u800c\u5012\u983b\u8b5c\u5e73\u5747\u503c\u8b8a\u7570\u6578\u6b63\u898f\u5316\u6cd5[13](\u4e0b\u7a31 CMVN)\u5247\u66f4\u9032\u4e00 \u7279\u5fb5\u9084\u539f\u81f3\u8207\u8a13\u7df4\u8a9e\u6599\u76f8\u540c\u7684\u7d71\u8a08\u5206\u4f48\u3002\u4ee4 (\u2022)\u70ba\u76ee\u524d\u8a9e\u53e5\u8a9e\u97f3\u7279\u5fb5\u6642\u9593\u5e8f\u5217* , -+\u7684 \u6b65\u5c07\u6b63\u898f\u5316\u7684\u7bc4\u570d\u64f4\u5c55\u81f3\u7b2c\u4e8c\u968e\u52d5\u5dee\uff0c\u4f7f\u4e0d\u540c\u8a9e\u53e5\u9593\u7684\u8b8a\u7570\u6578(variance)\u4e5f\u8b8a\u5f97\u4e00\u81f4\u3002\u4ee4 \u6a5f\u7387\u5206\u4f48(\u4ee5\u4e00\u500b\u5c07\u503c\u5c0d\u61c9\u5230 CDF \u7684\u51fd\u6578\u8868\u793a)\uff0c\u800c (\u2022)\u70ba\u6839\u64da\u6240\u6709\u8a13\u7df4\u8a9e\u6599\u7d71\u8a08\u51fa\u7684\u53c3 \u4e00\u8a9e\u53e5\u4e2d\uff0c\u67d0\u4e00\u7dad\u5ea6\u7684\u8a9e\u97f3\u7279\u5fb5\u6642\u9593\u5e8f\u5217\u70ba* , -+\uff0c\u03bc\u70ba* , -+\u7684\u671f\u671b\u503c\uff0c\u03c3 2 \u70ba\u5176\u8b8a\u7570\u6578\uff0c \u8003\u5206\u4f48\uff0c\u7d71\u8a08\u5716\u7b49\u5316\u6cd5\u6b63\u898f\u5316\u5f8c\u7684\u8a9e\u97f3\u7279\u5fb5\u53ef\u4ee5\u8868\u793a\u70ba\uff1a \u5247\u7d93\u6b64\u5169\u500b\u65b9\u6cd5\u6b63\u898f\u5316\u904e\u7684\u7279\u5fb5\u5206\u5225\u53ef\u4ee5\u8868\u793a\u70ba\uff1a \u0302C MS , -= , -\u2212 \u0302H EQ , -= \u22121 ( ( , -)) (3) (1) \uf9dd\u7528\u96e2\u6563\u9918\u5f26\u8f49\u63db(discrete cosine transform, DCT)\u6c42\u53d6\u7684\u8a9e\u97f3\u7279\u5fb5\uff0c\u5404\u7dad\u5ea6\u4e4b\u9593\u4ecd\u5177\u6709 \u90e8\u4efd\u7684\u76f8\u95dc\u6027\uff1b\u800c\u8a9e\u97f3\u662f\u96a8\u6642\u9593\u7de9\u6162\u8b8a\u5316\u7684\u8a0a\u865f\uff0c\u5728\u7d71\u8a08\u5716\u7b49\u5316\u6cd5\u4e2d\u5c07\u6bcf\u4e00\u500b\u97f3\u6846 \u0302C MVN , -= \u50b3\u7d71\u7684\u7d71\u8a08\u5716\u7b49\u5316\u6cd5\u901a\u5e38\u4ee5\u67e5\u8868\u6cd5(table lookup)\u63cf\u8ff0 (\u2022)\u51fd\u6578\u7684\u5c0d\u61c9\u95dc\u4fc2\uff0c\u4f46\u9019\u6a23 , -\u2212 (2) \u7684\u65b9\u6cd5\u4e0d\u50c5\u8f03\u8cbb\u6642\uff0c\u4e5f\u9700\u8981\u82b1\u8cbb\u8a31\u591a\u7a7a\u9593\u4f86\u8a18\u9304\u8868\u683c\u3002\u5728[33]\u4e2d\uff0c\u6211\u5011\u63d0\u51fa\u5229\u7528\u4e00\u591a\u9805 (frame)\u500b\u5225\u770b\u5f85\u7684\u65b9\u5f0f\u4e5f\u7121\u6cd5\u6709\u6548\u6293\u4f4f\u6642\u57df\u4e0a\u8207\u524d\u5f8c\u5176\u4ed6\u97f3\u6846\u7684\u76f8\u95dc\u6027\u3002\u91dd\u5c0d\u9019\u7a2e\u6bd4\u8f03 \u56b4\u683c\u7684\u5047\u8a2d\uff0c\u6709\u8a31\u591a\u4e0d\u540c\u7684\u65b9\u6cd5\u88ab\u63d0\u51fa \uff0c\u5982\u904b\u7528\u8ff4\u6b78 (regression)\u6280 \u8853\u6216\u6642\u57df\u5e73\u5747 (temporal average, TA)\u6280\u8853\u5f15\u5165\u524d\u5f8c\u6587\u8cc7\u8a0a[20,21]\uff0c\u6291\u6216\u662f\u5c07\u7a7a\u9593(spatial)\u57df\u53ca\u6642\u57df\u7684\u9ad8 \u4f4e\u983b\u6210\u4efd\u9032\u884c\u6b63\u898f\u5316\uff0c\u4ee5\u5206\u983b\u5e36\u7684\u65b9\u5f0f\u5f15\u5165\u6587\u8108\u8cc7\u8a0a(context information)[22,23]\u3002 \u5f0f\u51fd\u6578\u4f86\u903c\u8fd1 \u22121 (\u2022)\uff0c\u53ef\u4ee5\u964d\u4f4e\u8a08\u7b97\u6642\u9593\u8207\u5132\u5b58\u7a7a\u9593\uff0c\u540c\u6642\u7372\u5f97\u6bd4\u539f\u59cb\u7684 HEQ \u76f8\u4f3c\u6216 \u7531\u65bc\u901a\u9053\u6548\u61c9\u5728\u5012\u983b\u8b5c(cepstrum)\u4e0a\u8207\u539f\u672c\u7684\u8a9e\u97f3\u8a0a\u865f\u70ba\u76f8\u52a0\u7684\u95dc\u4fc2\uff0cCMS \u7684\u6b63\u898f \u8f03\u4f73\u7684\u8fa8\u8b58\u6b63\u78ba\u7387\u3002\u6b64\u65b9\u6cd5\u7a31\u70ba\u591a\u9805\u5f0f\u64ec\u5408\u7d71\u8a08\u5716\u7b49\u5316\u6cd5(polynomial-fit histogram \u5316\u53ef\u4ee5\u6709\u6548\u5730\u6d88\u53bb\u4e00\u4e9b\u7a69\u5b9a(stationary)\u7684\u901a\u9053\u6548\u61c9\uff0c\u800c\u4f7f\u5f97\u8a9e\u97f3\u8fa8\u8b58\u7684\u6b63\u78ba\u7387\u6709\u76f8\u7576\u660e \u986f\u7684\u6539\u5584\u3002\u53e6\u4e00\u65b9\u9762\uff0cCMVN \u5c0d\u8b8a\u7570\u6578\u7684\u6b63\u898f\u5316\uff0c\u66f4\u9032\u4e00\u6b65\u5730\u88dc\u511f\u4e86\u4e0d\u540c\u8a9e\u53e5\u7684\u8a9e\u97f3 equalization, PHEQ)\uff0c\u672c\u8ad6\u6587\u4e2d\u4e4b\u7d71\u8a08\u5716\u7b49\u5316\u6cd5\u7686\u4ee5\u6b64\u65b9\u5f0f\u5be6\u4f5c\uff0c\u5982\u4e0b\u5f0f\u6240\u793a\uff1a
" }, "TABREF1": { "type_str": "table", "html": null, "num": null, "text": "Proceedings of the Twenty-Fifth Conference on Computational Linguistics and Speech Processing(ROCLING 2013)", "content": "
\u57fa\u65bc\u6ffe\u6ce2\u5668\u7684\u6b63\u898f\u5316\u6280\u8853
\u9664\u4e86\u5728\u7d71\u8a08\u5206\u4f48\u4e0a\u9032\u884c\u8655\u7406\u5916\uff0c\u4e5f\u6709\u4e00\u4e9b\u8a9e\u97f3\u7279\u5fb5\u6b63\u898f\u5316\u7684\u65b9\u6cd5\u8a66\u5716\u5f9e\u6ffe\u6ce2\u5668\u7684\u8a2d\u8a08\u51fa
\u767c\u3002\u4f8b\u5982\u76f8\u5c0d\u983b\u8b5c\u6cd5(relative spectra, RASTA)[36]\u4fbf\u662f\u5229\u7528\u4eba\u985e\u8a9e\u97f3\u4e3b\u8981\u8cc7\u8a0a\u96c6\u4e2d\u5728\u7279
\u5b9a\u8abf\u8b8a\u983b\u8b5c\u983b\u5e36\u7684\u539f\u7406\uff0c\u8a2d\u8a08\u4e00\u5e36\u901a\u6ffe\u6ce2\u5668(band-pass filter)\uff0c\u85c9\u4ee5\u79fb\u9664\u8a9e\u97f3\u7279\u5fb5\u4e2d\u8207\u8a9e
