| { | |
| "paper_id": "O17-1020", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T07:59:26.105684Z" | |
| }, | |
| "title": "", | |
| "authors": [], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "", | |
| "pdf_parse": { | |
| "paper_id": "O17-1020", | |
| "_pdf_hash": "", | |
| "abstract": [], | |
| "body_text": [], | |
| "back_matter": [], | |
| "bib_entries": { | |
| "BIBREF0": { | |
| "ref_id": "b0", | |
| "title": "Mandarin Broadcast News Transcription", | |
| "authors": [], | |
| "year": 2005, | |
| "venue": "International Journal of Computational Linguistics & Chinese Language Processing", | |
| "volume": "10", | |
| "issue": "1", | |
| "pages": "1--18", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Mandarin Broadcast News Transcription,\" International Journal of Computational Linguistics & Chinese Language Processing, vol. 10, no. 1, 2005, pp. 1-18.", | |
| "links": null | |
| }, | |
| "BIBREF1": { | |
| "ref_id": "b1", | |
| "title": "Confidence measures for speech recognition: A survey", | |
| "authors": [ | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Jiang", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2005, | |
| "venue": "Speech Communication", | |
| "volume": "45", | |
| "issue": "4", | |
| "pages": "455--470", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "H. Jiang, \"Confidence measures for speech recognition: A survey,\" Speech Communication, vol. 45, no. 4, pp. 455 -470, 2005.", | |
| "links": null | |
| }, | |
| "BIBREF3": { | |
| "ref_id": "b3", | |
| "title": "The Kaldi Speech Recognition Toolkit", | |
| "authors": [ | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Motlicek", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "P", | |
| "middle": [], | |
| "last": "Qian", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Schwarz", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Silovsky", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Stemmer", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Vesely", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2011, | |
| "venue": "IEEE 2011 Workshop on Automatic Speech Recognition and Understanding", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Motlicek, YM Qian, P. Schwarz, J. Silovsky, G. Stemmer, K. Vesely, \"The Kaldi Speech Recognition Toolkit,\" in IEEE 2011 Workshop on Automatic Speech Recognition and Understanding, Big Island, Hawaii, 2011.", | |
| "links": null | |
| }, | |
| "BIBREF4": { | |
| "ref_id": "b4", | |
| "title": "Generating exact lattices in the WFST \u00b4 framework", | |
| "authors": [ | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Povey", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Hannemann", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Boulianne", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "L", | |
| "middle": [], | |
| "last": "Burget", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Ghoshal", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Janda", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Karafiat", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Kombrink", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "\u00b4", | |
| "middle": [ | |
| "P" | |
| ], | |
| "last": "Motl\u00b4\u0131cek", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Qian", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "N", | |
| "middle": [ | |
| "T" | |
| ], | |
| "last": "Vu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Riedhammer", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "\u02c7", | |
| "middle": [ | |
| "K" | |
| ], | |
| "last": "Vesely", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2011, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "D. Povey, M. Hannemann, G. Boulianne, L. Burget, A. Ghoshal, M. Janda, M. Karafiat, S. Kombrink, \u00b4 P. Motl\u00b4\u0131cek, Y. Qian, N. T. Vu, K. Riedhammer, and \u02c7 K. Vesely, \"Generating exact lattices in the WFST \u00b4 framework,\" 2011, submitted to ICASSP 2012.", | |
| "links": null | |
| }, | |
| "BIBREF6": { | |
| "ref_id": "b6", | |
| "title": "\u010cernock\u00fd Jan: RNNLM -Recurrent Neural Network Language Modeling Toolkit", | |
| "authors": [ | |
| { | |
| "first": "Mikolov", | |
| "middle": [], | |
| "last": "Tom\u00e1\u0161", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Kombrink", | |
| "middle": [], | |
| "last": "Stefan", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Deoras", | |
| "middle": [], | |
| "last": "Anoop", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Burget", | |
| "middle": [], | |
| "last": "Luk\u00e1\u0161", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2011, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Mikolov Tom\u00e1\u0161, Kombrink Stefan, Deoras Anoop, Burget Luk\u00e1\u0161, \u010cernock\u00fd Jan: RNNLM -Recurrent Neural Network Language Modeling Toolkit, In: ASRU 2011", | |
| "links": null | |
| }, | |
| "BIBREF7": { | |
| "ref_id": "b7", | |
| "title": "Demo Session", | |
| "authors": [], | |
| "year": null, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Demo Session.", | |
| "links": null | |
| }, | |
| "BIBREF8": { | |
| "ref_id": "b8", | |
| "title": "Quality Estimation for Automatic Speech Recognition", | |
| "authors": [ | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Negri", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Turchi", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Falavigna", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "C", | |
| "middle": [], | |
| "last": "Souza", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [ | |
| "G" | |
| ], | |
| "last": "", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2014, | |
