| { | |
| "paper_id": "O15-1011", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T08:09:57.366414Z" | |
| }, | |
| "title": "\u878d\u5408\u591a\u7a2e\u6df1\u5c64\u985e\u795e\u7d93\u7db2\u8def\u8072\u5b78\u6a21\u578b\u8207\u5206\u985e\u6280\u8853\u65bc\u83ef\u8a9e\u932f\u8aa4\u767c\u97f3\u6aa2 \u6e2c\u4e4b\u7814\u7a76 Exploring Combinations of Various Deep Neural Network based Acoustic Models and Classification Techniques for Mandarin Mispronunciation Detection", | |
| "authors": [ | |
| { | |
| "first": "Yao-Chi", | |
| "middle": [], | |
| "last": "\u8a31\u66dc\u9e92", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Ming-Han", | |
| "middle": [], | |
| "last": "Hsu", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Hsiao-Tsung", | |
| "middle": [], | |
| "last": "Yang", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "mh_yang@ntnu.edu.tw" | |
| }, | |
| { | |
| "first": "Yuwen", | |
| "middle": [], | |
| "last": "Hung", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Yao-Ting", | |
| "middle": [], | |
| "last": "Hsiung", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "ywhsiung@cycu.edu.tw" | |
| }, | |
| { | |
| "first": "Berlin", | |
| "middle": [], | |
| "last": "Sung", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "sungtc@ntnu.edu.tw" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Chen", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| } | |
| ], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c(mispronunciation detection)\u70ba\u96fb\u8166\u8f14\u52a9\u767c\u97f3\u8a13\u7df4(computer assisted pronunciation training, CAPT)\u7814\u7a76\u4e2d\u5341\u5206\u91cd\u8981\u7684\u4e00\u500b\u74b0\u7bc0\uff0c\u5176\u76ee\u7684\u662f\u56de\u994b\u7d66\u8a9e\u8a00\u5b78\u7fd2\u8005\u662f\u5426\u5728 \u5176\u8b80\u8aa6\u4e00\u6bb5\u8a71\u4e2d\u7684\u51fa\u73fe\u932f\u8aa4\u767c\u97f3\u3002\u4e00\u822c\u800c\u8a00\uff0c\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c\u6d41\u7a0b\u53ef\u5206\u70ba\u5169\u90e8\u5206\uff1a1)\u524d\u7aef \u7279\u5fb5\u64f7\u53d6\u6a21\u7d44\uff0c\u57fa\u65bc\u5b78\u7fd2\u8005\u6240\u5ff5\u8aa6\u7684\u97f3\u7d20\u6216\u8a9e\u53e5\u6bb5\u843d\u548c\u8072\u5b78\u6a21\u578b(acoustic model)\u7684\u6bd4\u5c0d \u4ee5\u64f7\u53d6\u5c0d\u61c9\u7684\u5177\u6709\u9451\u5225\u6027\u4e4b\u767c\u97f3\u6aa2\u6e2c\u7279\u5fb5\uff1b2)\u5f8c\u7aef\u5206\u985e\u6a21\u7d44\uff0c\u57fa\u65bc\u6240\u6c42\u5f97\u767c\u97f3\u6aa2\u6e2c\u7279\u5fb5\uff0c \u5224\u65b7\u97f3\u7d20\u6216\u8a9e\u53e5\u6bb5\u843d\u6240\u6b78\u5c6c\u985e\u5225(\u6b63\u78ba\u767c\u97f3\u6216\u932f\u8aa4\u767c\u97f3)\u3002\u5728\u672c\u7bc7\u8ad6\u6587\u5ef6\u7e8c\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c \u7814\u7a76\u800c\u4e3b\u8981\u6709\u4e09\u9805\u8ca2\u737b\uff1a1)\u6bd4\u8f03\u4e26\u7d50\u5408\u7576\u524d\u57fa\u65bc\u6df1\u5c64\u985e\u795e\u7d93\u7db2\u8def(deep neural networks, DNN)\u8207\u647a\u7a4d\u985e\u795e\u7d93\u7db2\u8def(convolutional neuron networks, CNN)\u4e4b\u5148\u9032\u7684\u8072\u5b78\u6a21\u578b\u4ee5\u7522\u751f \u66f4\u5177\u9451\u5225\u6027\u767c\u97f3\u6aa2\u6e2c\u7279\u5fb5\uff1b2)\u6211\u5011\u6bd4\u8f03\u4e26\u7d50\u5408\u4e0d\u540c\u5206\u985e\u65b9\u6cd5\uff0c\u4ee5\u671f\u80fd\u9054\u5230\u66f4\u4f73\u7684\u767c\u97f3\u6aa2 \u6e2c\u8868\u73fe\uff1b3)\u91dd\u5c0d\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c\u6240\u5305\u62ec\u7684\u6a21\u7d44\uff0c\u9032\u884c\u4e00\u7cfb\u5217\u5ee3\u6cdb\u4e14\u6df1\u5165\u7684\u5be6\u9a57\u5206\u6790\u8207\u8a0e\u8ad6\u3002 \u5f9e\u4e00\u5957\u4ee5\u83ef\u8a9e\u505a\u70ba\u7b2c\u4e8c\u8a9e\u5b78\u7fd2\u76ee\u6a19\u8a9e\u8a00\u7684\u5927\u91cf\u8a9e\u6599\u5eab\u4e4b\u5be6\u9a57\u7d50\u679c\u986f\u793a\uff0c\u6211\u5011\u6240\u63d0\u51fa\u878d\u5408 \u591a\u7a2e\u6df1\u5c64\u985e\u795e\u7d93\u7db2\u8def\u8072\u5b78\u6a21\u578b\u8207\u5206\u985e\u6280\u8853\u7684\u65b9\u6cd5\u7684\u78ba\u80fd\u8f03\u57fa\u790e\u65b9\u6cd5\u6709\u986f\u8457\u7684\u6548\u80fd\u63d0\u5347\u3002 \u95dc\u9375\u5b57\uff1a\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c\u3001\u81ea\u52d5\u8a9e\u97f3\u8fa8\u8b58\u3001\u6df1\u5c64\u985e\u795e\u7d93\u7db2\u8def\u3001\u647a\u7a4d\u985e\u795e\u7d93\u7db2\u8def", | |
| "pdf_parse": { | |
| "paper_id": "O15-1011", | |
| "_pdf_hash": "", | |
| "abstract": [ | |
| { | |
| "text": "\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c(mispronunciation detection)\u70ba\u96fb\u8166\u8f14\u52a9\u767c\u97f3\u8a13\u7df4(computer assisted pronunciation training, CAPT)\u7814\u7a76\u4e2d\u5341\u5206\u91cd\u8981\u7684\u4e00\u500b\u74b0\u7bc0\uff0c\u5176\u76ee\u7684\u662f\u56de\u994b\u7d66\u8a9e\u8a00\u5b78\u7fd2\u8005\u662f\u5426\u5728 \u5176\u8b80\u8aa6\u4e00\u6bb5\u8a71\u4e2d\u7684\u51fa\u73fe\u932f\u8aa4\u767c\u97f3\u3002\u4e00\u822c\u800c\u8a00\uff0c\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c\u6d41\u7a0b\u53ef\u5206\u70ba\u5169\u90e8\u5206\uff1a1)\u524d\u7aef \u7279\u5fb5\u64f7\u53d6\u6a21\u7d44\uff0c\u57fa\u65bc\u5b78\u7fd2\u8005\u6240\u5ff5\u8aa6\u7684\u97f3\u7d20\u6216\u8a9e\u53e5\u6bb5\u843d\u548c\u8072\u5b78\u6a21\u578b(acoustic model)\u7684\u6bd4\u5c0d \u4ee5\u64f7\u53d6\u5c0d\u61c9\u7684\u5177\u6709\u9451\u5225\u6027\u4e4b\u767c\u97f3\u6aa2\u6e2c\u7279\u5fb5\uff1b2)\u5f8c\u7aef\u5206\u985e\u6a21\u7d44\uff0c\u57fa\u65bc\u6240\u6c42\u5f97\u767c\u97f3\u6aa2\u6e2c\u7279\u5fb5\uff0c \u5224\u65b7\u97f3\u7d20\u6216\u8a9e\u53e5\u6bb5\u843d\u6240\u6b78\u5c6c\u985e\u5225(\u6b63\u78ba\u767c\u97f3\u6216\u932f\u8aa4\u767c\u97f3)\u3002\u5728\u672c\u7bc7\u8ad6\u6587\u5ef6\u7e8c\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c \u7814\u7a76\u800c\u4e3b\u8981\u6709\u4e09\u9805\u8ca2\u737b\uff1a1)\u6bd4\u8f03\u4e26\u7d50\u5408\u7576\u524d\u57fa\u65bc\u6df1\u5c64\u985e\u795e\u7d93\u7db2\u8def(deep neural networks, DNN)\u8207\u647a\u7a4d\u985e\u795e\u7d93\u7db2\u8def(convolutional neuron networks, CNN)\u4e4b\u5148\u9032\u7684\u8072\u5b78\u6a21\u578b\u4ee5\u7522\u751f \u66f4\u5177\u9451\u5225\u6027\u767c\u97f3\u6aa2\u6e2c\u7279\u5fb5\uff1b2)\u6211\u5011\u6bd4\u8f03\u4e26\u7d50\u5408\u4e0d\u540c\u5206\u985e\u65b9\u6cd5\uff0c\u4ee5\u671f\u80fd\u9054\u5230\u66f4\u4f73\u7684\u767c\u97f3\u6aa2 \u6e2c\u8868\u73fe\uff1b3)\u91dd\u5c0d\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c\u6240\u5305\u62ec\u7684\u6a21\u7d44\uff0c\u9032\u884c\u4e00\u7cfb\u5217\u5ee3\u6cdb\u4e14\u6df1\u5165\u7684\u5be6\u9a57\u5206\u6790\u8207\u8a0e\u8ad6\u3002 \u5f9e\u4e00\u5957\u4ee5\u83ef\u8a9e\u505a\u70ba\u7b2c\u4e8c\u8a9e\u5b78\u7fd2\u76ee\u6a19\u8a9e\u8a00\u7684\u5927\u91cf\u8a9e\u6599\u5eab\u4e4b\u5be6\u9a57\u7d50\u679c\u986f\u793a\uff0c\u6211\u5011\u6240\u63d0\u51fa\u878d\u5408 \u591a\u7a2e\u6df1\u5c64\u985e\u795e\u7d93\u7db2\u8def\u8072\u5b78\u6a21\u578b\u8207\u5206\u985e\u6280\u8853\u7684\u65b9\u6cd5\u7684\u78ba\u80fd\u8f03\u57fa\u790e\u65b9\u6cd5\u6709\u986f\u8457\u7684\u6548\u80fd\u63d0\u5347\u3002 \u95dc\u9375\u5b57\uff1a\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c\u3001\u81ea\u52d5\u8a9e\u97f3\u8fa8\u8b58\u3001\u6df1\u5c64\u985e\u795e\u7d93\u7db2\u8def\u3001\u647a\u7a4d\u985e\u795e\u7d93\u7db2\u8def", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Abstract", | |