\u97f3\u8f03\u4e0d\u76f8\u95dc\u7684\u6210\u4efd\uff1b\u800c\u5728[37]\u4e2d\uff0c\u5247\u662f\u4f7f\u7528\u4f4e\u901a\u6ffe\u6ce2\u5668(low-pass filter)\u5c0d\u7279\u5fb5\u9032\u884c\u5e73\u6ed1\u5316
(smoothing)\uff0c\u4ee5\u964d\u4f4e\u8a9e\u97f3\u7279\u5fb5\u4e2d\u4e0d\u7a69\u5b9a\u6216\u7a81\u767c\u7684\u96dc\u8a0a\u5c0d\u8a9e\u97f3\u7279\u5fb5\u9020\u6210\u7684\u5e72\u64fe\u3002\u503c\u5f97\u4e00\u63d0
\u7684\u662f\uff0c\u5f0f (1)\u4e5f\u53ef\u4ee5\u8996\u70ba\u662f\u4e00\u500b\u9ad8\u901a\u6ffe\u6ce2\u5668(high-pass filter)\u7684\u8108\u885d\u97ff\u61c9(impulse response)\uff0c
\u56e0\u6b64\u5f9e\u53e6\u4e00\u500b\u89d2\u5ea6\u4f86\u89e3\u8b80\uff0cCMS \u4ea6\u662f\u5229\u7528\u6ffe\u6ce2\u7684\u6982\u5ff5\u4f86\u79fb\u9664\u7a69\u5b9a\u901a\u9053\u6548\u61c9\u7684\u4e00\u7a2e\u6280
\u8853\u3002
\u4e09\u3001\u8abf\u8b8a\u983b\u8b5c\u65bc\u5f37\u5065\u6027\u8a9e\u97f3\u8fa8\u8b58\u4e4b\u7814\u7a76 (\u4e8c)\u8abf\u8b8a\u983b\u8b5c\u4e4b\u6b63\u898f\u5316
\u8abf\u8b8a\u983b\u8b5c\u6b63\u898f\u5316\u7684\u76f8\u95dc\u6280\u8853\uff0c\u65e8\u5728\u4f7f\u53d7\u5230\u74b0\u5883\u5e72\u64fe\u800c\u626d\u66f2\u7684\u8abf\u8b8a\u983b\u8b5c\u6062\u5fa9\u70ba\u672a\u53d7\u5e72\u64fe\u7684
(\u4e00)\u8abf\u8b8a\u983b\u8b5c\u4e4b\u5b9a\u7fa9\u8207\u7279\u6027 \u6a23\u8c8c\u3002\u91dd\u5c0d\u5f37\u5065\u6027\u8a9e\u97f3\u8fa8\u8b58\u6b63\u898f\u5316\u8abf\u8b8a\u983b\u8b5c\u7684\u904e\u7a0b\u5927\u81f4\u4e0a\u53ef\u4ee5\u5982\u4e0b\u4e09\u500b\u6b65\u9a5f\u8aaa\u660e\uff1a
\u4ee4\u4e00\u8a9e\u53e5\u4e2d\uff0c\u67d0\u4e00\u7279\u5b9a\u7dad\u5ea6\u4e4b\u8a9e\u97f3\u7279\u5fb5\u6642\u9593\u5e8f\u5217\u70ba* , -+\uff0c\u5176\u4e2d n \u70ba\u97f3\u6846(frame)\u7684\u7d22\u5f15 1) \u5206\u6790\uff1a\u5c07\u53d7\u5230\u74b0\u5883\u5e72\u64fe\u7684\u6574\u53e5\u8a9e\u53e5\u4e4b\u8a9e\u97f3\u7279\u5fb5\u6642\u9593\u5e8f\u5217* , -+\u9032\u884c\u96e2\u6563\u5085\u7acb\u8449\u8f49
\u503c\uff0c\u8a72\u8a9e\u97f3\u7279\u5fb5\u5e8f\u5217\u7684\u8abf\u8b8a\u983b\u8b5c\u53ef\u4ee5\u5b9a\u7fa9\u70ba\uff1a \u63db\uff0c\u5f97\u5230\u8a72\u8a9e\u53e5\u7684\u8abf\u8b8a\u983b\u8b5c* , -+\u3002\u4ee5\u96e2\u6563\u5085\u7acb\u8449\u8f49\u63db\u53d6\u5f97\u4e4b\u5e8f\u5217\u70ba\u4e00\u8907\u6578\u5e8f\u5217\uff0c
\u22121 \u53ef\u518d\u5206\u89e3\u6210\u8a72\u8abf\u8b8a\u983b\u8b5c\u7684\u5f37\u5ea6\u983b\u8b5c*| , -|+\u53ca\u76f8\u4f4d\u983b\u8b5c*\u2220 , -+\u3002 \u2212 2 , -= \u2211 , -2) \u6b63\u898f\u5316\uff1a\u91dd\u5c0d\u524d\u4e00\u6b65\u9a5f\u6240\u5f97\u5230\u7684\u5f37\u5ea6\u983b\u8b5c\u53ca\u76f8\u4f4d\u983b\u8b5c\u9032\u884c\u8655\u7406\u3002\u5176\u4e2d\u76f8\u4f4d\u983b\u8b5c\u901a (5) =0 \u5176\u4e2d = \u221a\u22121\u70ba\u865b\u6578\u55ae\u4f4d\uff0c\u5176\u4e2d \u70ba\u8abf\u8b8a\u983b\u7387\u7684\u7d22\u5f15\uff0cN \u70ba\u8a9e\u53e5\u4e2d\u97f3\u6846\u7684\u7e3d\u6578\uff0c\u6240\u5f97\u4e4b\u5e8f \u5e38\u7dad\u6301\u539f\u72c0\uff0c\u50c5\u6539\u8b8a\u5f37\u5ea6\u983b\u8b5c\u4e2d\u7684\u5f37\u5ea6\uff0c\u4e26\u5f97\u5230\u65b0\u7684\u5f37\u5ea6\u983b\u8b5c*| , -|+\u3002
\u5217* , -+\u5373\u70ba* , -+\u7684\u8abf\u8b8a\u983b\u8b5c\u3002\u5f0f(5)\u53ef\u4ee5\u8996\u70ba\u4e00\u96e2\u6563\u5085\u7acb\u8449\u8f49\u63db(discrete Fourier 3) \u9084\u539f\uff1a\u4f9d\u64da\u539f\u672c\u7684\u76f8\u4f4d\u983b\u8b5c*\u2220 , -+\u548c\u7b2c\u4e8c\u6b65\u9a5f\u4e2d\u6240\u5f97\u4e4b\u65b0\u7684\u5f37\u5ea6\u983b\u8b5c*| , -|+\uff0c
transform, DFT)\uff0c\u8abf\u8b8a\u983b\u8b5c\u4e2d\u7684\u983b\u7387\u7bc4\u570d\u8207\u8a9e\u97f3\u7279\u5fb5\u6642\u9593\u5e8f\u5217\u4e4b\u53d6\u6a23\u7387\u6709\u95dc\uff1a\u5728\u672c\u8ad6\u6587 \u9032\u884c\u53cd\u96e2\u6563\u5085\u7acb\u8449\u8f49\u63db(inverse discrete Fourier transform, IDFT)\uff0c\u53d6\u5f97\u9084\u539f\u5f8c\u7684\u8a9e
\u7684\u57fa\u790e\u8a9e\u97f3\u7279\u5fb5\u8a2d\u5b9a\u4e2d\uff0c\u6bcf\u5169\u500b\u76f8\u9130\u97f3\u6846\u4e4b\u9593\u9694\u70ba 10ms\uff0c\u4ea6\u5373\u8a9e\u97f3\u7279\u5fb5\u6642\u9593\u5e8f\u5217\u4e4b\u53d6 \u97f3\u7279\u5fb5\u6642\u9593\u5e8f\u5217\u3002
\u6a23\u7387\u70ba 100Hz\uff0c\u6839\u64da\u5948\u594e\u65af\u7279\u5b9a\u7406(Nyquist-Shannon sampling theorem)[38]\uff0c\u8abf\u8b8a\u983b\u8b5c\u4e4b
\u6700\u9ad8\u983b\u7387\u70ba 50Hz\u3002 \u82e5\u4e0a\u8ff0\u7b2c\u4e8c\u6b65\u9a5f\u4e2d\u7684\u5f37\u5ea6\u983b\u8b5c\u80fd\u5920\u88ab\u9069\u7576\u5730\u6b63\u898f\u5316\uff0c\u5247\u53ef\u4ee5\u6709\u6548\u964d\u4f4e\u74b0\u5883\u5e72\u64fe\u5c0d\u8abf
\u8b8a\u983b\u8b5c\u7684\u5931\u771f\uff0c\u9032\u800c\u4f7f\u9084\u539f\u5f8c\u7684\u8a9e\u97f3\u7279\u5fb5\u53c3\u6578\uff0c\u5728\u81ea\u52d5\u8a9e\u97f3\u8fa8\u8b58\u7cfb\u7d71\u4e2d\u5f97\u5230\u8f03\u597d\u7684\u8fa8\u8b58 \u8abf\u8b8a\u983b\u8b5c\u5728\u5206\u6790\u8a9e\u97f3\u7279\u5fb5\u4e4b\u6642\u57df\u7d50\u69cb\u4e0a\uff0c\u662f\u5f88\u6709\u7528\u7684\u5de5\u5177\uff1b\u904e\u53bb\u6709\u7814\u7a76[39]\u6307\u51fa\uff0c \u8abf\u8b8a\u983b\u7387\u5927\u7d04 1Hz \u5230 16Hz \u9593\u7684\u4f4e\u983b\u6210\u4efd\uff0c\u8207\u8a9e\u97f3\u8fa8\u8b58\u7684\u6b63\u78ba\u7387\u6709\u660e\u986f\u7684\u95dc\u806f\uff0c\u800c\u5176\u4e2d \u7cbe\u78ba\u7387\u3002\u4ee5\u4e0b\u5c07\u7c21\u8ff0\u6578\u7a2e\u8abf\u8b8a\u983b\u8b5c\u6b63\u898f\u5316\u7684\u65b9\u6cd5\uff1a