| "venue": "Proc. of COLING", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "1813--1823", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Negri, M., Turchi, M., Falavigna, D., C. de Souza, J.G., 2014. Quality Estimation for Automatic Speech Recognition, in: Proc. of COLING, Dublin, Ireland. pp. 1813-1823.", | |
| "links": null | |
| }, | |
| "BIBREF9": { | |
| "ref_id": "b9", | |
| "title": "A tutorial on support vector regression", | |
| "authors": [ | |
| { | |
| "first": "Alex", | |
| "middle": [ | |
| "J" | |
| ], | |
| "last": "", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Bernhard", | |
| "middle": [], | |
| "last": "Sch\u00f6lkopf", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2004, | |
| "venue": "Statistics and computing", | |
| "volume": "14", | |
| "issue": "", | |
| "pages": "199--222", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "0Smola, Alex J., and Bernhard Sch\u00f6lkopf. \"A tutorial on support vector regression.\" Statistics and computing 14.3 (2004): 199-222.", | |
| "links": null | |
| }, | |
| "BIBREF10": { | |
| "ref_id": "b10", | |
| "title": "Extremely randomized trees", | |
| "authors": [ | |
| { | |
| "first": "P", | |
| "middle": [], | |
| "last": "Geurts", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Ernst", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "L", | |
| "middle": [], | |
| "last": "Wehenkel", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2006, | |
| "venue": "Machine Learning", | |
| "volume": "63", | |
| "issue": "", | |
| "pages": "3--42", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Geurts, P., Ernst, D., & Wehenkel, L. (2006). Extremely randomized trees.Machine Learning, 63(1), 3-42.", | |
| "links": null | |
| }, | |
| "BIBREF11": { | |
| "ref_id": "b11", | |
| "title": "Translation Quality Estimation using Recurrent Neural Network", | |
| "authors": [ | |
| { | |
| "first": "Raj", | |
| "middle": [], | |
| "last": "Nath Patel", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Sasikumar", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2016, | |
| "venue": "Proceedings of the First Conference on Machine Translation", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "819--824", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Raj Nath Patel, Sasikumar M. 2016. Translation Quality Estimation using Recurrent Neural Network. In Proceedings of the First Conference on Machine Translation, pages 819-824, Berlin, Germany.", | |
| "links": null | |
| } | |
| }, | |
| "ref_entries": { | |
| "TABREF0": { | |
| "num": null, | |
| "content": "<table><tr><td>1. \u7c21\u4ecb \u884c\u9010\u5b57\u7a3f\u8f49\u5beb\u3002\u63a5\u8457\u628a\u81ea\u52d5\u7522\u751f\u7684\u9010\u5b57\u7a3f\u52a0\u5165\u8a13\u7df4\u8a9e\u6599\u5eab\uff0c\u91cd\u65b0\u8a13\u7df4\u65b0\u7684\u8a9e\u97f3\u8fa8\u8a8d\u6a21\u578b\uff0c 2. \u57fa\u65bc\u8a9e\u97f3\u54c1\u8cea\u4f30\u7b97\u4e4b\u534a\u76e3\u7763\u5f0f\u8a9e\u97f3\u8fa8\u8a8d\u6a21\u578b \u7d93\u7db2\u8def\u8a9e\u8a00\u6a21\u578b(Recurrent Neural Network Language Model, RNNLM) [7] \uff0c\u5728\u672c\u8ad6\u6587\u4e2d \u5176\u4e2d\u6df1\u5c64\u985e\u795e\u7d93\u7db2\u8def\u7684\u67b6\u69cb\u5305\u542b\u591a\u500b\u7bc0\u9ede\u6216\u795e\u7d93\u5143\u7684\u591a\u5c64\u6b21\uff0c\u67b6\u69cb\u5982\u5716 7 \u6240\u793a\uff0c\u5176 \u8868 2 \u7a2e\u5b50\u7cfb\u7d71\u8a13\u7df4\u8a9e\u6599 3.2. \u4e2d\u82f1\u593e\u96dc\u8a9e\u97f3\u8fa8\u8a8d\u5668 \u8868 6 \u7a2e\u5b50\u7cfb\u7d71\u8a9e\u8a00\u6a21\u578b\u8a13\u7df4\u6587\u672c trees, Extra-Tree)[10]\u8207(3)DNN\uff0c\u5206\u5225\u8a13\u7df4\u4e09\u7a2e QE \u932f\u8aa4\u7387\u9810\u6e2c\u6a21\u578b\uff0c\u9032\u884c\u932f\u8aa4\u7387 \u800c\u82e5\u53ea\u91dd\u5c0d\u6559\u80b2\u96fb\u81fa\u6e2c\u8a66\u8a9e\u6599\u4f86\u770b\u534a\u76e3\u7763\u5f0f\u5b78\u7fd2\u7684\u6548\u80fd\uff0c\u5be6\u9a57\u7d50\u679c\u986f\u793a\u56db\u7a2e\u65b9\u6cd5\u7684 QE Select data \u8868 15 \u52a0\u5165\u5b8c\u6574 Giga Word \u8a9e\u6599\u5f8c\u4e4b\u5ee3\u64ad\u8a9e\u6599\u8fa8\u8a8d\u932f\u8aa4\u76f8\u5c0d\u6539\u5584\u7387</td></tr><tr><td>\u5ee3\u64ad\u7bc0\u76ee\u7684\u8a9e\u97f3\u8cc7\u6599\u6e90\u6e90\u4e0d\u7d55\uff0c\u4f46\u56e0\u4eba\u529b\u3001\u8cc7\u6e90\u7b49\u56e0\u7d20\uff0c\u5ee3\u64ad\u7bc0\u76ee\u88fd\u4f5c\u5b8c\u6210\u5f8c\uff0c\u901a \u5e38\u53ea\u6709\u4fdd\u7559\u6700\u5f8c\u8981\u64ad\u51fa\u7684\u8a9e\u97f3\u8a0a\u865f\uff0c\u6c92\u6709\u5c07\u9304\u88fd\u904e\u7a0b\u4e2d\u7684\u7528\u5230\u7684\u76f8\u95dc\u8cc7\u6599\uff0c\u6574\u7406\u6210\u5f8c\u8a2d \u8cc7\u6599(metadata)\u3002 \u5c0e\u81f4\u7bc0\u76ee\u64ad\u51fa\u5f8c\uff0c\u5f88\u96e3\u6aa2\u518d\u6aa2\u7d22\u7bc0\u76ee\u5167\u5bb9\uff0c\u6216\u662f\u52a0\u4ee5\u7d44\u7e54\u518d\u5229\u7528\u3002\u56e0 \u6b64\u6211\u5011\u5e0c\u671b\u80fd\u5920\u8f49\u5beb\u5ee3\u64ad\u7bc0\u76ee\u7522\u751f\u8a9e\u97f3\u9010\u5b57\u7a3f\uff0c\u4ee5\u4fbf\u5c07\u5ee3\u64ad\u7bc0\u76ee\u7d44\u7e54\u6210\u6709\u8072\u66f8\uff0c\u8b93\u9019\u4e9b \u5927\u91cf\u7684\u8a9e\u97f3\u8cc7\u6599\u53ef\u4ee5\u6709\u66f4\u591a\u7684\u52a0\u503c\u904b\u7528\u3002\u9664\u4e86\u53ef\u4ee5\u8b93\u807d\u773e\u80fd\u5920\u5bb9\u6613\u5730\u4ee5\u6587\u5b57\u6aa2\u7d22\u7684\u65b9\u5f0f\uff0c \u53bb\u627e\u5230\u6700\u95dc\u9375\u7684\u8b1b\u8ff0\u5167\u5bb9\u90e8\u5206\uff0c\u5c24\u5176\u662f\u540d\u4eba\u5728\u7bc0\u76ee\u4e2d\u6240\u8aaa\u7684\u6545\u4e8b\u3001\u60f3\u6cd5\u601d\u7dad\u3001\u65b0\u77e5\u7b49\u7b49\uff0c \u4e5f\u53ef\u4ee5\u5229\u7528\u9010\u5b57\u7a3f\uff0c\u5c07\u5ee3\u64ad\u7bc0\u76ee\u7684\u8a9e\u97f3\u8f49\u6210\u5b57\u5e55\u6a94\uff0c\u8b8a\u6210\u591a\u5a92\u9ad4\u8996\u8a0a\u6a94\u6848\uff0c\u8b93\u807e\u80de\u4e5f\u53ef \u4ee5\u5f9e\u4e2d\u7372\u77e5\u5ee3\u64ad\u7bc0\u76ee\u5167\u5bb9\uff0c\u6216\u662f\u7576\u505a\u7b2c\u4e8c\u8a9e\u8a00\u5b78\u7fd2\u7528\u7684\u8a9e\u97f3\u7bc4\u4f8b\u3002 \u8981\u80fd\u9054\u5230\u5c07\u5ee3\u64ad\u7bc0\u76ee\u81ea\u52d5\u8f49\u5beb\u6210\u8a9e\u97f3\u9010\u5b57\u7a3f\u9019\u500b\u76ee\u7684\uff0c\u901a\u5e38\u9700\u8981\u5148\u64c1\u6709\u4e00\u500b\u9069\u5408\u8fa8 \u8a8d\u5ee3\u64ad\u7bc0\u76ee\u8a9e\u97f3\u7684\u5927\u8a5e\u5f59\u8a9e\u97f3\u8fa8\u8a8d(Large Vocabulary Continuous Speech Recognition, LVCSR)\u5668\uff0c\u4f46\u662f\u56e0\u70ba\u5ee3\u64ad\u7bc0\u76ee\u9304\u88fd\u6642\uff0c\u901a\u5e38\u4e0d\u6703\u5148\u7d66\u4e3b\u6301\u4eba\u8207\u4f86\u8cd3\u8b1b\u7a3f\uff0c\u5c24\u5176\u662f\u5c0d\u8a71 \u6027\u8cea\u7684\u7bc0\u76ee\uff0c\u554f\u7b54\u4e4b\u9593\u5e38\u8b93\u4f86\u8cd3\u81ea\u7531\u767c\u63ee\uff0c\u56e0\u6b64\u5ee3\u64ad\u7bc0\u76ee\u4e2d\u7684\u8a9e\u97f3\u901a\u5e38\u70ba\u8f03\u96a8\u8208\u7684\u53e3\u8a9e\uff0c \u5177\u6709\u5f37\u70c8\u7684\u81ea\u767c\u6027\u8a9e\u97f3(spontaneous speech)\u7279\u6027\u3002 \u4f46\u662f\uff0c\u76ee\u524d\u975e\u5e38\u7f3a\u4e4f\u9ad8\u6548\u80fd\u7684\u81ea\u767c\u6027\u8a9e\u97f3\u8fa8\u8a8d\u5668\u3002\u9019\u662f\u56e0\u70ba\u82e5\u8981\u63d0\u9ad8\u81ea\u767c\u6027\u8a9e\u97f3\u8fa8 \u8a8d\u5668\u7684\u6548\u80fd\uff0c\u9700\u8981\u76f4\u63a5\u4ee5\u5927\u91cf\u7684\u81ea\u767c\u6027\u8a9e\u97f3\u8a9e\u6599\u8207\u53e3\u8a9e\u6587\u5b57\u8a9e\u6599\u4f86\u8a13\u7df4\u8fa8\u8a8d\u5668\u4e2d\u7684\u8072\u5b78 \u6a21\u578b\u8207\u8a9e\u8a00\u6a21\u578b\u3002\u4f46\u9019\u5169\u7a2e\u8a9e\u6599\uff0c\u901a\u5e38\u5f88\u96e3\u53d6\u5f97\u3002\u5c24\u5176\u662f\u6a19\u8a3b\u597d\u9010\u5b57\u7a3f\u7684\u81ea\u767c\u6027\u8a9e\u97f3\u8a9e \u6599\u5eab\uff0c\u56e0\u5176\u9700\u8017\u8cbb\u5927\u91cf\u4eba\u5de5\u3001\u6642\u9593\u3001\u91d1\u9322\u6210\u672c\u624d\u80fd\u5b8c\u6210\uff0c\u56e0\u6b64\u901a\u5e38\u6709\u516c\u958b\u767c\u884c\u7684\u81ea\u767c\u6027 \u8a9e\u97f3\u8a9e\u6599\u5eab\u90fd\u5f88\u5c0f\uff0c\u53ea\u9069\u5408\u9032\u884c\u8a9e\u8a00\u5b78\u5206\u6790\uff0c\u63a2\u8a0e\u8a9e\u8a00\u73fe\u8c61\u3002\u800c\u82e5\u8981\u5efa\u7acb\u57fa\u65bc\u6df1\u5ea6\u5b78\u7fd2 \u7684\u9ad8\u6548\u80fd\u8a9e\u97f3\u6a21\u578b\uff0c\u5c31\u9700\u8981\u5f88\u5927\u7684\u6578\u64da\u91cf\uff0c\u901a\u5e38\u9700\u8981\u6578\u767e\u5c0f\u6642\uff0c\u6216\u662f\u6578\u5343\u5c0f\u6642\u6a19\u8a3b\u597d\u7684 \u81ea\u767c\u6027\u8a9e\u6599\u624d\u80fd\u9054\u6210\u3002\u56e0\u6b64\u5982\u4f55\u7372\u5f97\u8db3\u5920\u7684\u6709\u6a19\u8a18\u81ea\u767c\u6027\u8a9e\u97f3\u8a9e\u6599\uff0c\u662f\u76ee\u524d\u6025\u9700\u89e3\u6c7a\u7684 \u4e00\u5927\u554f\u984c\u3002 \u91dd\u5c0d\u6b64\u554f\u984c\uff0c\u4e00\u822c\u7684\u505a\u6cd5\u662f\u4ee5\u534a\u76e3\u7763\u5f0f\u5b78\u7fd2[1][2]\u65b9\u5f0f\u89e3\u6c7a\uff0c\u4f8b\u5982\u5716 1 \u6240\u793a\u7684\u67b6\u69cb\u3002 \u65b9\u6cd5\u662f\u5148\u5229\u7528\u8f03\u6613\u53d6\u5f97\u3001\u6709\u6b63\u78ba\u6a19\u8a3b\u7684\u8b80\u7a3f\u8a9e\u6599(reading speech) \uff0c\u5efa\u7acb\u4e00\u7a2e\u5b50\u8a9e\u97f3\u8fa8 \u8a8d\u5668\uff0c\u518d\u7528\u6b64\u7a2e\u5b50\u8a9e\u97f3\u8fa8\u8a8d\u5668\uff0c\u81ea\u52d5\u5c0d\u5927\u91cf\u672a\u7d93\u4eba\u5de5\u6a19\u8a3b\u7684\u5ee3\u64ad\u96fb\u81fa\u7bc0\u76ee\u8a9e\u97f3\u8a9e\u6599\uff0c\u9032 \u4ee5\u6539\u5584\u8fa8\u8a8d\u81ea\u767c\u6027\u8a9e\u97f3\u7684\u6548\u80fd\u3002 \u4ee5\u4e0b\u8aaa\u660e\u672c\u8ad6\u6587\u63d0\u51fa\u7684\u534a\u76e3\u7763\u5f0f\u8a9e\u97f3\u8fa8\u8a8d\u5668\u4e2d\u7684\u5404\u6a21\u7d44\uff0c\u5305\u62ec(1)\u7a2e\u5b50\u8a9e\u97f3\u8fa8\u8a8d \u6211\u5011\u5373\u4f7f\u7528 RNNLM \u53bb\u589e\u52a0\u8a9e\u8a00\u6a21\u578b\u7684\u6574\u9ad4\u6548\u80fd\uff0c\u4ee5\u6539\u5584\u8fa8\u8a8d\u7cfb\u7d71\u7684\u8fa8\u8a8d\u7387\u3002 2.2. \u534a\u76e3\u7763\u5f0f\u5b78\u7fd2\u8a13\u7df4 \u96b1\u85cf\u5c64\u9593\u7684\u795e\u7d93\u5143\u4e92\u4e0d\u9023\u7d50\uff0c\u6bcf\u500b\u795e\u7d93\u5143\u4f7f\u7528 Relu \u6fc0\u6d3b\u51fd\u6578\uff0c\u7528\u4f86\u89e3\u6c7a\u66f4\u65b0\u6b0a\u91cd\u503c\u6642 \u7684\u68af\u5ea6\u6d88\u5931\u554f\u984c\uff0c\u800c\u8a13\u7df4\u6642\u4f7f\u7528\u7684\u6210\u672c\u51fd\u6578\uff0c\u70ba\u4f9d\u64da\u4eba\u5de5\u6a19\u8a18\u7b97\u51fa\u7684\u771f\u6b63\u932f\u8aa4\u7387\uff0c\u8207 QE \u8a13\u7df4\u8a9e\u6599 \u6642\u6578 \u8a9e\u8005\u6578 \u8a9e\u53e5\u6578 TCC300 (all) 26.4 300 27,375 MATBN (train) 127.3 5,207 29,549 \u5728\u83ef\u8a9e\u570b\u5bb6\u7531\u65bc\u88ab\u570b\u969b\u8a9e\u8a00\u82f1\u6587\u7684\u5f71\u97ff\u72c0\u6cc1\u4e0b\u6642\u5e38\u5728\u8b1b\u8a71\u6642\u6703\u7a7f\u63d2\u4e00\u4e9b\u82f1\u6587\u5b57\u8a5e\uff0c \u800c\u55ae\u8a9e\u8a00\u8a9e\u97f3\u8fa8\u8a8d\u5668\u986f\u7136\u7121\u6cd5\u6b63\u5e38\u8fa8\u8b58\u591a\u8a9e\u8a00\u593e\u96dc\u7684\u8aaa\u8a71\u8a9e\u6d41\uff0c\u6240\u4ee5\u8a13\u7df4\u4e00\u500b\u591a\u8a9e\u8a00\u8a9e \u8a13\u7df4\u6587\u672c \u8a9e\u53e5\u6578 \u5b57\u8a5e\u6578 TCC300 (all) 27,375 186,369 MATBN (train) 29,549 1,264,625 \u9810\u6e2c\u6548\u80fd\u6bd4\u8f03\u3002\u5176\u4e2d\uff0cDNN \u7684\u8a13\u7df4\u53c3\u6578\u8a2d\u5b9a\u70ba learning rate= 0.001\u3001epochs=100\u3001 batch size=500\u3001dropout=0.1\uff0cDNN \u7684\u5c64\u6578\u5247\u5617\u8a66\u4f7f\u7528 1~3 \u5c64\u96b1\u85cf\u5c64\u3002 \u76f8\u5c0d\u8fa8\u8a8d\u6539\u5584\u7387\u5982\u8868 12 \u6240\u793a\u3002\u5176\u4e2d\u9304\u97f3\u54c1\u8cea\u8f03\u5dee\u7684 NER-set1 (\u6280\u8077\u6700\u524d\u7dda)\u7684\u6700\u4f73\u76f8 24 26 600 700 CER in % \u76f8\u5c0d\u6539\u5584\u7387 \u8fa8\u8a8d\u6a21\u578b NER-set1 NER-set2 NER-set1 NER-set2 \u5c0d\u6539\u5584\u7387\u4f86\u5230 4.48%(QE2 \u8207 CM2) \uff0c\u9304\u97f3\u54c1\u8cea\u8f03\u597d\u7684 NER-set2 (\u570b\u969b\u6559\u80b2\u5fc3\u52d5\u7dda) \u7684 22 500 \u7a2e\u5b50\u6a21\u578b 25.00 14.24 --\u6a21\u578b\u8207(2)\u534a\u76e3\u7763\u5f0f\u5b78\u7fd2\u8a13\u7df4\u7684\u4f5c\u6cd5\u3002 2.1. \u7a2e\u5b50\u8a9e\u97f3\u8fa8\u8a8d\u6a21\u578b \u672c\u8ad6\u6587\u5229\u7528 Kaldi speech recognition toolkit[4]\u74b0\u5883\uff0c\u5efa\u7acb\u7a2e\u5b50\u8a9e\u97f3\u8fa8\u8a8d\u7cfb\u7d71\uff0c\u5305\u62ec \u5ee3\u64ad\u96fb\u81fa\u7684\u8a9e\u97f3\u8cc7\u6599\u975e\u5e38\u9f90\u5927\uff0c\u4f46\u672a\u6a19\u8a18\u7684\u8a9e\u6599\u7121\u6cd5\u62ff\u53bb\u505a\u8072\u5b78\u6a21\u578b\u8a13\u7df4\u3002\u56e0\u6b64\u6211 \u5011\u63d0\u51fa\u4e00\u65b0\u7684\u534a\u76e3\u7763\u5f0f\u5b78\u7fd2\u65b9\u6cd5\uff0c\u5176\u5305\u542b\u8a13\u7df4 QE \u6a21\u578b\u8207\u6311\u9078\u8a9e\u6599\u5169\u90e8\u5206\uff0c\u6574\u9ad4\u67b6\u69cb\u5982 \u9810\u6e2c\u7684\u932f\u8aa4\u7387\u9593\u7684\u5747\u65b9\u5dee(Mean Squared Error, MSE)\u3002 OC16-CE80 (train) 63.8 1,163 58,132 SEAME 95.1 138 94,034 Librispeech (train-clean100hr) 100.6 251 28,539 \u97f3\u8fa8\u8a8d\u5668\u6703\u662f\u4e00\u500b\u7b26\u5408\u73fe\u4ee3\u4eba\u5728\u81ea\u7136\u8aaa\u8a71\u8a9e\u6d41\u7684\u8aaa\u8a71\u5f62\u5f0f\uff0c\u56e0\u6b64\u6211\u5011\u5c07\u4e2d\u3001\u82f1\u6587\u5229\u7528 X-SAMPA \u97f3\u7d20\u7de8\u78bc\u898f\u5247\u7522\u751f\u97f3\u7d20\uff0c\u4e26\u4e14\u4f7f\u7528\u97f3\u7d20\u5171\u4eab\u8207\u4e2d\u82f1\u6df7\u5408\u6a21\u578b\u4f86\u5b8c\u6210\u6211\u5011\u7684\u4e2d\u82f1 OC16-CE80 (train) 58,132 509,657 SEAME 94,034 1,200,121 Librispeech (train-960) 28,539 9,403,555 \u56e0\u8a13\u7df4\u8a9e\u6599\u8f03\u5c11\uff0c\u70ba\u516c\u5e73\u6bd4\u8f03\u4e09\u7a2e\u65b9\u5f0f\uff0c\u6211\u5011\u5229\u7528 Cross-validation \u8a13\u7df4\u8207\u6e2c\u8a66\u65b9 \u5f0f\uff0c\u5728\u6bcf\u4e00\u6b21\u7684\u8a13\u7df4\u5c07\u5176\u4e2d\u4e00\u500b\u7bc0\u76ee\u7576\u4f5c\u6e2c\u8a66\u8cc7\u6599\uff0c\u5176\u9918\u4e03\u500b\u7bc0\u76ee\u7576\u4f5c\u8a13\u7df4\u8cc7\u6599\uff0c\u7e3d\u5171 \u7df4\u8a9e\u6599\uff0c\u5982\u8868 10\u3002 \u8868 10 \u534a\u76e3\u7763\u5f0f\u5b78\u7fd2\u8a13\u7df4\u8a9e\u6599\u6311\u9078 \u6700\u4f73\u76f8\u5c0d\u6539\u5584\u7387\u4f86\u5230 7.94%(QE1)\u3002 16 18 20 300 400 CER (%) \u7a2e\u5b50\u6a21\u578b+LM2 24.54 13.27 1.84% 6.81% Hours QE1(598h) 23.61 13.24 5.56% 7.02% \u8868 12 \u5ee3\u64ad\u8a9e\u6599\u6e2c\u8a66\u7d50\u679c_\u76f8\u5c0d\u6539\u5584\u7387 14 200 QE1(598h)+LM2 23.25 12.63 7% 11.3% \u5716 1 \u5229\u7528\u672a\u6a19\u8a18\u8a9e\u6599\u8a13\u7df4\u8a9e\u97f3\u8fa8\u8a8d\u5668\u67b6\u69cb \u5176\u4e2d\uff0c\u56e0\u70ba\u81ea\u52d5\u8f49\u5beb\u51fa\u7684\u9010\u5b57\u7a3f\uff0c\u901a\u5e38\u6703\u6709\u932f\u8aa4\uff0c\u7121\u6cd5\u5b8c\u5168\u4fe1\u4efb\u3002\u56e0\u6b64\u50b3\u7d71\u4e0a\u6703\u518d \u5229\u7528\u5982\u5716 2 \u7684\u67b6\u69cb\uff0c\u589e\u52a0\u4e00\u500b\u4fe1\u5fc3\u503c\u4f30\u7b97(Confidence Measure)[3]\uff0c\u8a08\u7b97\u6bcf\u6bb5\u9010\u5b57\u7a3f \u7684\u8fa8\u8a8d\u4fe1\u5fc3\u503c\uff0c\u53ea\u6311\u9078\u8f03\u53ef\u9760\u7684\u8f49\u5beb\u7d50\u679c\uff0c\u52a0\u5165\u8a13\u7df4\u8a9e\u6599\u5eab\u3002 \u5716 2 \u50b3\u7d71\u534a\u76e3\u7763\u5f0f\u5b78\u7fd2\u8a9e\u97f3\u8fa8\u8a8d\u5668\u8a13\u7df4\u67b6\u69cb \u9010\u5b57\u7a3f\u7684\u4fe1\u5fc3\u503c\u4f30\u7b97\uff0c\u901a\u5e38\u662f\u4f9d\u8cf4\u7a2e\u5b50\u8fa8\u8a8d\u5668\u7684\u89e3\u78bc\u8f38\u51fa\u3002\u7136\u800c\u56e0\u70ba\u7a2e\u5b50\u8fa8\u8a8d\u5668\u4e00 \u822c\u662f\u7528\u8b80\u7a3f\u8a9e\u6599\u5efa\u7acb\uff0c\u8207\u81ea\u767c\u6027\u8a9e\u97f3\u6703\u6709\u8aaa\u8a71\u6a21\u5f0f(speaking style)\u4e0d\u5339\u914d\u7684\u554f\u984c\uff0c\u56e0 \u6b64\u7b97\u51fa\u4f86\u7684\u4fe1\u5fc3\u503c\u4f30\u7b97\u4e0d\u898b\u5f97\u53ef\u9760\u3002\u6240\u4ee5\u5728\u672c\u8ad6\u6587\u4e2d\uff0c\u5c07\u6539\u4ee5\u540c\u6642\u5229\u7528\u591a\u7a2e\u8a9e\u97f3\u54c1\u8cea\u7dda \u7d22\uff0c\u5305\u62ec\u8a9e\u97f3\u8a0a\u865f\u672c\u8eab\u7684\u7279\u5fb5\u53c3\u6578\uff0c\u9010\u5b57\u7a3f\u6587\u5b57\u5167\u5bb9\u7279\u5fb5\u53c3\u6578\uff0c\u8207\u591a\u7a2e\u8a0a\u865f\u8207\u6587\u5b57\u5167\u5bb9 \u6df7\u5408\u53c3\u6578\uff0c\u5efa\u7acb\u4e00\u8a9e\u97f3\u54c1\u8cea\u4f30\u7b97\u5668(Quality Estimation\uff0cQE) \uff0c\u76f4\u63a5\u9810\u6e2c\u672a\u6a19\u8a18\u8a9e\u6599\u81ea \u52d5\u8f49\u5beb\u9010\u5b57\u7a3f\u7684\u8fa8\u8b58\u5b57\u5143\u932f\u8aa4\u7387(CER) \uff0c\u4e26\u4e14\u53ea\u6311\u9078\u8fa8\u8a8d\u932f\u8aa4\u7387\u8f03\u4f4e\u7684\u8f49\u5beb\u7d50\u679c\uff0c\u52a0 \u5165\u8a13\u7df4\u8a9e\u6599\u5eab\u3002\u6211\u5011\u6240\u63d0\u51fa\u7684\u67b6\u69cb\u5982\u5716 3 \u6240\u793a\u3002 \u5716 3 \u4ee5\u9810\u6e2c\u932f\u8aa4\u7387\u6311\u9078\u534a\u76e3\u7763\u5f0f\u5b78\u7fd2\u8a13\u7df4\u8a9e\u6599\u67b6\u69cb \u8a9e\u97f3\u7279\u5fb5\u53c3\u6578\u64f7\u53d6\u3001\u8072\u5b78\u6a21\u578b\u8a13\u7df4\u3001\u8a9e\u8a00\u6a21\u578b\u8a13\u7df4\u3002\u6700\u5f8c\u4ee5\u52a0\u6b0a\u6709\u9650\u72c0\u614b\u8f49\u63db\u6a5f[5]\u7d50\u5408 \u8a9e\u8a00\u6a21\u578b\u8207\u8a5e\u5178\uff0c\u5c0d\u6559\u80b2\u96fb\u81fa\u5ee3\u64ad\u7bc0\u76ee\u97f3\u6a94\u9032\u884c\u8f49\u5beb\u9010\u5b57\u7a3f\uff0c\u6574\u9ad4\u67b6\u69cb\u5982\u5716 4 \u6240\u793a\u3002\u4ee5 \u4e0b\u8a73\u7d30\u8aaa\u660e\u5404\u6a21\u7d44\u3002 \u5716 4 \u7a2e\u5b50\u8a9e\u97f3\u8fa8\u8a8d\u7cfb\u7d71\u67b6\u69cb\u5716 2.1.1. \u8072\u5b78\u6a21\u578b LSTM \u7db2\u8def\u662f\u4e00\u7a2e\u7279\u6b8a\u7684 RNN \u7d50\u69cb[6]\uff0c\u53ef\u4ee5\u8a18\u61b6\u8f03\u9577\u7684\u6642\u9593\u8cc7\u8a0a\u3002\u5176\u4e2d\u6240\u6709\u6709\u95dc \u8a0a\u606f\u50b3\u905e\u7684\u904b\u4f5c\u90fd\u6c7a\u5b9a\u65bc\u9580(gates)\uff0c\u800c\u9019\u4e9b gates \u4f9d\u64da\u63a5\u6536\u5230\u7684\u4fe1\u865f\uff0c\u8a08\u7b97\u6fc0\u767c\u5f37\u5ea6\u4f86 \u6c7a\u5b9a\u8a0a\u606f\u662f\u5426\u901a\u904e\u6216\u662f\u88ab\u79fb\u9664\u3002LSTM \u7684\u7d50\u69cb\u80fd\u5920\u7528\u4f86\u9632\u6b62\u9577\u8ddd\u96e2\u4f9d\u8cf4\u554f\u984c\uff0c\u4e5f\u5c31\u662f\u53ef \u4ee5\u89e3\u6c7a\u68af\u5ea6\u6d88\u5931\u7684\u554f\u984c\uff0c\u5728\u6b64\u8ad6\u6587\u4e2d\uff0c\u6211\u5011\u4f7f\u7528\u6211\u5011\u6240\u64c1\u6709\u7684\u591a\u500b\u8a9e\u6599\u5eab\uff0c\u5305\u542b\u4e2d\u6587\u3001 \u82f1\u6587\u3001\u4e2d\u82f1\u593e\u96dc\u8a9e\u6599\uff0c\u5171\u7d04 400 \u5c0f\u6642\uff0c\u4f86\u8a13\u7df4\u905e\u8ff4\u5f0f\u985e\u795e\u7d93\u7db2\u8def\u8072\u5b78\u6a21\u578b\u3002 2.1.2. \u8a9e\u8a00\u6a21\u578b \u8a9e\u8a00\u6a21\u578b\u6700\u4e3b\u8981\u7684\u76ee\u6a19\u70ba\u4f7f\u7528\u8a5e\u5e8f\u5217\u4e2d\u5148\u524d\u51fa\u73fe\u7684\u8a5e\u4f86\u9810\u6e2c\u73fe\u5728\u6700\u6709\u53ef\u80fd\u7528\u5230\u7684 \u8a5e\u3002\u8f03\u5e38\u88ab\u4f7f\u7528\u7684\u8a9e\u8a00\u6a21\u578b\u70ba n \u9023\u8a9e\u8a00\u6a21\u578b(n-gram language model)\uff0c\u5176\u7d71\u8a08\u65b9\u5f0f\u70ba\u8a08 \u7b97\u8a5e\u8207\u8a5e\u4e4b\u9593\u9023\u63a5\u7684\u53ef\u80fd\u6027\u4ee5\u6311\u9078\u53ef\u80fd\u7684\u5b57\u8a5e\u3002\u76ee\u524d\u9032\u968e\u7684\u4f5c\u6cd5\u5247\u662f\uff0c\u4f7f\u7528\u905e\u8ff4\u5f0f\u985e\u795e \u5716 5 \u8207\u5716 6 \u6240\u793a\u3002 \u4e3b\u8981\u505a\u6cd5\u662f\u5148\u5229\u7528\u7a2e\u5b50\u8a9e\u97f3\u8fa8\u8a8d\u7cfb\u7d71\u81ea\u52d5\u8f49\u5beb\u672a\u6a19\u8a18\u7684\u5ee3\u64ad\u8a9e\u6599\uff0c\u518d\u4f7f\u7528\u8a13\u7df4\u597d\u7684 QE \u932f\u8aa4\u7387\u6a21\u578b\uff0c\u9810\u6e2c\u5176\u8fa8\u8a8d\u932f\u8aa4\u7387\uff0c\u6311\u9078\u932f\u8aa4\u7387\u8f03\u4f4e\u7684\u8a9e\u6599\u7576\u4f5c\u534a\u76e3\u7763\u5f0f\u5b78\u7fd2\u8a13\u7df4\u8a9e \u6599\uff0c\u52a0\u5165\u7a2e\u5b50\u6a21\u578b\u8a13\u7df4\u8a9e\u6599\u91cd\u65b0\u8a13\u7df4\u8072\u5b78\u6a21\u578b\u3002\u6211\u5011\u7684\u534a\u76e3\u7763\u5f0f\u5b78\u7fd2\u8207\u50b3\u7d71\u65b9\u6cd5\u6700\u5927\u4e0d \u540c\u7684\u5730\u65b9\uff0c\u5728\u65bc\u589e\u52a0\u4e00\u500b\u65b0\u7684 QE \u6a21\u578b\uff0c\u4ee5\u53d6\u4ee3\u50b3\u7d71\u7684 CM \u8a9e\u6599\u6311\u9078\u65b9\u6cd5\uff0c\u5176\u4e2d QE \u6a21 \u578b\u8a13\u7df4\uff0c\u662f\u4f9d\u64da\u5716 5 \u6240\u793a\u7684\u67b6\u69cb\uff0c\u57fa\u65bc\u76e3\u7763\u5f0f\u5b78\u7fd2\u8a13\u7df4\u800c\u6210\u3002\u4e00\u65b9\u9762\u5229\u7528\u7a2e\u5b50\u8a9e\u97f3\u8fa8\u8a8d \u7cfb\u7d71\u5c07\u73fe\u6709\u5df2\u6a19\u8a18\u7684\u8a9e\u6599\uff0c\u81ea\u52d5\u8f49\u5beb\u51fa\u9010\u5b57\u7a3f\uff0c\u4e26\u5229\u7528\u5df2\u6709\u7684\u4eba\u5de5\u6a19\u8a18\uff0c\u8a08\u7b97\u51fa\u9010\u5b57\u7a3f \u7684\u932f\u8aa4\u7387\u3002\u4e00\u65b9\u9762\u5f9e\u81ea\u52d5\u8f49\u5beb\u7684\u9010\u5b57\u7a3f\u64f7\u53d6\u6587\u5b57\u76f8\u95dc\u7279\u5fb5\u53c3\u6578\u3001\u8a0a\u865f\u76f8\u95dc\u53c3\u6578\uff0c\u4ee5\u53ca\u5229 \u7528\u7a2e\u5b50\u8a9e\u97f3\u8fa8\u8a8d\u7cfb\u7d71\u7522\u751f\u7684\u5207\u5272\u6642\u9593\uff0c\u8207\u5c0d\u61c9\u7684\u97f3\u6a94\u8a0a\u865f\uff0c\u63d0\u53d6\u6df7\u5408\u7279\u5fb5\u53c3\u6578\u3002\u518d\u4ee5\u5be6 \u969b\u932f\u8aa4\u7387\u70ba\u76ee\u6a19\uff0c\u8a13\u7df4\u51fa\u4e00\u500b QE \u932f\u8aa4\u7387\u9810\u6e2c\u6a21\u578b\u3002 \u5716 5 QE \u6a21\u578b\u8a13\u7df4\u67b6\u69cb \u5f85 QE \u6a21\u578b\u8a13\u7df4\u597d\u5f8c\uff0c\u5c31\u53ef\u4ee5\u4f9d\u5716 6 \u6240\u793a\u7684\u67b6\u69cb\uff0c\u5229\u7528 QE \u6a21\u578b\uff0c\u9810\u6e2c\u8a13\u7df4\u8a9e\u6599\u7684 \u8fa8\u8a8d\u932f\u8aa4\u7387\uff0c\u4ee5\u5f9e\u6e90\u6e90\u4e0d\u7d55\uff0c\u4f46\u672a\u6a19\u8a18\u7684\u5927\u91cf\u5ee3\u64ad\u8a9e\u6599\u4e2d\uff0c\u6311\u9078\u9069\u5408\u8a13\u7df4\u8a9e\u97f3\u8fa8\u8a8d\u5668\u7684 \u8a9e\u6599\uff0c\u9032\u884c\u534a\u76e3\u7763\u5f0f\u8072\u5b78\u6a21\u578b\u8a13\u7df4\u3002 \u5716 6 \u5229\u7528 QE \u540d\u884c\u6311\u9078\u534a\u76e3\u7763\u5f0f\u5b78\u7fd2\u8a13\u7df4\u8a9e\u6599\u67b6\u69cb 2.2.1. QE \u8a13\u7df4\u53c3\u6578\u64f7\u53d6 \u672c\u8ad6\u6587\u4e2d\uff0c\u6211\u5011\u5171\u5229\u7528 93 \u500b\u53c3\u6578\u4f86\u8a13\u7df4 QE \u6a21\u578b[8]\uff0c\u5982\u8868 1 \u6240\u793a\uff0c\u5305\u542b\u4e09\u985e\u7279\u5fb5 \u53c3\u6578\u3002\u5176\u4e2d 16 \u500b\u64f7\u53d6\u81ea\u8a9e\u97f3\u8a0a\u865f\u53c3\u6578\u300157 \u500b\u64f7\u53d6\u81ea\u8a9e\u8a00\u6a21\u578b\u53c3\u6578\u300120 \u500b\u64f7\u53d6\u81ea\u8a9e\u97f3\u8a0a \u865f\u8207\u8fa8\u8a8d\u7d50\u679c\u7684\u6df7\u5408\u53c3\u6578\u3002 \u8868 1 QE \u6a21\u578b\u8a13\u7df4\u53c3\u6578 Signal(16) Total segment duration (sec), Mean/Min/Max raw energy (dB), mean MFCC(12). 2.2.2. QE \u6a21\u578b\u8a13\u7df4 \u672c\u5be6\u9a57\u4e2d\u5229\u7528\u6df1\u5c64\u985e\u795e\u7d93\u7db2\u8def\u65b9\u6cd5\u4f86\u88fd\u505a QE \u6a21\u578b\u3002\u6df1\u5c64\u985e\u795e\u7d93\u7db2\u8def\u662f\u4e00\u7a2e\u5177\u5099\u81f3 \u5c11\u4e00\u500b\u96b1\u85cf\u5c64\u7684\u795e\u7d93\u7db2\u8def\u3002\u5176\u53ef\u4ee5\u900f\u904e\u96b1\u85cf\u5c64\u5c64\u6578\u7684\u589e\u52a0\uff0c\u63d0\u4f9b\u66f4\u8907\u96dc\u7684\u975e\u7dda\u6027\u8655\u7406\u80fd \u529b\uff0c\u56e0\u800c\u80fd\u63d0\u9ad8\u6a21\u578b\u7684\u80fd\u529b\u3002 \u5716 7 DNN-based QE \u67b6\u69cb 3. \u5be6\u9a57\u8a2d\u5b9a 3.1. \u8a9e\u6599\u5eab\u4ecb\u7d39 \u5728\u672c\u8ad6\u6587\u4e2d\uff0c\u6211\u5011\u5148\u4f7f\u7528\u591a\u500b\u5df2\u6709\u7684\u4e2d\u6587\u3001\u82f1\u6587\u53ca\u4e2d\u82f1\u593e\u96dc\u7684\u8a9e\u6599\u5eab\uff0c\u8a13\u7df4\u7a2e\u5b50\u8a9e \u97f3\u8fa8\u8a8d\u7cfb\u7d71\u3002\u518d\u5229\u7528\u6b64\u7a2e\u5b50\u8a9e\u97f3\u8fa8\u8a8d\u7cfb\u7d71\uff0c\u5c0d\u5927\u91cf\u672a\u6a19\u8a18\u7684\u5ee3\u64ad\u8a9e\u6599\uff0c\u9032\u884c\u81ea\u52d5\u8f49\u5beb\u9010 \u5b57\u7a3f\uff0c\u4e26\u5206\u5225\u5c0d\u81ea\u52d5\u8f49\u5beb\u51fa\u7684\u9010\u5b57\u7a3f\u505a\u9810\u6e2c\u932f\u8aa4\u7387\u8207\u4fe1\u5fc3\u5ea6\u8a55\u4f30 (Confidence Measure)\uff0c \u6311\u9078\u9069\u5408\u7684\u8a9e\u6599\uff0c\u52a0\u5165\u7a2e\u5b50\u7cfb\u7d71\u7684\u8a13\u7df4\u8a9e\u6599\u91cd\u65b0\u8a13\u7df4\u8072\u5b78\u6a21\u578b\uff0c\u4ee5\u6bd4\u8f03 QE \u65b9\u6cd5\u8207 CM \u65b9\u6cd5\u7684\u6548\u80fd\u3002 3.1.1. \u7a2e\u5b50\u8fa8\u8a8d\u5668\u8a13\u7df4\u8a9e\u6599\u53ca\u6e2c\u8a66\u8a9e\u6599 \u7a2e\u5b50\u8a9e\u97f3\u8fa8\u8a8d\u7cfb\u7d71\u7684\u8a13\u7df4\u8a9e\u6599\u5982\u8868 2 \u6240\u793a\uff0c\u5305\u542b 5 \u500b\u8a9e\u6599\u5eab\uff0c\u5171\u7d04 400 \u5c0f\u6642\u3002\u6e2c\u8a66 \u8a9e\u6599\u5982\u8868 3 \u6240\u793a\uff0c\u5305\u542b 7 \u7d44\u6e2c\u8a66\u8a9e\u6599\u3002\u5c24\u5176\u662f\u5728\u6e2c\u8a66\u8a9e\u6599\u90e8\u5206\uff0c\u589e\u52a0\u4e86\u5f9e\u6559\u80b2\u5ee3\u64ad\u96fb\u81fa \u7bc0\u76ee\u6311\u51fa\u9304\u97f3\u54c1\u8cea\u8f03\u5dee\uff0c\u8fa8\u8a8d\u7387\u8f03\u4e0d\u597d\u7684 NER-set1 (\u6280\u8077\u6700\u524d\u7dda)\uff0c\u4ee5 \u53ca\u9304\u97f3\u54c1\u8cea\u8f03\u4f73\uff0c \u8fa8\u8a8d\u7387\u8f03\u597d\u7684 NER-set2 (\u570b\u969b\u6559\u80b2\u5fc3\u52d5\u7dda)\uff0c\u4ee5\u6e2c\u8a66\u534a\u76e3\u7763\u5f0f\u8a13\u7df4\u5c0d\u4e0d\u540c\u5ee3\u64ad\u7bc0\u76ee\u7684\u8f49 \u5beb\u6548\u80fd\u3002 Total 413.2 7,059 237,629 \u8868 3 \u534a\u76e3\u7763\u5f0f\u8a13\u7df4\u7cfb\u7d71\u6e2c\u8a66\u8a9e\u6599 \u6e2c\u8a66\u8a9e\u6599 \u6642\u6578 \u8a9e\u8005\u6578 \u8a9e\u53e5\u6578 NER-set1 1.75 35 438 NER-set2 3.23 23 640 MATBN (test) 3.06 273 729 OC16-CE80 (test) 7.93 142 7,099 SEAME 13.70 18 12,104 Librispeech (test-other) 5.10 33 2,939 Librispeech (test-clean) 5.40 40 2,620 Total 40.17 564 26,569 3.1.2. QE \u6a21\u578b\u8a13\u7df4\u8a9e\u6599 \u5728\u6b64\u5229\u7528\u4eba\u5de5\u5148\u6a19\u8a18\u90e8\u5206\u6559\u80b2\u96fb\u81fa\u8a9e\u6599\u5eab\uff0c\u4f86\u9032\u884c QE \u932f\u8aa4\u7387\u9810\u6e2c\u6a21\u578b\u8a13\u7df4\u3002\u5176\u4e2d \u5305\u542b\u516b\u500b\u4e0d\u540c\u7bc0\u76ee\u5ee3\u64ad\u8a9e\u6599\uff0c\u7e3d\u8a08\u7d04 65 \u5c0f\u6642\u300110526 \u8a9e\u53e5\u6578\u3002\u8a73\u7d30\u8cc7\u6599\u5982\u8868 4 \u6240\u793a\u3002 \u8868 4 QE \u932f\u8aa4\u7387\u9810\u6e2c\u6a21\u578b\u8a13\u7df4\u8a9e\u6599 \u8a13\u7df4\u8a9e\u6599 \u6642\u6578 \u8a9e\u53e5\u6578 \u5275\u8a2d\u5e02\u96c6 On-Air 10.73 3,000 \u591a\u611b\u81ea\u5df1\u4e00\u9ede\u9ede 6.77 720 \u5152\u7ae5\u65b0\u805e 0.86 166 \u570b\u969b\u6559\u80b2\u5fc3\u52d5\u7dda 3.23 640 \u6280\u8077\u6700\u524d\u7dda 1.75 438 \u79d1\u5b78 So Easy 1.84 208 \u96d9\u8a9e\u65b0\u805e 34.49 4,042 \u6587\u6559\u65b0\u805e 5.46 1,312 Total 65.13 10,526 \u593e\u96dc\u8a9e\u97f3\u8fa8\u8a8d\u5668\u3002 3.2.1. \u97f3\u7d20\u5171\u4eab \u5728\u4e0d\u540c\u8a9e\u8a00\u4e4b\u9593\u5b58\u5728\u8457\u76f8\u8fd1\u97f3\u7684\u73fe\u8c61\uff0c\u65e2\u7136\u662f\u76f8\u4f3c\u7684\u97f3\u5c31\u4e0d\u9700\u8981\u56e0\u70ba\u8a9e\u8a00\u7684\u4e0d\u540c\u800c \u5206\u958b\u8a13\u7df4\uff0c\u800c\u662f\u5c07\u5176\u76f8\u8fd1\u7684\u767c\u8072\u97f3\u7d20\u5408\u4f75\u5efa\u6a21\u8a13\u7df4\uff0c\u5728\u6b64\u5be6\u9a57\u4e2d\u6240\u6709\u97f3\u7d20\u7de8\u78bc\u898f\u5247\u7686\u4f7f \u500b\u97f3\u7d20\u3002 \u8868 5 \u97f3\u7d20\u5171\u4eab\u8868 \u5171\u4eab(\u5b50\u97f3)\u97f3\u7d20 \u8a3b\u97f3 j\u3001w\u3001t\u3001s\u3001p\u3001n\u3001m\u3001l\u3001k\u3001f \u3127\u3001\u3128\u3001\u3109\u3001\u3119\u3001\u3105\u3001\u310b\u3001\u3107\u3001\u310c\u3001\u310d\u3001\u3108 3.2.2. \u4e2d\u82f1\u6df7\u5408\u5b57\u5178 \u6211\u5011\u8a2d\u5b9a\u4e86 X-SAMPA \u6709 190 \u500b\u97f3\u7d20\u8207\u6574\u7406\u5b8c\u8a13\u7df4\u6587\u672c\u5f8c\uff0c\u9700\u8981\u6574\u5408\u6210\u4e00\u500b\u4e2d\u82f1 \u6df7\u5408\u5b57\u5178\uff0c\u5f9e\u8a13\u7df4\u6587\u672c\u4e2d\u6311\u51fa\u6240\u6709\u4e0d\u91cd\u8907\u7684\u55ae\u8a5e\uff0c\u4e26\u4e14\u6309\u7167 X-SAMPA \u6a19\u8a18\u51fa\u97f3\u7d20\uff0c \u6211\u5011\u7684\u4e2d\u82f1\u6df7\u5408\u5b57\u5178\u6700\u7d42\u6709 455,715 \u500b\u5b57\u8a5e\u3002 3.2.3. \u4e2d\u82f1\u6df7\u5408\u8a9e\u8a00\u6a21\u578b \u8868 6 \u5247\u70ba\u7a2e\u5b50\u7cfb\u7d71\u8a9e\u8a00\u6a21\u578b\u7684\u8a13\u7df4\u6587\u672c\uff0c\u5728\u6b64\u5be6\u9a57\u6211\u5011\u96c6\u7d50\u8af8\u591a\u8a9e\u6599\u5eab\u7684\u6587\u672c\u9032\u884c \u8a13\u7df4\u3002\u4f46\u70ba\u4e86\u907f\u514d inside test \u60c5\u5f62\u767c\u751f\uff0c\u6211\u5011\u53ea\u62bd\u53d6\u8a9e\u6599\u5eab\u7684\u8a13\u7df4\u8a9e\u6599\u4e4b\u6587\u672c\u505a\u70ba\u8a13\u7df4 \u5411\u91cf\u56de\u6b78(Support Vector Regression, SVR)[9]\u3001( 2)\u6975\u7aef\u96a8\u6a5f\u6a39(Extremely randomized \u8cc7\u6599\uff0c\u4e26\u4e14\u4f7f\u7528 4-gram \u8207 RNNLM \u4f86\u5efa\u69cb\u8a9e\u8a00\u6a21\u578b\u3002 Giga Word 500,000 9,899,664 Total 737,629 22,463,991 4. \u5be6\u9a57\u7d50\u679c\u8207\u5206\u6790 4.1. \u5be6\u9a57\u4e00-\u7a2e\u5b50\u8a9e\u97f3\u8fa8\u8a8d\u7cfb\u7d71\u6548\u80fd \u6211\u5011\u5148\u6e2c\u8a66\u7a2e\u5b50\u8a9e\u97f3\u8fa8\u8a8d\u7cfb\u7d71\u7684\u8fa8\u8a8d\u6548\u80fd\uff0c\u5c24\u5176\u662f\u5c0d\u5f9e\u6559\u80b2\u5ee3\u64ad\u96fb\u81fa\u7bc0\u76ee\u4e2d\u6311\u51fa\uff0c \u8fa8\u8a8d\u7387\u8f03\u5dee\u7684 NER-set1 (\u6280\u8077\u6700\u524d\u7dda)\uff0c\u4ee5\u53ca\u8fa8\u8a8d\u7387\u8f03\u597d\u7684 NER-set2 (\u570b\u969b\u6559\u80b2\u5fc3\u52d5\u7dda)\uff0c \u5206\u5225\u505a\u6e2c\u8a66\uff0c\u4ee5\u6b64\u505a\u70ba Baseline \u7cfb\u7d71\u7684\u6548\u80fd\u53c3\u8003\u503c\u3002 \u8868 6 \u70ba\u7a2e\u5b50\u8fa8\u8a8d\u5668\u8fa8\u8a8d\u6548\u80fd\u5be6\u9a57\u7d50\u679c\u3002\u5f9e\u8868 7 \u53ef\u4ee5\u770b\u5230\uff0c\u96d6\u7136\u90fd\u662f\u4f86\u81ea\u6559\u80b2\u5ee3\u64ad\u96fb \u81fa\u7684\u8a9e\u6599\uff0c\u4f46\u662f\u56e0\u70ba\u7bc0\u76ee\u9304\u97f3\u54c1\u8cea\u4e0d\u540c\uff0c\u4e3b\u6301\u4eba\u53ca\u4f86\u8cd3\u7684\u53e3\u8a9e\u4e0d\u540c\uff0c\u6240\u8ac7\u8ad6\u7684\u8a71\u984c\u4e0d\u540c\uff0c \u5728\u6574\u9ad4\u7684\u8fa8\u8a8d\u4e0a\u9019\u5169\u500b\u6e2c\u8a66\u8a9e\u6599\u7684\u932f\u8aa4\u7387\u5dee\u8ddd\u76f8\u7576\u5927(\u76f8\u5dee\u7d04 10%) \u3002 \u8868 7 \u7a2e\u5b50\u8a9e\u97f3\u8fa8\u8a8d\u7cfb\u7d71\u8fa8\u8a8d\u7387 \u6e2c\u8a66\u8a9e\u6599 \u7a2e\u5b50\u6a21\u578b NER-set1 25.00 NER-set2 14.24 MATBN (test) 13.18 OC16-CE80 (test) 16.30 SEAME 36.32 Librispeech (test-other) 18.17 Librispeech (test-clean) 5.00 4.2. \u5be6\u9a57\u4e8c-QE \u6a21\u578b\u8a13\u7df4\u7d50\u679c Average2 (without SEAME) 13.21 13.08 12.93 12.92 Average 18.32 17.86 17.96 18.04 18.17 NER-set2(\u570b\u969b\u6559\u80b2\u5fc3\u52d5\u7dda)\u76f8\u5c0d\u6539\u5584\u7387\u4f86\u5230 11.3 %\u3002 13.09 \u6b64\u5be6\u9a57\u4e2d\u5c07\u64f7\u53d6\u7684 93 \u500b\u7279\u5fb5\u53c3\u6578\uff0c\u5229\u7528\u4e09\u7a2e\u4e0d\u540c\u8ff4\u6b78\u8a13\u7df4\u67b6\u69cb\uff0c\u5305\u542b(1)\u652f\u63f4 \u9032\u884c\u516b\u6b21\u8a13\u7df4\u8207\u6e2c\u8a66\uff0c\u6700\u5f8c\u518d\u5c07\u516b\u6b21\u6e2c\u8a66\u7d50\u679c\u6240\u8a08\u7b97\u7684 MAE\u3001MSE \u52a0\u7e3d\u5e73\u5747\u3002 \u4f7f\u7528\u8a9e\u6599\u6642\u6578 Total CM1 CM2 QE1 QE2 \u5ee3\u64ad\u7bc0\u76ee\u8a9e\u6599(hour) 377.58 209.25 38.28 38.44 38.21 CER in % \u76f8\u5c0d\u6539\u5584\u7387 12 100 \u8fa8\u8a8d\u6a21\u578b NER-set1 NER-set2 NER-set1 NER-set2 10 0 5. \u7d50\u8ad6 echo model training \u9996\u5148\uff0c\u7531\u65bc SVR \u8207 Extra-Trees \u662f\u5c6c\u65bc\u6dfa\u5c64\u5206\u6790\uff0c\u5728 DNN \u90e8\u5206\u5148\u53ea\u4f7f\u7528\u4e00\u5c64\u96b1\u85cf \u5176\u4e2d\uff0cCM1 \u662f\u4ee5\u6bcf\u4e00\u5b57\u8a5e\u7684\u4fe1\u5fc3\u5ea6\u8a55\u4f30\uff0c\u4f9d\u7167\u8f03\u9ad8\u7684\u8f49\u5beb\u4fe1\u5fc3\u7a0b\u5ea6( \u7a2e\u5b50\u6a21\u578b 25.00 14.24 --\u2265 0.9)\uff0c CM1(610h) 24.66 13.85 1.36% 2.74% \u57fa\u672c\u8a13\u7df4\u8a9e\u6599 QE\u8a13\u7df4\u8a9e\u6599 NER-set1 NER-set2 \u6211\u5011\u7528 QE \u6a21\u578b\u6311\u9078\u534a\u76e3\u7763\u5f0f\u5b78\u7fd2\u8a9e\u6599\uff0c\u5c0d\u8fa8\u8a8d\u7387\u8f03\u5dee\u7684 NER-set1 \u5b57\u5143\u932f\u8aa4\u7387 \u5c64\u8207 SVR \u548c Extra-Trees \u6bd4\u8f03\u3002\u5be6\u9a57\u7d50\u679c\u5982\u8868 8 \u6240\u793a\uff0c\u53ef\u4ee5\u770b\u5230\u4f7f\u7528 DNN \u67b6\u69cb\u8a13\u7df4\u51fa \u4f86\u7684\u9810\u6e2c\u6a21\u578b\uff0c\u5176\u932f\u8aa4\u7387\u8aa4\u5dee\u6700\u5c0f\u3002 \u8868 8 QE \u932f\u8aa4\u7387\u9810\u6e2c\u6a21\u578b\u4e09\u7a2e\u67b6\u69cb\u8a13\u7df4\u7d50\u679c\u6bd4\u8f03 \u5ee3\u64ad\u7bc0\u76ee 1 layer DNN SVR Extra-Trees MAE MSE MAE MSE MAE MSE \u591a\u611b\u81ea\u5df1\u4e00\u9ede\u9ede 0.0819 0.0101 0.0769 0.0110 0.0893 0.0194 \u5275\u8a2d\u5e02\u96c6 On-Air 0.1134 0.0201 0.1057 0.0230 0.0991 0.0196 \u5152\u7ae5\u65b0\u805e 0.0938 0.0141 0.1006 0.0157 0.0979 0.0147 \u570b\u969b\u6559\u80b2\u5fc3\u52d5\u7dda 0.0742 0.0100 0.0875 0.0129 0.0930 0.0138 \u6280\u8077\u6700\u524d\u7dda 0.0898 0.0128 0.0878 0.0138 0.0944 0.0153 \u79d1\u5b78 So Easy 0.0676 0.0069 0.0793 0.0098 0.0688 0.0096 \u96d9\u8a9e\u65b0\u805e 0.0989 0.0163 0.1103 0.0220 0.1078 0.0219 \u6587\u6559\u65b0\u805e 0.0966 0.0133 0.0932 0.0141 0.0965 0.0146 Average 0.0895 0.0129 0.0927 0.0153 0.0934 0.0161 \u7136\u5f8c\uff0c\u6211\u5011\u518d\u8a13\u7df4\u591a\u5c64 DNN\uff0c\u770b\u591a\u5c64 DNN \u662f\u5426\u53ef\u4ee5\u9032\u4e00\u6b65\u63d0\u5347\u9810\u6e2c\u6548\u679c\u3002\u5f9e\u5be6 \u9a57\u7d50\u679c\u4e2d\u4f7f\u7528 2 layer DNN \u6703\u6709\u8f03\u4f4e\u7684\u9810\u6e2c\u8aa4\u5dee\uff0c\u7e3d\u7d50\u5982\u4e0b\u8868 9 \u6240\u793a\u3002\u56e0\u6b64\uff0c\u5728\u4ee5\u4e0b \u7684\u975e\u76e3\u7763\u5f0f\u5b78\u7fd2\u5be6\u9a57\u4e2d\uff0c\u7686\u4f7f\u7528\u5169\u5c64\u7684 DNN \u6a21\u578b\u505a QE \u9810\u6e2c\u3002 \u8868 9 QE \u932f\u8aa4\u7387\u9810\u6e2c\u6a21\u578b\u6e2c\u8a66\u7d50\u679c \u5ee3\u64ad\u7bc0\u76ee 1 layer DNN 2 layer DNN 3 layer DNN MAE MSE MAE MSE MAE MSE Average 0.0894 0.0164 0.0855 0.0130 0.0865 0.0134 \u4ee5\u8a5e\u70ba\u55ae\u4f4d\u505a\u6311\u9078\u3002CM2 \u5247\u662f\u5148\u8a08\u7b97\u6bcf\u4e00\u53e5\u4e2d\u6240\u6709\u5b57\u8a5e\u4fe1\u5fc3\u5ea6\u8a55\u4f30\uff0c\u518d\u4ee5\u53e5\u70ba\u55ae\u4f4d\u53d6 \u5e73\u5747( CM2(438h) 23.88 13.26 4.48% 6.88% QE1(438h) 24.04 13.11 3.84% 7.94% (CER) \u4f86\u5230 23.61%\uff0c\u8fa8\u8a8d\u7387\u8f03\u597d\u7684 NER-set2 \u5b57\u5143\u932f\u8aa4\u7387 (CER) \u5176\u8fa8\u8a8d\u7387\u4f86\u5230 13.24%\u3002 \u5716 8 \u534a\u76e3\u7763\u5f0f\u5b78\u7fd2\u8072\u5b78\u6a21\u578b\u8a13\u7df4\u7d50\u679c(\u6539\u5584\u66f2\u7dda) \u2265 0.9)\uff0c\u9032\u884c\u6311\u9078\u3002QE1 \u4f7f\u7528 93 \u500b\u7279\u5fb5\u53c3\u6578\uff0c\u9810\u6e2c\u51fa\u932f\u8aa4\u7387\uff0c\u6311\u9078\u932f\u8aa4\u7387 ( QE2(438h) 23.88 13.35 4.48% 6.25% \u82e5\u589e\u52a0\u4e86\u66f4\u5b8c\u6574\u7684\u8a9e\u8a00\u6a21\u578b\u8a13\u7df4\u6587\u672c Giga Word2\uff0c\u80fd\u8b93 NER-set1 \u7684\u6700\u4f73\u8fa8\u8a8d\u7387\u4f86\u5230 4.5. \u5be6\u9a57\u4e94-\u8a9e\u8a00\u6a21\u578b\u6539\u5584 < 0.3)\u8f03\u4f4e\u7684\u8a9e\u53e5\u3002\u6700\u5f8c\uff0cQE2 \u662f\u5c07 CM \u503c\u518d\u52a0\u5165\u539f\u6709\u7684 93 \u500b\u7279\u5fb5\u53c3\u6578\uff0c\u8a13\u7df4\u51fa \u65b0\u7684 QE \u932f\u8aa4\u7387\u9810\u6e2c\u6a21\u578b\uff0c\u4e00\u6a23\u5f9e\u932f\u8aa4\u7387\u8f03\u4f4e( 4.4. \u5be6\u9a57\u56db-\u6311\u9078\u8a9e\u6599\u91cf\u8207\u6548\u80fd\u6bd4\u8f03 23.25%\uff0cNER-set2 \u7684\u6700\u4f73\u8fa8\u8a8d\u7387\u4f86\u5230 12.63%\u3002 \u5728\u4e0a\u4e00\u7bc0\u7684\u5be6\u9a57\u4e2d\uff0c\u91dd\u5c0d\u8072\u5b78\u6a21\u578b\u7684\u8a13\u7df4\u4e0a\u5f9e\u539f\u672c\u7684\u7a2e\u5b50\u8a13\u7df4\u8a9e\u6599 400 \u5c0f\u6642\u589e\u52a0\u4e86 < 0.3)\u7684\u8a9e\u53e5\u958b\u59cb\u6311\u9078\u3002 \u4ee5\u4e0b\u5be6\u9a57\u91dd\u5c0d QE1 \u932f\u8aa4\u7387\u9810\u6e2c\u6a21\u578b\u6240\u6311\u51fa\u7684\u8a9e\u6599\uff0c\u4ee5\u6f38\u9032\u7684\u65b9\u5f0f\u52a0\u5165\u534a\u76e3\u7763\u5f0f\u5b78 \u6700\u5f8c\uff0c\u6574\u9ad4\u6548\u80fd\u6539\u5584\u7e3d\u7d50\u5982\u5716 9 \u6240\u793a\uff0c\u5176\u986f\u793a\u4f7f\u7528 QE \u6a21\u578b\u4f86\u6311\u9078\u534a\u76e3\u7763\u5f0f\u5b78\u7fd2\u8a13 198 \u5c0f\u6642\u7684\u534a\u76e3\u7763\u5f0f\u5b78\u7fd2\u8a13\u7df4\u8a9e\u6599\u91cd\u65b0\u8a13\u7df4\uff0c\u6539\u5584\u7a0b\u5ea6\u5df2\u7d93\u8da8\u8fd1\u65bc\u6536\u6582\u72c0\u614b\u3002\u800c\u5728\u5be6\u9a57 \u4f9d\u6311\u9078\u7d50\u679c\uff0c\u6211\u5011\u5404\u81ea\u5c07\u6240\u6311\u9078\u51fa\u7684\u8a9e\u6599\u52a0\u5165\u539f\u5148\u7684\u8a13\u7df4\u8a9e\u6599\uff0c\u91cd\u65b0\u8a13\u7df4\u56db\u500b\u8072\u5b78 \u7fd2\u8a13\u7df4\u8a9e\u6599\uff0c\u91cd\u65b0\u8a13\u7df4\u8072\u5b78\u6a21\u578b\uff0c\u6e2c\u8a66\u6311\u9078\u8a9e\u6599\u91cf\u8207\u6548\u80fd\u7684\u5f71\u97ff\u3002\u8a9e\u6599\u6311\u9078\u9806\u5e8f\u70ba\u4f9d\u9810 \u7df4\u8a9e\u6599\uff0c\u91cd\u65b0\u8a13\u7df4\u8072\u5b78\u6a21\u578b\uff0c\u78ba\u5be6\u80fd\u6709\u6548\u63d0\u5347\u8072\u5b78\u6a21\u578b\u4e4b\u6548\u80fd\u3002 \u4e94\uff0c\u6211\u5011\u5c07\u52a0\u5165\u66f4\u8c50\u6c9b\u7684\u8a13\u7df4\u6587\u672c\uff0c\u4f86\u8a13\u7df4\u6211\u5011\u7684\u8a9e\u8a00\u6a21\u578b\u3002\u4e3b\u8981\u662f\u5728\u7a2e\u5b50\u7cfb\u7d71\u8a9e\u8a00\u6a21 \u6a21\u578b\u4ee5\u6e2c\u8a66\u56db\u7a2e\u6311\u9078\u534a\u76e3\u7763\u5f0f\u5b78\u7fd2\u8a13\u7df4\u8a9e\u6599\u65b9\u6cd5\u7684\u6548\u80fd\u3002 \u6e2c\u8a66\u7d50\u679c\u5982\u8868 11 \u6240\u793a\uff0c\u53ef\u4ee5\u767c\u73fe\u4f7f\u7528 QE \u6216\u662f CM \u503c\u6311\u9078\u8a9e\u6599\uff0c\u6240\u8a13\u7df4\u51fa\u4f86\u7684\u8072 \u5b78\u6a21\u578b\uff0c\u5728\u4e0d\u540c\u6e2c\u8a66\u8a9e\u6599\u7684\u8a9e\u97f3\u8fa8\u8a8d\u4e0a\u90fd\u80fd\u964d\u4f4e\u6574\u9ad4\u7684\u932f\u8aa4\u7387\uff0c\u4e26\u4e14\u7528 QE1 \u6216\u662f CM2 \u6311\u9078\u8a13\u7df4\u8a9e\u6599\uff0c\u8a13\u7df4\u51fa\u7684\u8072\u5b78\u6a21\u578b\uff0c\u5728\u5e73\u5747\u8a9e\u97f3\u8fa8\u8a8d\u932f\u8aa4\u7387\u4e0a\u90fd\u6709\u8f03\u597d\u7684\u8868\u73fe\u3002 \u8868 11 \u57fa\u65bc\u534a\u76e3\u7763\u5f0f\u5b78\u7fd2\u8072\u5b78\u6a21\u578b\u8a13\u7df4\u7d50\u679c \u6e2c\u8a66\u8a9e\u6599 \u7a2e\u5b50\u6a21\u578b CM1(610h) CM2(438h) QE1(438h) QE2(438h) NER-set1 25.00 24.66 23.88 24.04 23.88 NER-set2 14.24 13.85 13.26 13.11 13.35 MATBN (test) 13.18 13.19 13.00 13.00 13.24 OC16-CE80 (test) 16.30 15.96 16.10 16.08 15.95 SEAME 36.32 36.70 35.85 35.96 36.13 Librispeech (test-other) 18.17 18.00 17.83 17.87 Librispeech (test-clean) 5.00 5.18 5.02 4.96 Average1 (with SEAME) 20.53 20.57 20.19 20.22 20.39 Librispeech (test-other) 18.17 17.87 18.01 18.59 Librispeech (test-clean) 5.00 4.96 5.22 5.19 5.31 \u7684\u5f71\u97ff\uff0c\u8b93\u8fa8\u8a8d\u7387\u8f03\u5dee\u7684 NER-set1(\u6280\u8077\u6700\u524d\u7dda)\u76f8\u5c0d\u6539\u5584\u7387\u4f86\u5230 7%\uff0c\u8fa8\u8a8d\u7387\u8f03\u597d\u7684 18.51 5.2 SEAME 36.32 35.96 36.15 36.45 36.84 \u5be6\u9a57\u7684\u7d50\u679c\u5982\u8868 15 \u6240\u793a\uff0c\u53ef\u4ee5\u770b\u5230\u8a9e\u8a00\u6a21\u578b\u5efa\u6a21\u7684\u80fd\u529b\u5c0d\u8a9e\u97f3\u8b58\u5225\u7684\u7d50\u679c\u6709\u4e00\u5b9a 18.15 \u6e2c\u932f\u8aa4\u7387\u5f9e\u4f4e\u5230\u9ad8\uff0c\u6311\u51fa\u56db\u7d44\uff0c\u5404\u6709 38 \u5c0f\u6642\uff0c50 \u5c0f\u6642\u300150 \u5c0f\u6642\u8207 60 \u5c0f\u6642\u3002\u56e0\u6b64\u8a13\u7df4 \u8a9e\u6599\u7e3d\u6642\u6578\u5206\u8b8a\u6210\u70ba 438 \u5c0f\u6642\u3001488 \u5c0f\u6642\u3001538 \u5c0f\u6642\u3001598 \u5c0f\u6642\u56db\u7d44\u3002 \u6311\u9078\u8a9e\u6599\u91cf\u8207\u6548\u80fd\u5be6\u9a57\u7d50\u679c\u5982\u8868 13 \u6240\u793a\u3002\u53e6\u5916\uff0c\u5716 8 \u70ba\u53ea\u91dd\u5c0d\u6559\u80b2\u96fb\u81fa\u6e2c\u8a66\u8a9e\u6599 \u4f86\u770b\u534a\u76e3\u7763\u5f0f\u5b78\u7fd2\u7684\u6548\u80fd\u6539\u5584\u66f2\u7dda\uff0c\u5be6\u9a57\u7d50\u679c\u986f\u793a\u5c0d NER-set1\uff0c\u96a8\u8457\u52a0\u5165\u7684\u8a13\u7df4\u8a9e\u6599\u7684 \u6642\u6578\u589e\u52a0\uff0c\u6240\u8a13\u7df4\u51fa\u4f86\u7684\u8a9e\u97f3\u8fa8\u8a8d\u5668\uff0c\u7684\u78ba\u6709\u66f4\u597d\u7684\u8fa8\u8a8d\u6548\u80fd\u3002 \u8868 13 \u4e0d\u540c\u8a9e\u6599\u91cf\u5c0d\u534a\u76e3\u7763\u5f0f\u5b78\u7fd2\u578b\u8a13\u7df4\u7d50\u679c\u7684\u5f71\u97ff \u6e2c\u8a66\u8a9e\u6599 \u7a2e\u5b50\u6a21\u578b QE(438h) QE(488h) QE(538h) QE(598h) NER-set1 25.00 24.04 24.00 23.86 23.61 NER-set2 14.24 13.11 13.23 12.96 13.24 MATBN (test) 13.18 13.00 13.04 13.12 13.28 OC16-CE80 (test) 16.30 16.08 16.08 16.11 16.40 \u578b\u8a13\u7df4\u6587\u672c\u4e2d\uff0c\u589e\u52a0\u4e86\u66f4\u5b8c\u6574\u7684\u8a13\u7df4\u6587\u672c Giga Word2\u3002\u5176\u7d71\u8a08\u8cc7\u6599\u5982\u8868 14 \u6240\u793a\u3002 \u8868 14 \u8a9e\u8a00\u6a21\u578b\u8a13\u7df4\u6587\u672c \u8a13\u7df4\u6587\u672c \u8a9e\u53e5\u6578 \u5b57\u8a5e\u6578 TCC300 (all) 27,375 186,369 MATBN (train) 29,549 1,264,625 OC16-CE80 (train) 58,132 509,657 SEAME 94,034 1,200,121 Librispeech (train-960) 28,539 9,403,555 Giga Word 500,000 9,899,664 Giga Word2 16,500,000 441,889,701 Total 17,237,629 464,353,692 \u5716 9 \u5ee3\u64ad\u8a9e\u6599\u6e2c\u8a66\u7d50\u679c\u6539\u5584\u7387 \u53c3\u8003\u6587\u737b [1] 14.24 23.61 13.24 23.25 12.63 10 15 20 25 Wessel, 25 NER-set1 NER-set2 \u7a2e\u5b50\u8fa8\u8a8d\u5668 QE1(598h) QE1(598h)+LM2</td></tr></table>", | |
| "text": "\u500b\u4e0d\u540c\u7bc0\u76ee\uff0c\u7e3d\u8a08\u7d04\u70ba 377 \u5c0f\u6642\u7684\u672a\u6a19\u8a18\u8a9e\u6599\u3002\u7d93\u904e\u4fe1\u5fc3\u5ea6\u8a55\u4f30 CM \u53ca\u54c1\u8cea\u4f30\u7b97 QE\uff0c \u5206\u70ba QE1\uff0cQE2\uff0cCM1 \u8207 CM2 \u56db\u7a2e\u6311\u9078\u6a5f\u5236\u3002CM1 \u6311\u51fa\u7d04 210 \u5c0f\u6642(\u6b64\u90e8\u5206\u56e0\u8a9e\u6599\u6642 \u6578\u4e0d\u540c\uff0c\u50c5\u70ba\u53c3\u8003\u7528) \uff0c\u800c QE1\uff0cQE2 \u8207 CM2 \u5404\u6311\u9078\u51fa\u7d04\u83ab 38 \u5c0f\u6642\u7684\u534a\u76e3\u7763\u5f0f\u5b78\u7fd2\u8a13 F. and Ney, H., \"Unsupervised Training of Acoustic Models for Large Vocabulary Continuous Speech Recognition,\" IEEE Transactions on Speech and Audio Processing, vol. 13, no. 1, 2005, pp. 257-265.", | |
| "html": null, | |
| "type_str": "table" | |
| } | |
| } | |
| } | |
| } |