| "sec_num": null | |
| } | |
| ], | |
| "body_text": [ | |
| { | |
| "text": "(CNN), and compare their effectiveness for the extraction of discriminative pronunciation detection features. Second, we experiment with different types of classification methods and propose a novel integration of DNN-and CNN-based decision scores at the back-end. Third, we provide an extensive set of empirical evaluations on the aforementioned two modules and associated methods based on a recently compiled corpus for learning Mandarin Chinese as the second language. The experimental results reveal the performance utility of our approach in relation to several existing baselines. ", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "\u793a\u70ba\u7b2c \u5c64\uff0c\u8868\u793a\u6709 + 1\u5c64\u7684\u6df1\u5c64\u985e\u795e\u7d93\u7db2\u8def\uff0c\u6b64\u524d\u994b\u904b\u7b97\u53ef\u4ee5\u8868\u793a\u70ba\uff1a \u2113 \uff1d ( \u2113 )\uff1d (W \u2113 \u2113\u22121 + \u2113 ) , ( \u2113 = 0, 1, 2, 3, \u2026 , ) (1) ( ) = 1 1 + exp (\u2212 ) , 0 < ( ) < 1 (2) \u5f0f(1)\u4e2d\uff0c \u2113 \u2208 \u2115\uff0c\u70ba\u7b2c\u2113\u5c64\u7684\u795e\u7d93\u5143\u6578\u91cf\u3002 \u2113 \u2208 \u211d \u2113 \u00d71 \uff0cW \u2113 \u2208 \u211d \u2113 \u00d7 \u2113\u22121 \uff0c \u2113 \u2208 \u211d \u2113 \u00d71 \uff0c \u2113 \u70ba\u7b2c\u2113\u5c64\u7684\u8f38\u51fa\u5411\u91cf\uff0cW \u2113 \u70ba\u7b2c\u2113\u5c64\u7684\u6b0a\u91cd\u77e9\u9663\uff0c\u901a\u5e38\u63a1\u53d6\u96a8\u6a5f\u521d\u59cb\u5316(", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| } | |
| ], | |
| "back_matter": [], | |
| "bib_entries": { | |
| "BIBREF0": { | |
| "ref_id": "b0", | |
| "title": "40 million people worldwide study Chinese", | |
| "authors": [], | |
| "year": null, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "\"40 million people worldwide study Chinese,\" http://english.people.com.cn/90001/90782/90872/7112508.html .", | |
| "links": null | |
| }, | |
| "BIBREF1": { | |
| "ref_id": "b1", | |
| "title": "A DNN-based acoustic modeling of tonal language and its application to Mandarin pronunciation training", | |
| "authors": [ | |
| { | |
| "first": "W", | |
| "middle": [], | |
| "last": "Hu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Qian", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "F", | |
| "middle": [], | |
| "last": "Soong", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2013, | |
| "venue": "Proc. ICASSP", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "3230--3234", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "W. Hu, Y. Qian, and F. Soong, \"A DNN-based acoustic modeling of tonal language and its application to Mandarin pronunciation training,\" in Proc. ICASSP, pp. 3230-3234, 2013.", | |
| "links": null | |
| }, | |
| "BIBREF2": { | |
| "ref_id": "b2", | |
| "title": "Deep neural networks for acoustic modeling in speech recognition", | |
| "authors": [ | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Hinton", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "L", | |
| "middle": [], | |
| "last": "Deng", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [ | |
| "E" | |
| ], | |
| "last": "Dahl", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Mohamed", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "N", | |
| "middle": [], | |
| "last": "Jaitly", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Senior", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "V", | |
| "middle": [], | |
| "last": "Vanhoucke", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "P", | |
| "middle": [], | |
| "last": "Nguyen", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Sainath", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "B", | |
| "middle": [], | |
| "last": "Kingsbury", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2012, | |
| "venue": "IEEE Signal Processing Magazine", | |
| "volume": "29", | |
| "issue": "6", | |
| "pages": "82--97", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "G. Hinton, L. Deng, G.E. Dahl, A. Mohamed, N. Jaitly, A. Senior, V. Vanhoucke, P. Ngu- yen, T. Sainath, and B. Kingsbury, \"Deep neural networks for acoustic modeling in speech recognition,\" IEEE Signal Processing Magazine, vol. 29, no. 6, pp. 82-97, Nov. 2012.", | |
| "links": null | |
| }, | |
| "BIBREF3": { | |
| "ref_id": "b3", | |
| "title": "Practical recommendations for gradient-based ttraining of deep architectures", | |
| "authors": [ | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Bengio", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2013, | |
| "venue": "Neural Networks: Tricks of the Trade", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Y. Bengio, \"Practical recommendations for gradient-based ttraining of deep architectures,\" Neural Networks: Tricks of the Trade, K.R. Muller, G. Montavon, and G.B. Orr, eds., Springer 2013.", | |
| "links": null | |
| }, | |
| "BIBREF4": { | |
| "ref_id": "b4", | |
| "title": "Learning representations by backpropagating errors", | |
| "authors": [ | |
| { | |
| "first": "D", | |
| "middle": [ | |
| "E" | |
| ], | |
| "last": "Rumelhart", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [ | |
| "E" | |
| ], | |
| "last": "Hinton", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "R", | |
| "middle": [ | |
| "J" | |
| ], | |
| "last": "Williams", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1986, | |
| "venue": "Nature", | |
| "volume": "323", | |
| "issue": "", | |
| "pages": "533--536", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "D. E. Rumelhart, G.E. Hinton, and R. J. Williams, \"Learning representations by back- propagating errors,\" Nature, 1986, vol. 323, pp. 533-536.", | |
| "links": null | |
| }, | |
| "BIBREF5": { | |
| "ref_id": "b5", | |
| "title": "Applying convolutional neural networks concepts to hybrid NN-HMM model for speech recognition", | |
| "authors": [ | |
| { | |
| "first": "O", | |
| "middle": [], | |
| "last": "Hamid", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Mohamed", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Jiang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Penn", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2012, | |
| "venue": "Proc. ICASSP", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "4277--4280", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "O. Abdel-Hamid, A. Mohamed, H. Jiang, and G. Penn, \"Applying convolutional neural networks concepts to hybrid NN-HMM model for speech recognition,\" in Proc. ICASSP, pp. 4277-4280, 2012.", | |
| "links": null | |
| }, | |
| "BIBREF6": { | |
| "ref_id": "b6", | |
| "title": "Deep convolutional neural networks for LVCSR", | |
| "authors": [ | |
| { | |
| "first": "T", | |
| "middle": [ | |
| "N" | |
| ], | |
| "last": "Sainath", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Mohamed", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "B", | |
| "middle": [], | |
| "last": "Kingsbury", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "B", | |
| "middle": [], | |