\u4ee5 4Hz \u9644\u8fd1\u6240\u5305\u542b\u7684\u8cc7\u8a0a\u6700\u70ba\u91cd\u8981\u3002\u95dc\u65bc\u4eba\u985e\u807d\u89ba\u7684\u7814\u7a76[40]\u4e5f\u4e0d\u7d04\u800c\u540c\u5730\u767c\u73fe\uff1a4Hz \u7684\u8abf\u8b8a\u983b\u7387\u5728\u4eba\u985e\u7684\u807d\u89ba\u611f\u77e5\u4e2d\u4f54\u6709\u5f88\u91cd\u8981\u7684\u5730\u4f4d\u3002 1. \u5f37\u5ea6\u983b\u8b5c\u6bd4\u4f8b\u6b63\u898f\u5316\u6cd5(magnitude ratio equalization, MRE)
\u7576\u8a9e\u97f3\u8a0a\u865f\u53d7\u5230\u96dc\u8a0a\u5e72\u64fe\u6642\uff0c\u4e0d\u53ea\u5176\u8a9e\u97f3\u7279\u5fb5\u6642\u9593\u5e8f\u5217\u7684\u5206\u4f48\u7279\u6027\u6703\u6539\u8b8a\uff0c\u5176\u6642\u57df \u6b64\u6280\u8853[26]\u8a08\u7b97\u8abf\u8b8a\u983b\u8b5c\u4e2d\u4f4e\u983b\u6210\u4efd\u5f37\u5ea6\u548c\u9ad8\u983b\u6210\u4efd\u5f37\u5ea6\u7684\u6bd4\u4f8b\uff0c\u5728\u8a9e\u53e5\u53d7\u5230\u74b0\u5883\u5e72\u64fe
\u7d50\u69cb\u4e5f\u6703\u6709\u4e00\u5b9a\u7a0b\u5ea6\u7684\u626d\u66f2\uff0c\u4ea6\u5373\u4f7f\u5176\u8abf\u8b8a\u983b\u8b5c\u7522\u751f\u5931\u771f\u3002\u4e00\u4e9b\u904e\u53bb\u91dd\u5c0d\u8abf\u8b8a\u983b\u8b5c\u7684\u7814 \u6642\uff0c\u5c07\u6b64\u6bd4\u4f8b\u8abf\u6574\u56de\u672a\u53d7\u5e72\u64fe\u60c5\u6cc1\u4e0b\u7684\u6bd4\u4f8b\u3002\u7531\u65bc\u8abf\u8b8a\u983b\u8b5c\u53d7\u74b0\u5883\u5e72\u64fe\u6642\u300c\u4f4e\u983b\u4e0b\u964d\uff0c
\u7a76[25,30]\u767c\u73fe\uff0c\u8a9e\u97f3\u8a0a\u865f\u53d7\u5230\u74b0\u5883\u5e72\u64fe\u7684\u5f71\u97ff\u8d8a\u5287\u70c8\uff0c\u4ea6\u5373\u8a0a\u566a\u6bd4(signal-to-noise ratio, \u9ad8\u983b\u62ac\u5347\u300d\u7684\u73fe\u8c61\u5341\u5206\u660e\u986f\uff0c\u82e5\u80fd\u627e\u5230\u9ad8\u983b\u6210\u4efd\u548c\u4f4e\u983b\u6210\u4efd\u9593\u9069\u7576\u7684\u754c\u7dda\uff0c\u6b64\u65b9\u6cd5\u80fd\u6709
SNR)\u8d8a\u4f4e\u7684\u6642\u5019\uff0c\u8abf\u8b8a\u983b\u8b5c\u4e2d\u5c0d\u8a9e\u97f3\u8fa8\u8b58\u6700\u91cd\u8981\u7684 1Hz \u5230 16Hz \u6210\u4efd\u5f37\u5ea6\u8d8a\u53d7\u5230\u58d3\u6291\uff0c \u4e0d\u932f\u7684\u6210\u6548\uff0c\u4e14\u904b\u7b97\u5341\u5206\u5feb\u901f\u3002
\u800c\u504f\u96e2\u4e7e\u6de8\u72c0\u6cc1\u7684\u8abf\u8b8a\u983b\u8b5c\u8d8a\u9060\u3002\u8209\u4f8b\u4f86\u8aaa\uff0c\u5716\u4e00\u662f Aurora-2 \u8a9e\u6599\u5eab\u6240\u6709\u6e2c\u8a66\u96c6\u6885\u723e \u5012\u983b\u8b5c\u7cfb\u6578(Mel-frequency cepstral coefficients, MFCC)[41]\u4e2d c1 \u7cfb\u6578\u7684\u8abf\u8b8a\u983b\u8b5c\u3002\u7531\u65bc 2. \u8abf\u8b8a\u983b\u8b5c\u7d71\u8a08\u5716\u7b49\u5316\u6cd5(spectral histogram equalization, SHE)
\u9664\u4e86\u74b0\u5883\u5e72\u64fe\u5916\uff0c\u5c1a\u6709\u500b\u5225\u8a9e\u8005\u7684\u5dee\u7570\u7b49\u56e0\u7d20\uff0c\u56e0\u6b64\u6b64\u5716\u63a1\u7528\u6e2c\u8a66\u96c6\u4e2d\u6240\u6709\u53e5\u8a9e\u53e5\u8abf\u8b8a
\u983b\u8b5c\u4e4b\u5e73\u5747\u503c\uff0c\u4ee5\u7a81\u986f\u74b0\u5883\u689d\u4ef6\u7684\u4e0d\u540c\uff0c\u964d\u4f4e\u500b\u5225\u8a9e\u53e5\u5dee\u7570\u9020\u6210\u7684\u5f71\u97ff\u3002\u5f9e\u6b64\u5716\u4e2d\u53ef\u4ee5
\u89c0\u5bdf\u5230\uff0c\u7576\u8a0a\u566a\u6bd4\u964d\u4f4e\u6642\uff0c\u6574\u500b\u8abf\u8b8a\u983b\u8b5c\u7684\u6240\u6709\u983b\u5e36\u90fd\u6703\u7522\u751f\u5931\u771f\uff0c\u5c24\u5176\u4ee5\u5305\u542b\u6700\u591a\u8a9e
\u97f3\u5167\u5bb9\u8cc7\u8a0a\u7684\u983b\u5e36\u70ba\u751a\u3002
" }, "TABREF3": { "type_str": "table", "html": null, "num": null, "text": "\uff0c\u6700\u5f8c 13 \u7dad\u5247\u70ba\u524d 13 \u7dad\u7684\u4e8c\u968e\u5dee\u91cf\u4fc2\u6578(acceleration coefficient)\u3002\u672c \u8ad6\u6587\u7684\u5be6\u9a57\u4e2d\uff0c\u64f7\u53d6\u7279\u5fb5\u7684\u904e\u7a0b\u5171\u4f7f\u7528 23 \u7d44\u6885\u723e\u6ffe\u6ce2\u5668(Mel filter)\u3002 PHEQ)\u5177\u6709\u826f\u597d\u7684\u4e92\u88dc\u6027[29]\u3002\u9032\u4e00\u6b65\u5c07 PSHE \u904b\u7528\u5728\u7d93 CMVN \u6216 HEQ \u6b63\u898f\u5316\u5f8c\u7684\u7279\u5fb5\u4e0a\uff0c\u53ef\u4ee5\u7372\u5f97\u76f8\u7576\u7a81\u51fa\u7684\u6210\u679c\uff0c\u5176\u6548\u80fd\u751a\u81f3\u9ad8\u65bc ST-PHEQ\u3002\u4f9d\u9019\u6a23\u7684 \u7d50\u679c\u4f86\u770b\uff0c\u986f\u7136\u4f7f\u7528\u8abf\u8b8a\u983b\u8b5c\u9019\u7a2e\u63cf\u8ff0\u8a9e\u53e5\u6574\u9ad4\u8b8a\u5316\u8cc7\u8a0a\u7684\u8868\u793a\u6cd5\u662f\u6709\u5176\u91cd\u8981\u6027\u7684\u3002\u53e6 \u5916\uff0c\u5728\u96dc\u8a0a\u7684\u5e72\u64fe\u76f8\u7576\u56b4\u91cd\u7684\u74b0\u5883\u4e0b(\u5982\u8a0a\u566a\u6bd4\u70ba-5dB \u7684\u60c5\u6cc1)\uff0c\u61c9\u7528 PSHE \u5f8c\uff0c\u5176\u6539\u5584 \u7684\u8f3b\u5ea6\u591a\u65bc\u5728\u6240\u6709\u74b0\u5883\u4e0b\u7684\u5e73\u5747\u60c5\u6cc1\uff0c\u751a\u81f3\u5728\u540c\u6642\u61c9\u7528 HEQ+PSHE \u7684\u60c5\u6cc1\u4e0b\uff0c\u8a0a\u566a\u6bd4 -5dB \u7684\u8fa8\u8b58\u6b63\u78ba\u7387\u9ad8\u9054\u539f\u59cb MFCC \u7279\u5fb5\u7684 8 \u500d\u4ee5\u4e0a\u3002\u6b64\u7d50\u679c\u8aaa\u660e\u4e86\u8abf\u8b8a\u983b\u8b5c\u78ba\u5be6\u80fd\u6355 \u6349\u5230\u4e00\u4e9b\u7121\u6cd5\u76f4\u63a5\u900f\u904e\u6b63\u898f\u5316\u8a9e\u97f3\u7279\u5fb5\u6539\u5584\u7684\u554f\u984c\uff0c\u5c24\u4ee5\u5728\u96dc\u8a0a\u8f03\u5f37\u6642\u70ba\u751a\u3002 \u672c\u8ad6\u6587\u6240\u63d0\u51fa\u7684\u65b9\u6cd5\uff0c\u5176\u5be6\u9a57\u7d50\u679c\u5247\u5217\u5728\u8868\u4e8c\u4e2d\u3002\u8207\u539f\u672c\u7684 PSHE \u76f8\u8f03\uff0c\u91dd\u5c0d\u5176\u6b63 \u898f\u5316\u5f8c\u7684\u7279\u5fb5\u9032\u884c\u5206\u983b\u5e36\u7684\u6b63\u898f\u5316\uff0c\u7121\u8ad6\u4ee5\u4f55\u7a2e\u9806\u5e8f\u7d44\u5408\u6642\u57df\u8207\u7a7a\u9593\u57df\u5169\u500b\u5143\u7d20\uff0c\u90fd\u80fd \u53d6\u5f97\u66f4\u597d\u7684\u7d50\u679c\uff0c\u9019\u986f\u793a\u4e86 PSHE \u96d6\u7136\u80fd\u5920\u4f7f\u8abf\u8b8a\u983b\u8b5c\u4e0a\u7684\u5206\u4f48\u8b8a\u5f97\u4e00\u81f4\uff0c\u4f46\u5728\u6642\u57df\u8207 \u7a7a\u9593\u57df\u9ad8\u4f4e\u983b\u6210\u4efd\u7684\u8abf\u8b8a\u983b\u8b5c\u4e2d\u4ecd\u7136\u5b58\u5728\u8457\u4e00\u4e9b\u672a\u88ab\u6d88\u9664\u7684\u5e72\u64fe\uff0c\u85c9\u7531\u5c07\u9019\u4e9b\u6210\u4efd\u4e5f\u7d0d \u5165\u6b63\u898f\u5316\u7684\u7bc4\u570d\uff0c\u53ef\u4ee5\u88dc\u8db3 PSHE \u9019\u4e00\u9ede\u4e0d\u8db3\u4e4b\u8655\u3002\u5728\u5716\u4e09\u4e2d\uff0c\u6211\u5011\u4ee5\u7a7a\u9593\u57df\u9ad8\u983b\u6210\u4efd Proceedings of the Twenty-Fifth Conference on Computational Linguistics and Speech Processing (ROCLING 2013) \u70ba\u4f8b\uff0c\u986f\u793a\u4e86\u5373\u4f7f PSHE \u5df2\u5c07\u5168\u983b\u5e36\u7279\u5fb5\u7684\u8abf\u8b8a\u983b\u8b5c\u8b8a\u5f97\u8f03\u70ba\u4e00\u81f4\uff0c\u5728\u5b50\u983b\u5e36\u7279\u5fb5\u7684\u8abf \u8b8a\u983b\u8b5c\u4e2d\uff0c\u4ecd\u7136\u5b58\u5728\u8457\u56e0\u70ba\u96dc\u8a0a\u800c\u7522\u751f\u7684\u5931\u771f\uff1b\u800c\u9019\u500b\u5931\u771f\u5728\u7d93\u904e ST-PSHE \u7684\u8655\u7406\u4ee5 \u5f8c\uff0c\u5247\u6709\u986f\u8457\u7684\u6539\u5584\uff0c\u4e26\u9054\u5230\u8ddf\u5168\u983b\u5e36\u7684\u8abf\u8b8a\u983b\u8b5c\u76f8\u8fd1\u7684\u4e00\u81f4\u7a0b\u5ea6\u3002\u53e6\u5916\uff0c\u55ae\u7368\u5728\u7a7a\u9593 \u57df\u4e0a\u6216\u662f\u6642\u57df\u4e0a\u9032\u884c\u5206\u983b\u7684\u6b63\u898f\u5316\uff0c\u90fd\u80fd\u5920\u76f8\u5c0d\u5730\u6e1b\u5c11\u5927\u7d04 1.3%\u7684\u5b57\u932f\u8aa4\u7387(word error rate)\uff0c\u800c\u4f9d\u7167\u7a7a\u9593\u57df-\u6642\u57df\u7684\u9806\u5e8f\u9032\u884c\u5206\u983b\u6b63\u898f\u5316\uff0c\u66f4\u80fd\u5920\u76f8\u5c0d\u6e1b\u5c11 2.8%\u7684\u932f\u8aa4\u3002\u4f46 \u82e5\u5c07\u9806\u5e8f\u53cd\u904e\u4f86\uff0c\u4f9d\u7167\u6642\u57df-\u7a7a\u9593\u57df\u7684\u9806\u5e8f\u9032\u884c\uff0c\u5247\u6539\u9032\u7684\u8f3b\u5ea6\u53cd\u800c\u8b8a\u5f97\u975e\u5e38\u6709\u9650\u3002 \u524d\u6587\u4e2d\u63d0\u5230\u5728\u8abf\u8b8a\u983b\u8b5c\u4e0a\u7684\u6b63\u898f\u5316\u65b9\u6cd5\uff0c\u82e5\u8207\u5728\u7279\u5fb5\u6642\u57df\u4e0a\u7684\u6b63\u898f\u5316\u65b9\u6cd5\u7d50\u5408\uff0c\u6703 \u7522\u751f\u5f88\u660e\u986f\u7684\u4e92\u88dc\u6548\u61c9\uff0c\u800c\u4f7f\u8fa8\u8b58\u7387\u5927\u8f3b\u4e0a\u5347\u3002\u56e0\u6b64\u5728\u8868\u4e09\u7576\u4e2d\uff0c\u6211\u5011\u4e5f\u5617\u8a66\u5c07 ST-PSHE \u8207 CMVN\u3001HEQ \u4ee5\u53ca\u540c\u6a23\u61c9\u7528\u6642\u57df\u53ca\u7a7a\u9593\u57df\u6587\u8108\u8cc7\u8a0a\u9032\u884c\u5206\u983b\u7684 ST-PHEQ \u9032\u884c\u7d50\u5408\uff0c \u63a2\u7d22\u8207\u9019\u4e9b\u65b9\u6cd5\u7d50\u5408\u7684\u6548\u679c\u3002\u7531\u65bc\u8abf\u8b8a\u983b\u8b5c\u96d6\u7136\u6293\u4f4f\u4e86\u6574\u500b\u8a9e\u53e5\u7684\u7279\u5fb5\u8b8a\u5316\u6a21\u5f0f\uff0c\u4f46\u5c0d \u65bc\u6bd4\u8f03\u5340\u57df\u6027\u7684\u96dc\u8a0a\u5e72\u64fe\u53ca\u500b\u5225\u97f3\u6846\u7684\u626d\u66f2\u5247 \u8f03\u96e3\u8a73\u76e1\u5730\u63cf\u8ff0\uff0c\u56e0\u6b64\u82e5\u80fd\u5728\u9032\u884c ST-PSHE \u524d\u5148\u5229\u7528\u7279\u5fb5\u4e0a\u7684\u6b63\u898f\u5316\u65b9\u6cd5 CMVN \u53ca HEQ \u8655\u7406\u904e\uff0c\u5247\u80fd\u540c\u6642\u6b63\u898f\u5316\u6574\u9ad4 \u8b8a\u5316\u6a21\u5f0f\u53ca\u500b\u5225\u97f3\u6846\u7684\u6578\u503c\uff0c\u8207\u55ae\u7d14\u8655\u7406\u8abf\u8b8a\u983b\u8b5c\u76f8\u8f03\uff0c\u53ef\u4ee5\u53d6\u5f97\u8d85\u904e 36%\u7684\u76f8\u5c0d\u5b57\u932f \u8aa4\u7387\u6e1b\u5c11\u3002\u800c\u82e5\u5728\u9032\u884c ST-PSHE \u4e4b\u524d\u5148\u4f7f\u7528 ST-PHEQ \u8655\u7406\u904e\u4e00\u6b21\uff0c\u96d6\u7136\u540c\u6a23\u662f\u904b\u7528\u5206 \u983b\u53d6\u5f97\u6587\u8108\u7684\u6982\u5ff5\u9032\u884c\uff0c\u4f46\u7531\u65bc\u8655\u7406\u7684\u9762\u5411\u4e0d\u540c\uff0c\u56e0\u6b64\u4ecd\u7136\u6709\u5f88\u5927\u7684\u4e92\u88dc\u6210\u4efd\u5b58\u5728\uff0c\u5176 \u7d50\u679c\u8f03\u55ae\u7368\u4f7f\u7528 ST-PSHE \u76f8\u5c0d\u6e1b\u5c11\u4e86 36.8%\u7684\u8fa8\u8b58\u932f\u8aa4\uff0c\u8207 ST-PHEQ \u6bd4\u8f03\u4e5f\u76f8\u5c0d\u964d\u4f4e \u4e86 14.4%\u7684\u5b57\u932f\u8aa4\u7387\u3002 \u6700\u5f8c\uff0c\u6211\u5011\u4e5f\u5c07\u672c\u8ad6\u6587\u6240\u63d0\u51fa\u7684\u65b9\u6cd5\u8207\u6b50\u6d32\u96fb\u4fe1\u5354\u6703(European telecommunications standards institute, ETSI)\u767c\u5c55\u7684 AFE (advanced front end)[44]\u9032\u884c\u6bd4\u8f03\u3002\u5982\u8868\u56db\u6240\u793a\uff0c\u7531 \u65bc AFE \u5305\u542b\u4e86\u8f03\u8907\u96dc\u7684\u8a9e\u97f3\u6d3b\u52d5\u5075\u6e2c(voice-activity detection, VAD)\u53ca\u566a\u97f3\u6291\u5236(noise reduction)\u7684\u6280\u8853\uff0cAFE \u7684\u8fa8\u8b58\u6b63\u78ba\u7387\u76f8\u8f03\u65bc ST-PSHE \u660e\u986f\u662f\u8f03\u597d\u7684\uff1b\u4f46\u9032\u4e00\u6b65\u5c07 AFE \u7684\u7279\u5fb5\u65bd\u4ee5 ST-PSHE \u7684\u8655\u7406\uff0c\u4e26\u5c07\u4e4b\u8207\u539f\u672c\u7684 AFE \u7279\u5fb5\u7dda\u6027\u7d50\u5408\u4e4b\u5f8c\uff0c\u4ecd\u7136\u80fd\u5920\u76f8\u5c0d \u5730\u6e1b\u5c11\u5927\u7d04 2.7%\u7684\u8fa8\u8b58\u932f\u8aa4\uff0c\u986f\u793a\u9019\u5169\u6a23\u6280\u8853\u5f7c\u6b64\u4ecd\u7136\u6709\u80fd\u5920\u4e92\u88dc\u7684\u5c64\u9762\u5b58\u5728\u3002\u503c\u5f97 \u6ce8\u610f\u7684\u662f\uff0c\u4ee5 ST-PSHE \u8655\u7406\u5f8c\u7684 MFCC \u7279\u5fb5\u96d6\u7136\u5e73\u5747\u7684\u8fa8\u8b58\u6b63\u78ba\u7387\u4e0d\u5982 AFE\uff0c\u4f46\u5728\u6975 \u7aef\u7684\u566a\u97f3\u74b0\u5883\u4e0b(\u8a0a\u566a\u6bd4-5dB)\u53cd\u800c\u80fd\u53d6\u5f97\u8f03\u597d\u7684\u6548\u679c\uff0c\u518d\u6b21\u986f\u793a\u8abf\u8b8a\u983b\u8b5c\u7684\u6b63\u898f\u5316\u5c0d\u65bc \u56b4\u91cd\u7684\u96dc\u8a0a\u5e72\u64fe\u662f\u5f88\u6709\u6548\u7684\u3002 \u516d\u3001\u7d50\u8ad6 \u5728\u672c\u8ad6\u6587\u4e2d\uff0c\u6211\u5011\u63a2\u8a0e\u4e86\u4f7f\u7528\u5c07\u8a9e\u97f3\u7279\u5fb5\u5728\u6642\u57df\u8207\u7a7a\u9593\u57df\u9032\u884c\u5206\u983b\u7684\u65b9\u5f0f\u4ee5\u53d6\u5f97\u6587\u8108\u8cc7 \u8a0a\uff0c\u9032\u800c\u6e1b\u7de9\u50b3\u7d71 SHE \u4ee5\u53ca PSHE \u7684\u56b4\u683c\u9650\u5236\u3002ST-PSHE \u548c\u50b3\u7d71\u7684\u65b9\u6cd5\u76f8\u8f03\uff0c\u4e0d\u50c5\u5168 \u983b\u5e36\u7684\u8abf\u8b8a\u983b\u8b5c\u5177\u6709\u4e00\u81f4\u7684\u5206\u4f48\uff0c\u9ad8\u983b\u6210\u4efd\u8207\u4f4e\u983b\u6210\u4efd\u7684\u8abf\u8b8a\u983b\u8b5c\u5206\u4f48\u4e5f\u7d0d\u5165\u6b63\u898f\u5316\u7684 \u7bc4\u570d\uff0c\u9032\u4e00\u6b65\u5730\u6e1b\u5c11\u4e86\u96dc\u8a0a\u5c0d\u8abf\u8b8a\u983b\u8b5c\u7684\u5e72\u64fe\u3002\u5be6\u9a57\u7684\u7d50\u679c\u4e5f\u8aaa\u660e\u4e86\u672c\u8ad6\u6587\u6240\u63d0\u51fa\u7684\u65b9 \u6cd5\u78ba\u5be6\u80fd\u5920\u9054\u6210\u8f03\u9ad8\u7684\u8fa8\u8b58\u6b63\u78ba\u7387\u8868\u73fe\uff0c\u4e26\u80fd\u5920\u8207\u5176\u4ed6\u7279\u5fb5\u6b63\u898f\u5316\u7684\u65b9\u6cd5\u4e92\u88dc\u3002 \u5c55\u671b\u672a\u4f86\u7814\u7a76\uff0c\u6211\u5011\u63d0\u51fa\u5169\u9ede\u53ef\u80fd\u7684\u65b9\u5411\u3002\u7b2c\u4e00\u662f\u5c07\u6b64\u6280\u8853\u61c9\u7528\u5230\u66f4\u8907\u96dc\u7684\u8a9e\u97f3\u8fa8 \u8b58\u4efb\u52d9\u4e0a\uff0c\u5982\u5c6c\u65bc\u5927\u8a5e\u5f59\u9023\u7e8c\u8a9e\u97f3\u8fa8\u8b58(large vocabulary continuous speech recognition, LVCSR)\u7684 Aurora-4 \u8a9e\u6599\u5eab[45]\u548c MATBN \u8a9e\u6599\u5eab[46]\u4e0a\uff0c\u4ee5\u66f4\u9032\u4e00\u6b65\u9a57\u8b49\u6211\u5011\u6240\u63d0\u51fa\u4e4b \u65b9\u6cd5\u662f\u5426\u5728\u8f03\u8907\u96dc\u7684\u8a9e\u97f3\u8fa8\u8b58\u4efb\u52d9\u4e0a\u4e5f\u80fd\u5920\u6709\u76f8\u540c\u7684\u8868\u73fe\u3002\u7b2c\u4e8c\u662f\u5728\u6574\u500b\u8a9e\u53e5\u7684\u8abf\u8b8a\u983b \u8b5c\u4e4b\u5916\uff0c\u66f4\u6df1\u5165\u5730\u63a2\u8a0e\u904b\u7528\u4e0d\u540c\u7684\u5206\u6790\u55ae\u4f4d\u8655\u7406\u8abf\u8b8a\u983b\u8b5c\uff0c\u4ee5\u671f\u80fd\u6355\u6349\u66f4\u591a\u5c64\u9762\u7684\u8cc7\u8a0a \u800c\u9032\u4e00\u6b65\u63d0\u5347\u8a9e\u97f3\u8fa8\u8b58\u7684\u5f37\u5065\u6027\uff0c\u4e26\u4f7f\u6b64\u65b9\u6cd5\u80fd\u5920\u61c9\u7528\u5728\u5be6\u6642(real-time)\u7684\u7cfb\u7d71\u4e2d\u3002", "content": "
\u8868\u4e00\u3001\u5404\u7a2e\u57fa\u790e\u7279\u5fb5\u53ca\u5f37\u5065\u6027\u6280\u8853\u7684\u8fa8\u8b58\u6b63\u78ba\u7387(%) \u8a0a\u566a\u6bd4 \u4e7e\u6de8 20dB 15dB 10dB 5dB 0dB -5dB 99.71 92.44 80.56 58.61 30.04 9.31 3.39 99.72 98.13 94.27 80.45 50.64 23.81 13.04 \u8868\u4e09\u3001ST-PSHE \u8207\u5176\u4ed6\u5f37\u5065\u6027\u6280\u8853\u7d50\u5408\u4e4b\u8fa8\u8b58\u6b63\u78ba\u7387(%) MFCC \u7279\u5fb5 CMS \u7279\u5fb5 \u8a0a\u566a\u6bd4 \u4e7e\u6de8 20dB 15dB 10dB 5dB 0dB (delta coefficient)\u8a55\u4f30\u8a9e\u97f3\u7279\u5fb5\u6240\u4f7f\u7528\u7684\u8072\u5b78\u6a21\u578b\u8a13\u7df4\u53ca\u8fa8\u8b58\uff0c\u7686\u4f7f\u7528 HTK \u5957\u4ef6[43]\u5b8c\u6210\u3002\u5176\u4e2d\u6bcf \u5e73\u5747\u503c 54.19 \u5e73\u5747\u503c -5dB CMVN+ST-PSHE 99.45 98.44 96.82 92.8 82.01 58.44 29.39 \u500b\u6578\u5b57\u7686\u7531\u4e00\u500b\u7531\u5de6\u5230\u53f3\u5f62\u5f0f\u7684\u9023\u7e8c\u5bc6\u5ea6\u96b1\u85cf\u5f0f\u99ac\u53ef\u592b\u6a21\u578b(continuous density hidden 85.70 Markov model, CDHMM)\u8868\u793a\uff0c\u6bcf\u500b\u6a21\u578b\u6263\u9664\u524d\u5f8c\u4e4b\u929c\u63a5\u7528\u72c0\u614b(state)\u5171\u6709 16 \u500b\u72c0\u614b\uff0c 69.46 PHEQ+ST-PSHE 99.41 98.28 96.59 92.44 82.03 59.13 29.32 85.69 \u6bcf\u500b\u72c0\u614b\u4ee5\u542b 20 \u500b\u9ad8\u65af\u6df7\u5408(Gaussian mixture)\u7684\u9ad8\u65af\u6df7\u5408\u6a21\u578b(Gaussian mixture model,