| "last": "Ramabhadran", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2013, | |
| "venue": "Proc. ICASSP", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "8614--8618", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "T. N. Sainath, A. Mohamed, B. Kingsbury, and B. Ramabhadran, \"Deep convolutional neural networks for LVCSR,\" in Proc. ICASSP, pp. 8614-8618, 2013.", | |
| "links": null | |
| }, | |
| "BIBREF7": { | |
| "ref_id": "b7", | |
| "title": "Maximum F1-score discriminative training criterion for automatic mispronunciation detection", | |
| "authors": [ | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Huang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Xu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "X", | |
| "middle": [], | |
| "last": "Wang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "W", | |
| "middle": [], | |
| "last": "Silamu", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2015, | |
| "venue": "IEEE Trans. Audio, Speech, and Language Processing", | |
| "volume": "23", | |
| "issue": "5", | |
| "pages": "787--797", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "H. Huang, H. Xu, X. Wang, and W. Silamu, \"Maximum F1-score discriminative training criterion for automatic mispronunciation detection,\" IEEE Trans. Audio, Speech, and Language Processing, vol. 23, no. 5 pp. 787-797, April. 2015.", | |
| "links": null | |
| }, | |
| "BIBREF8": { | |
| "ref_id": "b8", | |
| "title": "Automatic detection of phone-level mispronunciation for language learning", | |
| "authors": [ | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Franco", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "L", | |
| "middle": [], | |
| "last": "Neumeyer", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Ramos", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Bratt", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1999, | |
| "venue": "Proc. Eurospeech", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "851--854", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "H. Franco, L. Neumeyer, M. Ramos, and H. Bratt, \"Automatic detection of phone-level mispronunciation for language learning,\" in Proc. Eurospeech, pp. 851-854, 1999.", | |
| "links": null | |
| }, | |
| "BIBREF9": { | |
| "ref_id": "b9", | |
| "title": "Phone-level pronunciation scoring and assessment for interactive language learning", | |
| "authors": [ | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Witt", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Young", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2000, | |
| "venue": "Speech Communication", | |
| "volume": "30", | |
| "issue": "2-3", | |
| "pages": "95--108", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "S. Witt and S. Young, \"Phone-level pronunciation scoring and assessment for interactive language learning,\" Speech Communication, vol. 30, no. 2-3, pp. 95-108, 2000.", | |
| "links": null | |
| }, | |
| "BIBREF10": { | |
| "ref_id": "b10", | |
| "title": "Automatic mispronunciation detection for Mandarin", | |
| "authors": [ | |
| { | |
| "first": "F", | |
| "middle": [], | |
| "last": "Zhang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "C", | |
| "middle": [], | |
| "last": "Huang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "F", | |
| "middle": [ | |
| "K" | |
| ], | |
| "last": "Soong", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Chu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "R", | |
| "middle": [ | |
| "H" | |
| ], | |
| "last": "Wang", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2008, | |
| "venue": "Proc. ICASSP", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "5077--5080", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "F. Zhang, C. Huang, F. K. Soong, M. Chu, and R. H. Wang, \"Automatic mispronunciation detection for Mandarin,\" in Proc. ICASSP, pp. 5077-5080, 2008.", | |
| "links": null | |
| }, | |
| "BIBREF11": { | |
| "ref_id": "b11", | |
| "title": "Improved approaches of modeling and detecting error patterns with empirical analysis for computer-aided pronunciation training", | |
| "authors": [ | |
| { | |
| "first": "Y", | |
| "middle": [ | |
| "B" | |
| ], | |
| "last": "Wang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "L", | |
| "middle": [ | |
| "S" | |
| ], | |
| "last": "Lee", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2012, | |
| "venue": "Proc. ICASSP", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "5049--5052", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Y.B. Wang and L.S. Lee, \"Improved approaches of modeling and detecting error patterns with empirical analysis for computer-aided pronunciation training,\" in Proc. ICASSP, pp. 5049-5052, 2012.", | |
| "links": null | |
| }, | |
| "BIBREF12": { | |
| "ref_id": "b12", | |
| "title": "Pronunciation error detection method based on error rule clustering using a decision tree", | |
| "authors": [ | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Ito", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Lim", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Suzuki", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Makino", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2005, | |
| "venue": "Proc. EuroSpeech", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "173--176", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Ito, A., Lim, Y., Suzuki, M., Makino, S., \"Pronunciation error detection method based on error rule clustering using a decision tree\", in Proc. EuroSpeech, pp. 173-176, 2005.", | |
| "links": null | |
| }, | |
| "BIBREF13": { | |
| "ref_id": "b13", | |
| "title": "Automatic pronunciation error detection: an acoustic-phonetic approach", | |
| "authors": [ | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Truong", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Neri", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "C", | |
| "middle": [], | |
| "last": "Cuchiarini", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Strik", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2004, | |
| "venue": "Proc. of the InSTIL/ICALL Symposium", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "135--138", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "K. Truong, A. Neri, C. Cuchiarini, and H. Strik, \"Automatic pronunciation error detection: an acoustic-phonetic approach,\" in Proc. of the InSTIL/ICALL Symposium, pp. 135-138, 2004.", | |
| "links": null | |
| }, | |
| "BIBREF14": { | |
| "ref_id": "b14", | |
| "title": "A new method for mispronunciation detection using support vector machine based on pronunciation space models", | |
| "authors": [ | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Wei", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Hu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Hu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "R", | |
| "middle": [ | |
| "H" | |
| ], | |
| "last": "Wang", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2009, | |
| "venue": "Speech Communication", | |
| "volume": "51", | |
| "issue": "10", | |