CMVN ST-PHEQ+ST-PSHE 99.37 98.12 96.42 92.28 82.16 60.08 30.98 99.69 97.97 94.98 87.25 67.52 34.87 13.73 GMM)\u8868\u793a\u3002\u975c\u97f3(silence)\u6a21\u578b\u5247\u70ba 3 \u500b\u72c0\u614b\u548c 36 \u500b\u9ad8\u65af\u6df7\u5408\u300276.52 85.81
MVA PHEQ (\u4e09)\u8fa8\u8b58\u6548\u80fd\u8a55\u4f30\u65b9\u5f0f 99.66 97.96 95.98 90.27 76.46 50.70 22.86 99.65 98.52 96.56 91.19 75.78 45.39 18.14 ST-PHEQ 99.58 98.59 96.99 92.26 78.95 50.36 20.04 \u8868\u56db\u3001ST-PSHE \u8207 AFE \u6bd4\u8f03\u53ca\u7d50\u5408\u7684\u8fa8\u8b58\u6b63\u78ba\u7387(%) PSHE 99.47 97.55 94.29 86.54 68.54 37.58 16.09 CMVN+PSHE 99.56 98.38 96.59 92.26 80.63 56.24 26.93 \u7279\u5fb5 \u672c\u8ad6\u6587\u8fa8\u8b58\u6548\u80fd\u8a55\u4f30\u7684\u65b9\u6cd5\u63a1\u7528\u7f8e\u570b\u6a19\u6e96\u8207\u79d1\u6280\u7d44\u7e54(The National Institute of Standards 82.27 81.49 83.43 76.90 \u8a0a\u566a\u6bd4 and Technology, NIST)\u6240\u8a02\u5b9a\u4e4b\u7528\u4ee5\u8a55\u4f30\u8f49\u8b6f\u6587\u53e5\u8207\u6b63\u78ba\u6587\u53e5\u6bd4\u8f03\u7684\u6a19\u6e96\u3002\u8a55\u4f30\u7684\u6307\u6a19 \u5e73\u5747\u503c \u4e7e\u6de8 20dB 15dB 10dB 5dB 0dB -5dB \u70ba\u8a5e\u6b63\u78ba\u7387(word accuracy)\uff0c\u8a08\u7b97\u65b9\u5f0f\u5982\u4e0b\uff1a 84.82 PHEQ+PSHE 99.45 98.39 96.61 92.71 82.05 58.75 28.34 AFE 99.74 98.89 97.68 94.27 85.47 62.54 30.26 87.77 85.70 AFE+ST-PSHE 99.70 98.82 97.64 94.28 85.89 63.86 32.22 88.10 \u8a5e\u6b63\u78ba\u7387 = \u8a5e\u6b63\u78ba\u8fa8\u8b58\u500b\u6578 \u2212 \u8a5e\u63d2\u5165\u500b\u6578 \u2212 \u8a5e\u522a\u9664\u500b\u6578 \u6b64\u53e5\u4e2d\u8a5e\u7684\u7e3d\u6578 (14)
, -\u5247\u5206\u5225\u4ee3\u8868\u7a7a\u9593\u57df\u9ad8\u983b\u3001\u7a7a\u9593\u57df \u5c0d\u65bc\u6bcf\u4e00\u500b\u8a9e\u53e5\uff0c\u5728\u9032\u884c\u4e86\u4e00\u6b21 PSHE \u4e4b\u5f8c\uff0c\u5176\u5168\u983b\u5e36(full-band)\u7684\u8abf\u8b8a\u983b\u8b5c\u5df2\u7d93 \u4f4e\u983b\u3001\u6642\u57df\u9ad8\u983b\u3001\u6642\u57df\u4f4e\u983b\u7684\u5b50\u983b\u5e36\u6210\u4efd\u7279\u5fb5\u3002 \u5177\u6709\u548c\u8a13\u7df4\u8a9e\u6599\u7684\u8abf\u8b8a\u983b\u8b5c\u76f8\u540c\u7684\u5206\u4f48\uff0c\u4f46\u6642\u57df\u6216\u7a7a\u9593\u57df\u4e0a\u7684\u9ad8\u4f4e\u983b\u6210\u4efd\u537b\u9084\u662f\u6709\u4e00\u90e8 \u4efd\u7684\u4e0d\u5339\u914d\u73fe\u8c61\u3002\u56e0\u6b64\u5728\u9032\u884c PSHE \u4ee5\u5f8c\uff0c\u8981\u5c07\u8655\u7406\u5f8c\u7684\u7279\u5fb5\u4f9d\u5f0f(8)\u53ca\u5f0f(9)\u5728\u7a7a\u9593\u57df \u4e0a\u5206\u70ba\u9ad8\u983b\u7279\u5fb5\u8207\u4f4e\u983b\u7279\u5fb5\uff0c\u5c07\u6b64\u5169\u500b\u983b\u5e36\u7684\u7279\u5fb5\u5206\u5225\u6c42\u53d6\u5176\u8abf\u8b8a\u983b\u8b5c\u4e26\u4ee5 PSHE \u6b63\u898f \u5316\u4e26\u7531\u8abf\u8b8a\u983b\u8b5c\u9084\u539f\u56de\u7279\u5fb5\u57df\u4e4b\u5f8c\uff0c\u518d\u4f9d\u4e0b\u5f0f\u5c07\u7a7a\u9593\u57df\u9ad8\u4f4e\u983b\u6210\u4efd\u7d50\u5408\uff1a \u0302, -=\u0302s ,hp , -+\u0302s ,lp , -(12) \u5176\u4e2d\u0302s ,hp , -\u70ba\u7a7a\u9593\u57df\u9ad8\u983b\u6210\u4efd\u7d93 PSHE \u6b63\u898f\u5316\u5f8c\u4e4b\u7279\u5fb5\uff0c\u0302s ,lp , -\u5247\u70ba\u7a7a\u9593\u57df\u4f4e\u983b\u6210\u4efd \u7d93 PSHE \u6b63\u898f\u5316\u5f8c\u4e4b\u7279\u5fb5\u3002\u7531\u65bc\u5f0f(8)\u8207\u5f0f(9)\u7684\u8a2d\u8a08\u4f7f\u5f97\u6b64\u5169\u500b\u983b\u5e36\u5177\u6709\u4e92\u88dc\u95dc\u4fc2\uff0c\u6545 \u5728\u50b3\u7d71\u7684 HEQ \u6216\u662f SHE \u4e2d\uff0c\u90fd\u5047\u8a2d\u96dc\u8a0a\u5c0d\u65bc\u8a9e\u97f3\u53ea\u5177\u6709\u55ae\u8abf(monotonic)\u7684\u7684\u5e72\u64fe\uff0c \u4ea6\u5373\u6703\u6539\u8b8a\u7279\u5fb5\u6216\u8abf\u8b8a\u983b\u8b5c\u4e2d\u6240\u6709\u6578\u503c\u7684\u5927\u5c0f\uff0c\u4f46\u5404\u6578\u503c\u4e4b\u9593\u7684\u76f8\u5c0d\u6392\u5e8f(ordering)\u662f\u7dad \u6301\u4e0d\u8b8a\u7684\u3002ST-PSHE \u9664\u4e86\u6253\u7834\u6642\u57df\u53ca\u7a7a\u9593\u57df\u4e0a\u7684\u7368\u7acb\u5047\u8a2d\u4ee5\u5916\uff0c\u6b64\u7a2e\u5c07\u9ad8\u4f4e\u983b\u5206\u5225\u6b63\u898f \u5316\u518d\u7d50\u5408\u7684\u65b9\u5f0f\u4e5f\u53ef\u80fd\u6703\u6539\u8b8a\u8abf\u8b8a\u983b\u8b5c\u4e0d\u540c\u983b\u7387\u5f37\u5ea6\u7684\u5927\u5c0f\u9806\u5e8f\uff0c\u800c\u4f7f\u5f97\u975e\u55ae\u8abf\u7684\u5e72\u64fe \u8868\u4e8c\u3001PSHE \u7d50\u5408\u7a7a\u9593\u57df\u6216\u6642\u57df\u6587\u8108\u8cc7\u8a0a\u7684\u8fa8\u8b58\u6b63\u78ba\u7387(%) \u7279\u5fb5 \u8a0a\u566a\u6bd4 \u53e6\u5916\uff0c\u672c\u8ad6\u6587\u4e2d\u975c\u97f3\u8a5e(silence \u548c short pause)\u5c07\u4e0d\u5217\u5165\u8a5e\u6b63\u78ba\u7387\u7684\u8a08\u7b97\u3002\u800c\u5728 Aurora-2 \u4e94\u3001\u5be6\u9a57\u8207\u5206\u6790 \u8a9e\u6599\u5eab\u7684\u8a2d\u5b9a\u4e2d\uff0c\u6bcf\u4e00\u500b\u6e2c\u8a66\u5b50\u96c6\u7684\u5e73\u5747\u8fa8\u8b58\u7387\uff0c\u53ea\u4ee5 0dB(\u542b)\u5230 20dB(\u542b)\u9593\u7684\u8fa8\u8b58\u7cbe \u5e73\u5747\u503c \u4e7e\u6de8 20dB 15dB 10dB 5dB 0dB -5dB S-PSHE 99.39 97.29 93.69 85.73 68.77 40.17 17.75 \u78ba\u7387\u8a08\u7b97\u5e73\u5747\u3002\u672c\u8ad6\u6587\u4ea6\u4ee5\u6b64\u8a08\u7b97\u65b9\u5f0f\u8a55\u4f30\u8fa8\u8b58\u6548\u80fd\u3002 (\u4e00)\u5be6\u9a57\u8a9e\u6599\u5eab 77.19 T-PSHE 99.41 97.31 93.78 85.90 68.89 40.18 17.68 \u672c\u8ad6\u6587\u7684\u5be6\u9a57\u6240\u4f7f\u7528\u7684\u8a9e\u6599\u5eab\u70ba Aurora-2 \u82f1\u6587\u9023\u7e8c\u6578\u5b57\u8a9e\u6599\u5eab[42]\uff0c\u6b64\u8a9e\u6599\u5eab\u7531\u6b50\u6d32\u96fb (\u56db)\u5be6\u9a57\u7d50\u679c\u8207\u8a0e\u8ad6 77.21 TS-PSHE 99.45 97.10 93.44 85.50 68.65 39.96 17.13 76.93 ST-PSHE 99.28 97.28 94.21 86.70 69.48 40.06 17.71 77.55 \u4fe1\u6a19\u6e96\u5354\u6703(European Telecommunications Standards Institute, ESTI)\u6240\u767c\u884c\uff0c\u5167\u5bb9\u7686\u662f\u7531 \u9996\u5148\uff0c\u4f5c\u70ba\u6bd4\u8f03\u7684\u57fa\u6e96\uff0c\u6211\u5011\u5728\u8868\u4e00\u4e2d\u5217\u51fa\u4e86 MFCC \u7279\u5fb5\u53ca\u4e00\u4e9b\u57fa\u790e\u5f37\u5065\u6027\u8a9e\u97f3\u8fa8\u8b58 \u7f8e\u570b\u6210\u5e74\u4eba\u9304\u88fd\u7684\u9023\u7e8c\u6578\u5b57\u3002\u6b64\u8a9e\u6599\u5eab\u5305\u542b G.712 \u548c MIRS \u5169\u7a2e\u4e0d\u540c\u7684\u901a\u9053\u6548\u61c9\uff0c\u53ca\u6a5f \u6280\u8853\u7684\u8fa8\u8b58\u6b63\u78ba\u7387\u3002\u5176\u4e2d PHEQ \u53ca PSHE \u7684\u591a\u9805\u5f0f\u968e\u6578\u5747\u662f\u6839\u64da Aurora-2 \u8a9e\u6599\u5eab\u9032\u884c \u5834\u3001\u4eba\u8072\u3001\u6c7d\uf902\u3001\u5c55\u89bd\u6703\ufa2c\u3001\u9910\u5ef3\u3001\u5730\u4e0b\u9435\u3001\u8857\u9053\u3001\u706b\uf902\u7ad9\u7b49\u516b\u7a2e\u52a0\u6210\u6027\u566a\u97f3\uff0c\u52a0\u6210\u6027 \u6311\u9078\u4e4b\u6700\u4f73\u8a2d\u5b9a\u503c\uff0c\u672c\u8ad6\u6587\u5f8c\u7e8c\u5be6\u9a57\u7686\u4f9d\u5faa\u6b64\u7d44\u8a2d\u5b9a\uff0c\u800c\u4e0d\u53e6\u884c\u6700\u4f73\u5316\u591a\u9805\u5f0f\u968e\u6578\u3002\u800c \u566a\u97f3\u5206\u5225\u4ee5\u4e7e\u6de8\u300120dB\u300115dB\u300110dB\u30015dB\u30010dB\u3001-5dB \u7b49\u4e03\u7a2e\u4e0d\u540c\u7684\u8a0a\u566a\u6bd4\u6df7\u5165\u8a9e\u97f3 \u7531\u8868\u4e00\u4e2d\u4e5f\u53ef\u4ee5\u767c\u73fe\uff1a\u7531\u65bc PHEQ \u975e\u7dda\u6027\u8f49\u63db\u7684\u7279\u6027\uff0c\u6bd4\u8d77\u4f7f\u7528\u7dda\u6027\u8f49\u63db\u7684 CMS \u53ca \u4e2d\u3002\u6b64\u8a9e\u6599\u5eab\u542b\u6709\u5169\u7d44\u4e0d\u540c\u7684\u8a13\u7df4\u8a9e\u6599\uff0c\u5206\u5225\u6709 8,440 \u53e5\u7684\u8a13\u7df4\u8a9e\u53e5\u3002\u5728\u4e7e\u6de8\u8a13\u7df4 CMVN \u80fd\u5920\u88dc\u511f\u66f4\u591a\u96dc\u8a0a\u9020\u6210\u7684\u5e72\u64fe\uff0c\u5728\u8fa8\u8b58\u6b63\u78ba\u7387\u4e0a\u6709\u8f03\u597d\u7684\u8868\u73fe\uff0c\u800c\u540c\u6a23\u5f15\u5165\u6642 (clean-condition training) \u8a9e \u6599 \u4e2d \uff0c \u6240 \u6709 \u8a9e \u53e5 \u7686 \u4e7e \u6de8 \u4e0d \u542b \u4efb \u4f55 \u566a \u97f3 \uff1b \u800c \u5728 \u8907 \u5408 \u60c5 \u5883 \u57df\u53ca\u7a7a\u9593\u57df\u6587\u8108\u8cc7\u8a0a\u9032\u884c\u5206\u983b\u7684 ST-PHEQ\uff0c\u76f8\u8f03\u65bc\u539f\u672c\u7684 PHEQ \u4ea6\u6709\u5927\u8f3b\u7684\u6539\u9032\uff0c\u986f (multi-condition training)\u8a13\u7df4\u8a9e\u6599\u4e2d\uff0c\u542b\u6709\u53ca\u5730\u4e0b\u9435\u3001\u4eba\u8072\u3001\u6c7d\uf902\u3001\u5c55\u89bd\u6703\ufa2c\u7b49\u56db\u7a2e\u566a \u97f3\uff0c\u5176\u8a0a\u566a\u6bd4\u7531 5dB \u5230 20dB \u5916\u52a0\u4e7e\u6de8\u8a9e\u97f3\uff0c\u5169\u7d44\u8a13\u7df4\u8a9e\u6599\u7686\u542b G.712 \u901a\u9053\u6548\u61c9\u3002\u672c\u8ad6 \u793a\u9019\u4e9b\u6587\u8108\u8cc7\u8a0a\u5c0d\u65bc\u8a9e\u97f3\u8fa8\u8b58\u7684\u5f37\u5065\u6027\u6709\u5de8\u5927\u7684\u5e6b\u52a9\u3002 \u80fd\u5920\u4e00\u4f75\u88ab\u8003\u616e\u9032\u4f86\u3002\u6709\u9451\u65bc\u6b64\uff0c\u5728\u8a13\u7df4\u968e\u6bb5\u7d71\u8a08\u6642\u57df\u5206\u983b\u90e8\u4efd\u7684\u53c3\u8003\u5206\u4f48\u6642\uff0c\u9700\u8981\u4f7f \u6587\u4e2d\u7684\u5be6\u9a57\u4e00\u5f8b\u4f7f\u7528\u4e7e\u6de8\u8a13\u7df4\u8a9e\u6599\u9032\u884c\u8a13\u7df4\u3002 \u800c\u5728\u8abf\u8b8a\u983b\u8b5c\u7684\u6b63\u898f\u5316\u65b9\u9762\uff0c\u96d6\u7136\u55ae\u7368\u4f7f\u7528 PSHE \u6c92\u6709\u592a\u7a81\u51fa\u7684\u8868\u73fe\uff0c\u4f46\u7531\u65bc PSHE \u7528\u7a7a\u9593\u57df\u5206\u983b\u90e8\u4efd\u5df2\u7d93\u6b63\u898f\u5316\u904e\u7684\u8a9e\u97f3\u7279\u5fb5\u9032\u884c\u7d71\u8a08\uff0c\u800c\u975e\u539f\u59cb\u672a\u7d93\u6b63\u898f\u5316\u7684\u8a9e\u97f3\u7279 \u5fb5\u3002 \u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u672c\u8ad6\u6587\u4e2d\u6642\u57df\u5206\u983b\u7684\u65b9\u6cd5\uff0c\u5176\u6982\u5ff5\u8207\u524d\u4eba\u91dd\u5c0d SHE \u6240\u63d0\u51fa\u7684\u5206\u983b \u6b63\u898f\u5316\u7684\u662f\u6574\u500b\u8a9e\u53e5\u4e2d\u7279\u5fb5\u8b8a\u5316\u7684\u8da8\u52e2\u8207\u898f\u5f8b\uff0c\u8207\u5176\u4ed6\u76f4\u63a5\u8abf\u6574\u8a9e\u97f3\u7279\u5fb5\u6578\u503c\u7684\u65b9\u6cd5 \u5728\u6e2c\u8a66\u8a9e\u6599\u90e8\u4efd\uff0c\u8a0a\u566a\u6bd4\u7bc4\u570d\u7686\u662f\u7531-5dB \u5230 20dB \u5916\u52a0\u4e7e\u6de8\u8a9e\u97f3\u3002\u6e2c\u8a66\u96c6 A \u6709 28,028 \u53e5\uff0c\u5206\u70ba\u56db\u500b\u5b50\u96c6\uff0c\u542b\u6709\u548c\u8907\u5408\u60c5\u5883\u8a13\u7df4\u8a9e\u6599\u4e2d\u76f8\u540c\u7684\u566a\u97f3\u548c\u901a\u9053\u6548\u61c9\uff1b\u6e2c\u8a66\u96c6 B \u6709 (\u5982 CMVN \u8207