| "pages": "896--905", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "S. Wei, G. Hu, Y. Hu, and R. H. Wang, \"A new method for mispronunciation detection using support vector machine based on pronunciation space models,\" Speech Communi- cation., vol. 51, no. 10, pp. 896-905, 2009.", | |
| "links": null | |
| }, | |
| "BIBREF15": { | |
| "ref_id": "b15", | |
| "title": "Improved mispronunciation detection with deep neural network trained acoustic models and transfer learning based logistic regression classifiers", | |
| "authors": [ | |
| { | |
| "first": "W", | |
| "middle": [], | |
| "last": "Hu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Qian", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "F", | |
| "middle": [ | |
| "K" | |
| ], | |
| "last": "Soong", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Wang", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2015, | |
| "venue": "Speech Communication", | |
| "volume": "67", | |
| "issue": "", | |
| "pages": "154--166", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "W. Hu, Y. Qian, F. K. Soong, and Y. Wang, \"Improved mispronunciation detection with deep neural network trained acoustic models and transfer learning based logistic regres- sion classifiers,\" Speech Communication, vol. 67, pp. 154-166, 2015.", | |
| "links": null | |
| }, | |
| "BIBREF16": { | |
| "ref_id": "b16", | |
| "title": "Automatic text-independent pronunciation scoring of foreign language student speech", | |
| "authors": [ | |
| { | |
| "first": "L", | |
| "middle": [], | |
| "last": "Neumeyer", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Franco", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Weintraub", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "P", | |
| "middle": [], | |
| "last": "Price", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1996, | |
| "venue": "Proc. Int. Conf. Spoken Lang. Process", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "1457--1460", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "L. Neumeyer, H. Franco, M. Weintraub, and P. Price, \"Automatic text-independent pro- nunciation scoring of foreign language student speech,\" in Proc. Int. Conf. Spoken Lang. Process., pp. 1457-1460, 1996.", | |
| "links": null | |
| }, | |
| "BIBREF17": { | |
| "ref_id": "b17", | |
| "title": "Automatic pronunciation scoring using learning to rank and DP-based score segmentation", | |
| "authors": [ | |
| { | |
| "first": "L", | |
| "middle": [ | |
| "Y" | |
| ], | |
| "last": "Chen", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [ | |
| "S R" | |
| ], | |
| "last": "Jang", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2010, | |
| "venue": "Proc. Interspeech", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "761--764", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "L. Y. Chen and J. S. R. Jang, \"Automatic pronunciation scoring using learning to rank and DP-based score segmentation,\" in Proc. Interspeech, pp. 761-764, 2010.", | |
| "links": null | |
| }, | |
| "BIBREF18": { | |
| "ref_id": "b18", | |
| "title": "Improvement in automatic pronunciation scoring using additional basic scores and learning to rank", | |
| "authors": [ | |
| { | |
| "first": "L", | |
| "middle": [ | |
| "Y" | |
| ], | |
| "last": "Chen", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [ | |
| "S R" | |
| ], | |
| "last": "Jang", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2012, | |
| "venue": "Proc. Interspeech", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "L. Y. Chen and J. S. R. Jang, \"Improvement in automatic pronunciation scoring using additional basic scores and learning to rank,\" in Proc. Interspeech, 2012.", | |
| "links": null | |
| }, | |
| "BIBREF19": { | |
| "ref_id": "b19", | |
| "title": "Automatic pronunciation scoring with score combination by learning to rank and class-normalized DP-based quantization", | |
| "authors": [ | |
| { | |
| "first": "L", | |
| "middle": [ | |
| "Y" | |
| ], | |
| "last": "Chen", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [ | |
| "S R" | |
| ], | |
| "last": "Jang", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2015, | |
| "venue": "IEEE Trans. Audio, Speech, and Language Processing", | |
| "volume": "23", | |
| "issue": "11", | |
| "pages": "787--797", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "L. Y. Chen and J. S. R. Jang, \"Automatic pronunciation scoring with score combination by learning to rank and class-normalized DP-based quantization,\" IEEE Trans. Audio, Speech, and Language Processing, vol. 23, no. 11 pp. 787-797, November. 2015.", | |
| "links": null | |
| }, | |
| "BIBREF20": { | |
| "ref_id": "b20", | |
| "title": "An application of pretrained deep neural networks to large vocabulary speech recognition", | |
| "authors": [ | |
| { | |
| "first": "N", | |
| "middle": [], | |
| "last": "Jaitly", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "P", | |
| "middle": [], | |
| "last": "Nguyen", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Senior", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "V", | |
| "middle": [], | |
| "last": "Vanhoucke", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": null, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "N. Jaitly, P. Nguyen, A. Senior, and V. Vanhoucke, \"An application of pretrained deep neural networks to large vocabulary speech recognition,\" submitted for publication.", | |
| "links": null | |
| }, | |
| "BIBREF21": { | |
| "ref_id": "b21", | |
| "title": "New types of deep neural network learning for speech recognition and related applications: An overview", | |
| "authors": [ | |
| { | |
| "first": "L", | |
| "middle": [], | |
| "last": "Deng", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Hinton", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "B", | |
| "middle": [], | |
| "last": "Kingsbury", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2013, | |
| "venue": "Proc. ICASSP", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "L. Deng, G. Hinton, and B. Kingsbury, \"New types of deep neural network learning for speech recognition and related applications: An overview,\" in Proc. ICASSP, 2013.", | |
| "links": null | |
| }, | |
| "BIBREF22": { | |
| "ref_id": "b22", | |
| "title": "The use of DBN-HMMs for mispronunciation detection and diagnosis in L2 English to support computer-aided pronunciation training", | |
| "authors": [ | |
| { | |
| "first": "X", | |
| "middle": [], | |
| "last": "Qian", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Meng", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "F", | |
| "middle": [], | |
| "last": "Soong", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2012, | |
| "venue": "Proc. Interspeech", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "775--778", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "X. Qian, H. Meng, and F. Soong, \"The use of DBN-HMMs for mispronunciation detec- tion and diagnosis in L2 English to support computer-aided pronunciation training,\" in Proc. Interspeech, pp. 775-778, 2012.", | |
| "links": null | |