\u5c07\u5169\u500b\u983b\u5e36\u7684\u7279\u5fb5\u76f4\u63a5\u76f8\u52a0\u5373\u53ef\u9084\u539f\u56de\u539f\u672c\u5168\u983b\u5e36\u7684\u7279\u5fb5\u3002\u9032\u884c\u5b8c\u7a7a\u9593\u57df\u4e0a\u7684\u5206\u983b\u6b63\u898f \u8655\u7406\u985e\u4f3c\uff0c\u4e26\u5177\u6709\u76f8\u4eff\u7684\u6210\u6548\uff1a\u5728[25]\u4e2d\uff0c\u8abf\u8b8a\u983b\u8b5c\u88ab\u4f9d\u7b49\u6bd4\u97f3\u7a0b(octave)\u7684\u6bd4\u4f8b\u5206\u70ba\u82e5 28,028 \u53e5\uff0c\u5206\u70ba\u56db\u500b\u5b50\u96c6\uff0c\u542b\u6709\u9910\u5ef3\u3001\u6a5f\u5834\u3001\u8857\u9053\u3001\u706b\uf902\u7ad9\u7b49\u56db\u7a2e\u566a\u97f3\uff0c\u4ee5\u53ca\u548c\u8a13\u7df4\u8a9e
\u5316\u4ee5\u5f8c\uff0c\u5c07\u7d50\u5408\u5f8c\u7684\u5168\u983b\u5e36\u7279\u5fb5\u518d\u6b21\u4f9d\u64da\u5f0f(10)\u53ca\u5f0f(11)\u5728\u6642\u57df\u4e0a\u5206\u70ba\u9ad8\u983b\u7279\u5fb5\u8207\u4f4e\u983b \u5e72\u500b\u983b\u5e36\uff0c\u8d8a\u4f4e\u983b\u7684\u6210\u4efd\u8d8a\u52a0\u7d30\u5206\uff0c\u4e26\u91dd\u5c0d\u6bcf\u4e00\u500b\u983b\u5e36\u9032\u884c\u7368\u7acb\u7684 SHE \u8655\u7406\uff1b\u800c\u5728[30] \u6599\u76f8\u540c\u7684\u901a\u9053\u6548\u61c9\uff1b\u6e2c\u8a66\u96c6 C \u6709 14,014 \u53e5\uff0c\u5206\u70ba\u5169\u500b\u5b50\u96c6\uff0c\u542b\u6709\u5730\u4e0b\u9435\u548c\u8857\u9053\u5169\u7a2e\u566a
\u7279\u5fb5\uff0c\u540c\u6a23\u5c07\u6b64\u4e8c\u983b\u5e36\u5206\u5225\u9032\u884c PSHE \u5f8c\uff0c\u5229\u7528\u8207\u7a7a\u9593\u57df\u9ad8\u4f4e\u983b\u7d50\u5408\u76f8\u540c\u7684\u65b9\u5f0f\uff0c\u4f9d\u4e0b \u4e2d\uff0c\u8abf\u8b8a\u983b\u8b5c\u88ab\u756b\u5206\u70ba\u5169\u500b\u983b\u5e36\u7368\u7acb\u9032\u884c SHE \u8655\u7406\uff0c\u800c\u5283\u5206\u7684\u983b\u7387\u5247\u70ba\u53ef\u8abf\u6574\u4e4b\u53c3\u6578\u3002 \u97f3\uff0c\u901a\u9053\u6548\u61c9\u70ba MIRS\u3002\u7531\u65bc\u672c\u8ad6\u6587\u4f7f\u7528\u4e7e\u6de8\u8a13\u7df4\u8a9e\u6599\uff0c\u6240\u6709\u52a0\u6210\u6027\u566a\u97f3\u7686\u662f\u8a13\u7df4\u8a9e\u6599
\u5f0f\u6240\u793a\u5c07\u6642\u57df\u4e4b\u9ad8\u4f4e\u983b\u6210\u4efd\u7d50\u5408\uff1a \u5728\u6b64\u5169\u7a2e\u6280\u8853\u4e2d\uff0c\u5c0d\u983b\u5e36\u7684\u756b\u5206\u90fd\u662f\u76f4\u63a5\u5c07\u67d0\u500b\u7279\u5b9a\u983b\u7387\u4ee5\u4e0b\u53ca\u4ee5\u4e0a\u7684\u6210\u4efd\u756b\u5206\u70ba\u4e0d\u540c \u4e2d\u672a\u66fe\u898b\u904e\uff0c\u800c\u53ea\u6709\u6e2c\u8a66\u96c6 C \u7684\u901a\u9053\u6548\u61c9\u8207\u8a13\u7df4\u8a9e\u6599\u4e0d\u540c\u3002
\u0303, -=\u0303t \u7684\u983b\u5e36\uff1b\u7136\u800c\u672c\u8ad6\u6587\u4e2d\u9032\u884c\u5206\u983b\u7684\u6ffe\u6ce2\u5668\u5728\u9ad8\u983b\u5e36\u8207\u4f4e\u983b\u5e36\u4e4b\u9593\u6709\u91cd\u758a\uff0c\u5728\u9ad8\u4f4e\u983b\u4e4b\u9593 ,hp , -+\u0303t ,lp , -(13) \u6c92\u6709\u4e00\u500b\u78ba\u5207\u7684\u5206\u5272\u9ede\uff0c\u5c07\u9ad8\u4f4e\u983b\u7d50\u5408\u5f8c\u4e5f\u4e0d\u6703\u7522\u751f\u660e\u986f\u7684\u4e0d\u9023\u7e8c\u73fe\u8c61\u3002\u53e6\u5916\uff0c\u672c\u8ad6\u6587 (\u4e8c)\u57fa\u790e\u5be6\u9a57\u8a2d\u5b9a
\u5176\u4e2d\u0303t \u4e2d\u5206\u983b\u7684\u6ffe\u6ce2\u5668\u70ba\u6709\u9650\u8108\u885d\u97ff\u61c9(finite impulse response, FIR)\u6ffe\u6ce2\u5668\uff0c\u5206\u983b\u7684\u904e\u7a0b\u4e0d\u9700\u8f49 ,hp , -\u70ba\u6642\u57df\u9ad8\u983b\u6210\u4efd\u7d93 PSHE \u6b63\u898f\u5316\u5f8c\u4e4b\u7279\u5fb5\uff0c\u0303t ,lp , -\u5247\u70ba\u6642\u57df\u4f4e\u983b\u6210\u4efd\u7d93 \u672c \u8ad6 \u6587 \u7684 \u57fa \u790e \u5be6 \u9a57 \u662f \u63a1 \u7528 \u6885 \u723e \u5012 \u983b \u8b5c \u4fc2 \u6578 [41] \u505a \u70ba \u8a9e \u97f3 \u7279 \u5fb5 \u53c3 \u6578 \uff0c \u5176 \u4e2d \u9810 \u5f37 \u8abf
PSHE \u6b63\u898f\u5316\u5f8c\u4e4b\u7279\u5fb5\uff0c\u7d93\u904e\u6b64\u4e00\u904e\u7a0b\u7522\u751f\u6700\u7d42\u7d93 ST-PSHE \u8655\u7406\u5f8c\u7684\u7279\u5fb5\u3002\u5176\u4e2d\uff0c\u4ea6\u53ef \u63db\u81f3\u8abf\u8b8a\u983b\u8b5c\uff0c\u53ef\u76f4\u63a5\u5728\u7279\u5fb5\u4e0a\u5feb\u901f\u4e26\u7a69\u5b9a(numerical stability)\u5730\u9032\u884c\u5be6\u4f5c\u3002 (pre-emphasis)\u53c3\u6578\u8a2d\u70ba 0.97\uff0c\u7a97\u51fd\u6578(window function)\u70ba\u6f22\u660e\u7a97(Hamming window)\uff0c\u5176
\u4ee5\u9078\u64c7\u8df3\u904e\u6642\u57df\u5206\u983b\u7684\u90e8\u4efd(\u7a31\u70ba S-PSHE)\u3001\u8df3\u904e\u7a7a\u9593\u57df\u5206\u983b\u7684\u90e8\u4efd(\u7a31\u70ba T-PSHE)\u3001\u6216 \u53c3\u6578\u8a2d\u70ba 0.46\uff0c\u53d6\u6a23\u97f3\u6846\u9577\u5ea6\u70ba 25 \u6beb\u79d2\uff0c\u97f3\u6846\u9593\u8ddd(frame shift)\u70ba 10 \u6beb\u79d2\u3002\u6bcf\u500b\u97f3\u6846\u5167
\u662f\u5c07\u6642\u57df\u5206\u983b\u8207\u7a7a\u9593\u57df\u5206\u983b\u5169\u90e8\u4efd\u8abf\u63db\u9806\u5e8f(\u7a31\u70ba TS-PSHE)\uff0c\u6b64\u90e8\u4efd\u7684\u5dee\u7570\u5c07\u65bc\u7b2c\u4e94\u7ae0 \u7684\u8cc7\u8a0a\uff0c\u5728\u5b8c\u6210\u7279\u5fb5\u64f7\u53d6\u4ee5\u5f8c\u7531 39 \u7dad\u7684\u8a9e\u97f3\u7279\u5fb5\u5411\u91cf\u8868\u793a\u3002\u5176\u4e2d\u524d 13 \u7dad\u70ba\u6885\u723e\u5012\u983b\u8b5c
\u4e2d\u63a2\u8a0e\u3002 \u4fc2\u6578\u7684\u524d 12 \u9805(c1~c12)\u53ca\u7b2c\u96f6\u5012\u983b\u8b5c\u4fc2\u6578(c0)\uff0c14 \u7dad\u5230 26 \u7dad\u70ba\u524d 13 \u7dad\u7684\u4e00\u968e\u5dee\u91cf\u4fc2\u6578
" } } } }