| }, | |
| "BIBREF23": { | |
| "ref_id": "b23", | |
| "title": "Acoustic model optimization for automatic pronunciation quality assessment", | |
| "authors": [ | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Ke", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2014, | |
| "venue": "Proc. ICMFI", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Y. Ke. \"Acoustic model optimization for automatic pronunciation quality assessment,\" in Proc. ICMFI, 2014.", | |
| "links": null | |
| }, | |
| "BIBREF24": { | |
| "ref_id": "b24", | |
| "title": "ImageNet classification with deep convolutional neural networks", | |
| "authors": [ | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Krizhevsky", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "I", | |
| "middle": [], | |
| "last": "Sutskever", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Hinton", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2012, | |
| "venue": "Proc. Neural Information and Processing Systems", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "A. Krizhevsky, I. Sutskever, and G. Hinton, \"ImageNet classification with deep convo- lutional neural networks,\" Proc. Neural Information and Processing Systems, 2012.", | |
| "links": null | |
| }, | |
| "BIBREF25": { | |
| "ref_id": "b25", | |
| "title": "Convolutional neural networks for speech recognition", | |
| "authors": [ | |
| { | |
| "first": "O", | |
| "middle": [], | |
| "last": "Hamid", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [ | |
| "R" | |
| ], | |
| "last": "Mohamed", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Jiang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "L", | |
| "middle": [], | |
| "last": "Deng", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Penn", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Yu", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2014, | |
| "venue": "IEEE/ACM Trans. Audio, Speech, Lang. Process", | |
| "volume": "22", | |
| "issue": "10", | |
| "pages": "1533--1545", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "O. Abdel-Hamid, A.R. Mohamed, H. Jiang, L. Deng, G. Penn, and D. Yu, \"Convolutional neural networks for speech recognition,\" IEEE/ACM Trans. Audio, Speech, Lang. Pro- cess., vol. 22, no. 10, pp. 1533-1545, Oct. 2014.", | |
| "links": null | |
| }, | |
| "BIBREF26": { | |
| "ref_id": "b26", | |
| "title": "Convolutional networks for images, speech, and time series", | |
| "authors": [ | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Le Cun", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Bengio", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1995, | |
| "venue": "The Handbook of Brain Theory and Neural Networks, M. A. Arbib", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "255--258", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Y. Le Cun and Y. Bengio, \"Convolutional networks for images, speech, and time series,\" in The Handbook of Brain Theory and Neural Networks, M. A. Arbib, Ed. Cambridge, MA: MIT Press, pp. 255-258, 1995.", | |
| "links": null | |
| }, | |
| "BIBREF27": { | |
| "ref_id": "b27", | |
| "title": "Exploring hierarchical speech representations using a deep convolutional neural network", | |
| "authors": [ | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Hau", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Chen", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2011, | |
| "venue": "Proc. UKCI", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "D. Hau and K. Chen, \"Exploring hierarchical speech representations using a deep con- volutional neural network,\" in Proc. UKCI, 2011.", | |
| "links": null | |
| }, | |
| "BIBREF28": { | |
| "ref_id": "b28", | |
| "title": "Automatic speech recognition -a deep learning approach", | |
| "authors": [ | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Yu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "L", | |
| "middle": [], | |
| "last": "Deng", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2014, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "D. Yu and L. Deng, \"Automatic speech recognition -a deep learning approach\", Springer, 2014.", | |
| "links": null | |
| }, | |
| "BIBREF29": { | |
| "ref_id": "b29", | |
| "title": "Deep neural networks for acoustic modeling in speech recognition", | |
| "authors": [ | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Hinton", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "L", | |
| "middle": [], | |
| "last": "Deng", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [ | |
| "E" | |
| ], | |
| "last": "Dahl", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Mohamed", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "N", | |
| "middle": [], | |
| "last": "Jaitly", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Senior", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "V", | |
| "middle": [], | |
| "last": "Vanhoucke", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "P", | |
| "middle": [], | |
| "last": "Nguyen", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Sainath", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "B", | |
| "middle": [], | |
| "last": "Kingsbury", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2012, | |
| "venue": "IEEE Signal Processing Magazine", | |
| "volume": "29", | |
| "issue": "6", | |
| "pages": "82--97", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "G. Hinton, L. Deng, G.E. Dahl, A. Mohamed, N. Jaitly, A. Senior, V. Vanhoucke, P. Ngu- yen, T. Sainath, and B. Kingsbury, \"Deep neural networks for acoustic modeling in speech recognition,\" IEEE Signal Processing Magazine, vol. 29, no. 6, pp. 82-97, Nov. 2012.", | |
| "links": null | |
| }, | |
| "BIBREF30": { | |
| "ref_id": "b30", | |
| "title": "Phone recognition with the mean-covariance restricted boltzmann machine", | |
| "authors": [ | |
| { | |
| "first": "G", | |
| "middle": [ | |
| "E" | |
| ], | |
| "last": "Dahl", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Ranzato", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Mohamed", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [ | |
| "E" | |
| ], | |
| "last": "Hinton", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2010, | |
| "venue": "Advances in Neural Information Processing Systems", | |
| "volume": "23", | |
| "issue": "", | |
| "pages": "469--477", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "G. E. Dahl, M. Ranzato, A. Mohamed, and G. E. Hinton, \"Phone recognition with the mean-covariance restricted boltzmann machine,\" in Advances in Neural Information Processing Systems 23, J. Lafferty, C. K. I. Williams, J. ShaweTaylor, R.S. Zemel, and A. Culotta, Eds. Cambridge, MA: MIT Press, pp. 469-477, 2010.", | |
| "links": null | |
| }, | |
| "BIBREF31": { | |
| "ref_id": "b31", | |
| "title": "Deep boltzmann machines", | |
| "authors": [ | |
| { | |
| "first": "R", | |
| "middle": [], | |
| "last": "Salakhutdinov", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [ | |
| "E" | |
| ], | |
| "last": "Hinton", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2009, | |
| "venue": "Proc. AISTATS", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "448--455", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "R. Salakhutdinov and G.E. Hinton, \"Deep boltzmann machines,\" in Proc. AISTATS, pp. 448-455, 2009.", | |
| "links": null | |
| }, | |
| "BIBREF32": { | |
| "ref_id": "b32", | |
| "title": "Unsupervised feature learning for audio classification using convolutional deep belief networks", | |
| "authors": [ | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Lee", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "P", | |
| "middle": [], | |
| "last": "Pham", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Largman", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Ng", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2009, | |
| "venue": "Advances in Neural Information Processing Systems", | |
| "volume": "22", | |
| "issue": "", | |
| "pages": "1096--1104", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "H. Lee, P. Pham, Y. Largman, and A. Ng, \"Unsupervised feature learning for audio clas- sification using convolutional deep belief networks,\" in Advances in Neural Information Processing Systems 22, Y. Bengio, D. Schuurmans, J. Lafferty, C. K. I. Williams, and A. Culotta, Eds. Cambridge, MA: MIT Press, pp. 1096-1104, 2009.", | |
| "links": null | |
| }, | |
| "BIBREF33": { | |
| "ref_id": "b33", | |
| "title": "Understanding how deep belief networks perform acoustic modelling", | |
| "authors": [ | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Mohamed", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Hinton", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Penn", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2012, | |
| "venue": "Proc. ICASSP", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "4273--4276", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "A. Mohamed, G. Hinton, and G. Penn, \"Understanding how deep belief networks per- form acoustic modelling,\" in Proc. ICASSP, pp. 4273-4276, 2012.", | |
| "links": null | |
| }, | |
| "BIBREF34": { | |
| "ref_id": "b34", | |
| "title": "Context-dependent pretrained deep neural networks for large-vocabulary speech recognition", | |
| "authors": [ | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Dahl", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Yu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "L", | |
| "middle": [], | |
| "last": "Deng", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Acero", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2012, | |
| "venue": "IEEE Trans. Audio Speech Lang. Processing", | |
| "volume": "20", | |
| "issue": "1", | |
| "pages": "30--42", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "G. Dahl, D. Yu, L. Deng, and A. Acero, \"Context-dependent pretrained deep neural net- works for large-vocabulary speech recognition,\" IEEE Trans. Audio Speech Lang. Pro- cessing, vol. 20, no. 1, pp. 30-42, Jan. 2012.", | |
| "links": null | |
| }, | |
| "BIBREF35": { | |
| "ref_id": "b35", | |
| "title": "Deep belief networks using discriminative features for phone recognition", | |
| "authors": [ | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Mohamed", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "T", | |
| "middle": [ | |
| "N" | |
| ], | |
| "last": "Sainath", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [ | |
| "E" | |
| ], | |
| "last": "Dahl", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "B", | |
| "middle": [], | |
| "last": "Ramabhadran", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [ | |
| "E" | |
| ], | |
| "last": "Hinton", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Picheny", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2011, | |
| "venue": "Proc. ICASSP", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "5060--5063", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "A. Mohamed, T. N. Sainath, G. E. Dahl, B. Ramabhadran, G. E. Hinton, and M. Picheny, \"Deep belief networks using discriminative features for phone recognition,\" in Proc. ICASSP, pp. 5060-5063, 2011.", | |
| "links": null | |
| }, | |
| "BIBREF36": { | |
| "ref_id": "b36", | |
| "title": "Receptive fields, binocular in teraction, and functional architecture in the cat's visual cortex", | |
| "authors": [ | |
| { | |
| "first": "D", | |
| "middle": [ | |
| "H" | |
| ], | |
| "last": "Hubel", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "T", | |
| "middle": [ | |
| "N" | |
| ], | |
| "last": "Wiesel", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1962, | |
| "venue": "J. Physiology (London)", | |
| "volume": "160", | |
| "issue": "", | |
| "pages": "106--154", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "D. H. Hubel and T. N. Wiesel, \"Receptive fields, binocular in teraction, and functional architecture in the cat's visual cortex,\" J. Physiology (London), vol. 160, pp. 106-154, 1962.", | |
| "links": null | |
| }, | |
| "BIBREF37": { | |
| "ref_id": "b37", | |
| "title": "Notes on convolutional neural networks", | |
| "authors": [ | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Bouvrie", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2006, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "J. Bouvrie, \"Notes on convolutional neural networks,\" 2006.", | |
| "links": null | |
| }, | |
| "BIBREF38": { | |
| "ref_id": "b38", | |
| "title": "Evaluation of pooling operations in convolutional architectures for object recognition", | |
| "authors": [ | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Scherer", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Muller", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Behnke", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2010, | |
| "venue": "Proc. ICANN", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "92--101", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "D. Scherer, A. Muller, and S. Behnke, \"Evaluation of pooling operations in convolutional architectures for object recognition,\" in Proc. ICANN, pp. 92-101, 2010.", | |
| "links": null | |
| }, | |
| "BIBREF39": { | |
| "ref_id": "b39", | |
| "title": "Python -scikit-learn", | |
| "authors": [], | |
| "year": null, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Python -scikit-learn. http://scikit-learn.org/dev/index.html", | |
| "links": null | |
| }, | |
| "BIBREF40": { | |
| "ref_id": "b40", | |
| "title": "Discriminatively trained acoustic model for improving mispronunciation detection and diagnosis in computer-aided pronunciation training (CAPT)", | |
| "authors": [ | |
| { | |
| "first": "X", | |
| "middle": [], | |
| "last": "Qian", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [ | |
| "M" | |
| ], | |
| "last": "Meng", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "F", | |
| "middle": [ | |
| "K" | |
| ], | |
| "last": "Soong", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2010, | |
| "venue": "Proc. Interspeech", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "X. Qian, H. M. Meng, and F. K. Soong, \"Discriminatively trained acoustic model for improving mispronunciation detection and diagnosis in computer-aided pronunciation training (CAPT),\" in Proc. Interspeech, 2010.", | |
| "links": null | |
| }, | |
| "BIBREF41": { | |
| "ref_id": "b41", | |
| "title": "CDF-matching for automatic tone error detection in Mandarin CALL system", | |
| "authors": [ | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Wei", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Wang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Q", | |
| "middle": [], | |
| "last": "Liu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "R", | |
| "middle": [], | |
| "last": "Wang", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2007, | |
| "venue": "Proc. ICASSP", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "205--208", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "S. Wei, H. Wang, Q. Liu, and R. Wang, \"CDF-matching for automatic tone error detection in Mandarin CALL system,\" in Proc. ICASSP, pp. 205-208, 2007.", | |
| "links": null | |
| }, | |
| "BIBREF42": { | |
| "ref_id": "b42", | |
| "title": "Automatic tone assessment of non-native Mandarin speakers", | |
| "authors": [ | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Cheng", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2013, | |
| "venue": "Proc. Interspeech", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "1299--1302", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "J. Cheng, \"Automatic tone assessment of non-native Mandarin speakers,\" in Proc. Inter- speech, pp. 1299-1302, 2013.", | |
| "links": null | |
| }, | |
| "BIBREF43": { | |
| "ref_id": "b43", | |
| "title": "The kaldi speech recognition toolkit", | |
| "authors": [ | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Povey", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Ghoshal", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Boulianne", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "L", | |
| "middle": [], | |
| "last": "Burget", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "O", | |
| "middle": [], | |
| "last": "Glembek", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "N", | |
| "middle": [], | |
| "last": "Goel", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Hannemann", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "P", | |
| "middle": [], | |
| "last": "Motl\u00b4\u0131cek", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Qian", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "P", | |
| "middle": [], | |
| "last": "Schwarz", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Silovsky", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Stemmer", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Vesely", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2011, | |
| "venue": "Proc. IEEE ASRU", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "D. Povey, A. Ghoshal, G. Boulianne, L. Burget, O. Glembek, N. Goel, M. Hannemann, P. Motl\u00b4\u0131cek, Y. Qian, P. Schwarz, J. Silovsky, G. Stemmer, and K. Vesely, \"The kaldi speech recognition toolkit,\" in Proc. IEEE ASRU, 2011.", | |
| "links": null | |
| } | |
| }, | |
| "ref_entries": { | |
| "TABREF0": { | |
| "num": null, | |
| "type_str": "table", | |
| "html": null, | |
| "content": "<table><tr><td>\u8f38\u51fa\u767c\u97f3\u6aa2\u6e2c\u5206\u6578\u5c0d\u61c9\u4e4b\u6392\u5e8f\u53d6\u8abf\u548c\u5e73\u5747\u505a\u70ba\u7d50\u5408\u5f8c\u7684\u5206\u985e\u7d50\u679c\uff1b\u6700\u5f8c\uff0c\u5728\u7b2c\u516d\u5c0f\u7bc0\uff0c</td></tr><tr><td>\u6211\u5011\u63d0\u51fa\u7d50\u8ad6\u8207\u4e00\u4e9b\u672a\u4f86\u53ef\u80fd\u7684\u7814\u7a76\u65b9\u5411\u3002</td></tr><tr><td>\u4e8c\u3001 \u76f8\u95dc\u7814\u7a76</td></tr><tr><td>\u4e00\u3001 \u7dd2\u8ad6 \u73fe\u4eca\u5168\u7403\u5316\u7684\u6642\u4ee3\u88e1\uff0c\u7cbe\u901a\u5169\u7a2e\u6216\u5169\u7a2e\u4ee5\u4e0a\u7684\u8a9e\u8a00\u4e0d\u50c5\u662f\u512a\u52e2\u66f4\u662f\u5fc5\u8981\u7684\u80fd\u529b\u3002\u5728 \u5341\u5e7e\u5e74\u4ee5\u524d\uff0c\u82f1\u8a9e\u9084\u662f\u570b\u969b\u901a\u7528\u7684\u8a9e\u8a00\uff1b\u4f46\u8fd1\u5e74\u4f86\uff0c\u7531\u65bc\u4e2d\u570b\u5e02\u5834\u7684\u5feb\u901f\u767c\u5c55\uff0c\u5168\u7403\u83ef \u8a9e\u5b78\u7fd2\u71b1\u6f6e\u5e2d\u6372\u800c\u4f86\uff0c\u5b78\u7fd2\u83ef\u8a9e\u7684\u4eba\u6578\u9810\u4f30\u5df2\u7d93\u8d85\u904e\u4e00\u5104\uff0c\u5728\u8a31\u591a\u975e\u83ef\u8a9e\u8a9e\u7cfb\u7684\u4e9e\u6d32\u3001 \u6b50\u6d32\u4ee5\u53ca\u7f8e\u6d32\u570b\u5bb6\uff0c\u83ef\u8a71\u5df2\u7d93\u9010\u6f38\u6210\u70ba\u4e00\u7a2e\u5fc5\u9808\u5b78\u7fd2\u7684\u8a9e\u8a00[1][2]\u3002\u8a9e\u8a00\u5b78\u7fd2\u53c8\u5206\u70ba\u807d (listening)\u800c\u672c\u7bc7\u8ad6\u6587\u5c07\u805a\u7126\u5728\u5982\u4f55\u6539\u5584\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c\u4e4b\u6548\u80fd\u3002\u76ee\u524d\uff0c\u5728\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c\u7684\u8a55\u4f30\u65b9\u5f0f \u5728\u5927\u591a\u6578\u7684\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c\u7814\u7a76\u4e2d\u5e7e\u4e4e\u90fd\u662f\u4ee5\u81ea\u52d5\u8a9e\u97f3\u8fa8\u8b58\u70ba\u524d\u7aef\uff0c\u800c\u5c07\u5f8c\u7aef\u8996\u70ba\u5206 \u4e2d\uff0c\u53ec\u56de\u7387(recall)\u548c\u7cbe\u6e96\u5ea6(precision)\u7684\u66f2\u7dda\u8207\u63a5\u6536\u8005\u64cd\u4f5c\u7279\u5fb5\u66f2\u7dda(receiver operating characteristic curve, ROC)\u662f\u6700\u5e38\u88ab\u63a1\u7528\u4f86\u8a55\u4f30\u6548\u80fd\u4e4b\u512a\u52a3\u3002\u6211\u5011\u8a8d\u70ba\u76f8\u8f03\u65bc\u6b63\u78ba\u767c\u97f3\u6aa2 \u6e2c(correct pronunciation detection)\uff0c\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c\u5c0d\u65bc\u5b78\u7fd2\u8005\u800c\u8a00\u662f\u8f03\u70ba\u91cd\u8981\uff1b\u6240\u4ee5\uff0c\u672c \u7bc7\u8ad6\u6587\u5f8c\u7e8c\u5728\u53ec\u56de\u7387\u548c\u7cbe\u6e96\u5ea6\u66f2\u7dda\u7684\u8a55\u4f30\u5be6\u9a57\u4e2d\uff0c\u6211\u5011\u5c07\u96c6\u4e2d\u8a0e\u8ad6\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c\u7684\u6548\u80fd \u8868\u73fe\u3002 \u81ea\u52d5\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c\u7684\u7814\u7a76\u5927\u90e8\u5206\u662f\u57fa\u65bc\u73fe\u6709\u7684\u8a9e\u97f3\u8fa8\u8b58\u6280\u8853\u800c\u767c\u5c55\uff0c\u5e0c\u671b\u80fd\u9054\u5230\u50cf \u5c08\u696d\u8a9e\u8a00\u6559\u5e2b\u4e00\u6a23\u5730\u7d66\u4e88\u8a9e\u8a00\u5b78\u7fd2\u8005\u6240\u5ff5\u8aa6\u8a9e\u53e5\u9069\u7576\u7684\u767c\u97f3\u8a55\u4f30\u3002\u5728\u672c\u8ad6\u6587\u4e2d\uff0c\u6211\u5011\u5c07 \u8a9e\u97f3\u8fa8\u8b58\u6a21\u7d44\u8996\u70ba\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c\u7cfb\u7d71\u7684\u524d\u7aef(front-end)\uff0c\u800c \u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c(\u5206\u985e)\u6a21\u7d44\u8996\u70ba \u7cfb\u7d71\u7684\u5f8c\u7aef(back-end)\u3002\u524d\u7aef\u7684\u8a9e\u97f3\u8fa8\u8b58\u6a21\u7d44\u5982\u679c\u80fd\u85c9\u7531\u8072\u5b78\u6a21\u578b\u7684\u4f7f\u7528\uff0c\u7522\u751f\u97f3\u6846 (frame)\u6216\u8005\u6bb5\u843d(segment)\u5c64\u6b21\u7684\u4e8b\u5f8c\u6a5f\u7387\u4f86\u505a\u70ba\u5177\u9451\u5225\u6027\u7684\u767c\u97f3\u6aa2\u6e2c\u7279\u5fb5\uff0c\u5247\u5f8c\u7aef\u5075\u6e2c \u932f\u8aa4\u767c\u97f3\u6642\u5c31\u80fd\u57fa\u65bc\u9019\u4e9b\u767c\u97f3\u6aa2\u6e2c\u7279\u5fb5\u4f86\u7cbe\u6e96\u5730\u5224\u65b7\u5b78\u7fd2\u8005\u7684\u767c\u97f3\u6b63\u78ba\u8207\u5426\u3002\u56e0\u6b64\uff0c\u8a9e \u97f3\u8fa8\u8b58\u6a21\u7d44\u4e2d\u8072\u5b78\u6a21\u578b\u6240\u7522\u751f\u7684\u56de\u994b\u5c07\u662f\u6211\u5011\u8a55\u65b7\u767c\u97f3\u597d\u58de\u8207\u5426\u7684\u91cd\u8981\u4f9d\u64da\u3002\u5728\u8a9e\u97f3\u8fa8 \u8b58\u7814\u7a76\u4e0a\uff0c\u6709\u5225\u65bc\u50b3\u7d71\u4f7f\u7528\u6885\u723e\u5012\u983b\u8b5c\u4fc2\u6578(mel-frequency cepstral coefficients, MFCC)\u4e4b \u8a9e\u97f3\u7279\u5fb5\u7684\u9ad8\u65af\u6df7\u5408\u6a21\u578b-\u96b1\u85cf\u5f0f\u99ac\u53ef\u592b\u6a21\u578b(gaussian mixture model-hidden markov model, GMM-HMM)\u7684\u8072\u5b78\u6a21\u578b\uff0c\u8fd1\u5e74\u4f86\u7531\u65bc\u6a5f\u5668\u5b78\u7fd2\u6f14\u7b97\u6cd5[3][4][5]\u8207\u96fb\u8166\u786c\u9ad4\u7684\u9032 \u6b65\uff0c\u8a13\u7df4\u591a\u96b1\u85cf\u5c64(hidden layers)\u53ca\u5927\u91cf\u8f38\u51fa\u795e\u7d93\u5143(neurons)\u985e\u795e\u7d93\u7db2\u8def\u7684\u65b9\u6cd5\u4e5f\u66f4\u6709\u6548 \u7387\u5728\u5b78\u8853\u754c\u8207\u5be6\u52d9\u754c\u6fc0\u8d77\u4e86\u6df1\u5c64\u5b78\u7fd2(deep learning)\u7684\u6d6a\u6f6e\uff0c\u985b\u8986\u4e86\u5e7e\u5341\u5e74\u4f86\u7684\u7814\u7a76\u751f \u614b\u3002\u8a31\u591a\u5b78\u8005\u8207\u5be6\u52d9\u5bb6\u7814\u7a76\u5c07\u6df1\u5c64\u985e\u795e\u7d93\u7db2\u8def(deep neural networks, DNN)\u7576\u4f5c\u8a9e\u97f3\u8fa8\u8b58 \u7684\u8072\u5b78\u6a21\u578b\u7684\u91cd\u8981\u7d44\u6210\uff0c\u53d6\u4ee3\u50b3\u7d71 GMM \u7684\u89d2\u8272\u4f86\u8a08\u7b97\u6bcf\u500b\u97f3\u6846\u6240\u5c0d\u61c9 HMM \u72c0\u614b\u7684\u89c0 \u6e2c\u6a5f\u7387(observation probability)\u6216\u76f8\u4f3c\u5ea6\u503c(likelihood)\u3002\u96d6\u7136 DNN \u5728\u8a9e\u97f3\u8fa8\u8b58\u9818\u57df\u5df2\u7d93 \u6709\u76f8\u7576\u512a\u7570\u7684\u6548\u679c\uff0c\u4f46\u4e5f\u6709\u8a31\u591a\u7814\u7a76\u6307\u51fa\u647a\u7a4d\u985e\u795e\u7d93\u7db2\u8def(convolutional neuron networks, CNN)\u5728\u97f3\u7d20\u8fa8\u8b58[6]\u4ee5\u53ca\u5927\u8a5e\u5f59\u9023\u7e8c\u8a9e\u97f3\u8fa8\u8b58[7]\u7684\u4efb\u52d9\u4e0a\u7684\u8868\u73fe\u66f4\u512a\u65bc DNN\uff1b\u9019\u53ef\u6b78 \u529f\u65bc CNN \u80fd\u5f9e\u8a9e\u97f3\u7279\u5fb5\u4e2d\u64f7\u53d6\u51fa\u767c\u97f3\u4e2d\u7d30\u5fae\u7684\u4f4d\u79fb\u4e0d\u8b8a(shift invariance)\u7684\u7279\u6027\u3002\u900f\u904e CNN \u4f86\u505a\u70ba\u767c\u97f3\u6aa2\u6e2c\u7279\u5fb5\u7684\u64f7\u53d6\u6a21\u7d44\uff0c\u671f\u671b\u80fd\u5920\u5f9e\u4e0d\u540c\u570b\u5bb6\u7684\u83ef\u8a9e\u5b78\u7fd2\u8005\u4e4b\u767c\u97f3\u8a0a\u865f \u4e2d\u6c42\u53d6\u51fa\u5c0d\u767c\u97f3\u6aa2\u6e2c\u6709\u5e6b\u52a9\u3001\u5177\u9451\u5225\u6027\u7684\u767c\u97f3\u6aa2\u6e2c\u7279\u5fb5(\u80fd\u63d0\u4f9b\u66f4\u5177\u9451\u5225\u529b\u7684\u4e8b\u5f8c\u6a5f\u7387 \u4f86\u5e6b\u52a9\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c)\uff0c\u63d0\u5347\u81ea\u52d5\u6aa2\u6e2c\u932f\u8aa4\u767c\u97f3\u7684\u80fd\u529b\u3002\u672c\u7bc7\u8ad6\u6587\u5c0d\u65bc\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c\u7814 \u7a76\u6709\u4e09\u9805\u4e3b\u8981\u8ca2\u737b\uff1a\u9996\u5148\uff0c\u6211\u5011\u6bd4\u8f03\u4e26\u7d50\u5408\u7576\u524d\u57fa\u65bc\u6df1\u5c64\u985e\u795e\u7d93\u7db2\u8def(DNN)\u8207\u647a\u7a4d\u985e\u795e \u7d93\u7db2\u8def(CNN)\u4e4b\u5148\u9032\u7684\u8072\u5b78\u6a21\u578b\u4ee5\u7522\u751f\u66f4\u5177\u9451\u5225\u6027\u767c\u97f3\u6aa2\u6e2c\u7279\u5fb5\uff1b\u518d\u8005\uff0c\u6211\u5011\u6bd4\u8f03\u4e26\u7d50 \u5408\u4e0d\u540c\u5206\u985e\u65b9\u6cd5\uff0c\u4ee5\u671f\u80fd\u9054\u5230\u66f4\u4f73\u7684\u767c\u97f3\u6aa2\u6e2c\u8868\u73fe\uff1b\u6700\u5f8c\uff0c\u91dd\u5c0d\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c\u4e4b\u69cb\u6210\u6a21 \u7d44\uff0c\u9032\u884c\u4e00\u7cfb\u5217\u5ee3\u6cdb\u4e14\u6df1\u5165\u7684\u5be6\u9a57\u5206\u6790\u8207\u8a0e\u8ad6\u3002 \u985e\u554f\u984c[8]\u3002\u4f8b\u5982\uff0cFranco \u7b49\u4eba[9]\u4f7f\u7528\u6bcd\u8a9e\u8005\u7684 HMM \u4e4b\u5c0d\u6578\u76f8\u4f3c\u5ea6\u503c(log-likelihood)\u8207 \u975e\u6bcd\u8a9e\u8005\u7684 HMM \u4e4b\u5c0d\u6578\u76f8\u4f3c\u5ea6\u503c\u8a08\u7b97\u6bd4\u503c\uff0c\u7a31\u70ba\u5c0d\u6578\u76f8\u4f3c\u5ea6\u6bd4\u503c(log-likelihood ratio, LLR)\uff0c\u8a72\u8ad6\u6587\u7684\u5be6\u9a57\u986f\u793a\u4f7f\u7528\u5c0d\u6578\u76f8\u4f3c\u5ea6\u6bd4\u503c(LLR)\u5c0d\u65bc\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c\u4e4b\u8868\u73fe\u52dd\u904e\u76f4\u63a5 \u4f7f\u7528\u5c0d\u6578\u76f8\u4f3c\u5ea6\u503c\u3002Witt \u7b49\u4eba[10]\u63d0\u51fa GOP \u4f5c\u70ba\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c\u4e4b\u8a55\u4f30\u65b9\u5f0f\uff0c\u8a72\u65b9\u6cd5\u57fa \u65bc\u8072\u5b78\u6a21\u578b\u6240\u7522\u751f\u7684\u4e8b\u5f8c\u6a5f\u7387(posterior probability)\u5c0d\u97f3\u7d20\u5c64\u6b21\u7684\u767c\u97f3\u8a08\u7b97\u8a55\u4f30\u5206\u6578\uff0c\u4e26 \u8a02\u5b9a\u9580\u6abb\u503c(threshold)\u4f86\u5340\u5206\u6b63\u78ba\u767c\u97f3\u8207\u932f\u8aa4\u767c\u97f3\uff1b\u9678\u7e8c\u4e5f\u6709\u5176\u5b83\u7814\u7a76\u662f\u57fa\u65bc GOP \u7684\u65b9 \u6cd5\u9032\u884c\u6539\u826f[11][12]\u3002\u53e6\u4e00\u65b9\u9762\uff0cHuang \u7b49\u4eba[8]\u5c07\u9451\u5225\u5f0f\u8a13\u7df4\u61c9\u7528\u5728 GOP \u4f30\u6e2c\uff0c\u4ee5\u6700\u5c0f \u5927\u5316 F \u5ea6\u91cf(F-measure)\u70ba\u76ee\u6a19\u4f5c\u9451\u5225\u5f0f\u8a13\u7df4\u3002Ito \u7b49\u4eba[13]\u4f7f\u7528\u6c7a\u7b56\u6a39(decision tree)\u7684\u65b9 \u6cd5\u4e26\u91dd\u5c0d\u4e0d\u540c\u932f\u8aa4\u767c\u97f3\u7684\u60c5\u6cc1\u5b9a\u7fa9\u5404\u81ea\u7684\u9580\u6abb\u503c\u4f86\u9032\u884c\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c\uff1b\u8a72\u8ad6\u6587\u7684\u5be6\u9a57\u8b49 \u660e\u5176\u6548\u679c\u52dd\u904e\u6240\u6709\u767c\u97f3\u5171\u7528\u76f8\u540c\u7684\u9580\u6abb\u503c\u3002Truong \u7b49\u4eba[14]\u6bd4\u8f03\u6c7a\u7b56\u6a39\u8207\u7dda\u6027\u9451\u5225\u5206\u6790 (linear discriminant analysis, LDA)\u7528\u65bc\u8377\u862d\u8a9e\u5b78\u7fd2\u8005\u7684\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c\u4efb\u52d9\u3002\u5ee3\u7fa9\u4e0a\u4f86\u770b\uff0c GOP \u4e5f\u5c6c\u65bc\u4e00\u7a2e\u4e8c\u5143\u5206\u985e\u7684\u65b9\u6cd5\uff0c\u4f46 GOP \u53ea\u6709\u8003\u616e\u5230\u76ee\u6a19(\u6b63\u78ba)\u97f3\u7d20\u8207\u5b83\u7684\u6df7\u6dc6\u97f3\u7d20 \u7684\u5c0d\u6578\u76f8\u4f3c\u5ea6\u503c\u3002\u6709\u9452\u65bc\u6b64\uff0cWei \u7b49\u4eba[15]\u4f7f\u7528\u76ee\u6a19\u97f3\u7d20\u8207\u5176\u5b83\u6240\u6709\u97f3\u7d20\u7684\u5c0d\u6578\u76f8\u4f3c\u5ea6 \u503c\u505a\u70ba\u8f38\u5165\u5206\u985e\u5668\u7684\u767c\u97f3\u6aa2\u6e2c\u7279\u5fb5\uff0c\u4e26\u5c07 SVM \u505a\u70ba\u5206\u985e\u5668\u4f86\u8fa8\u8a8d\u97f3\u7d20\u7279\u5fb5\u5c0d\u61c9\u7684\u8f38\u51fa \u70ba\u6b63\u78ba\u767c\u97f3\u6216\u932f\u8aa4\u767c\u97f3\u6a19\u8a18\u3002\u4f46\u9664\u4e86\u6bcf\u4e00\u500b\u97f3\u7d20\u7684\u5c0d\u6578\u76f8\u4f3c\u5ea6\u503c\u4f86\u4f5c\u70ba\u767c\u97f3\u6aa2\u6e2c\u7279\u5fb5\uff0c Hu \u7b49\u4eba[16]\u4e0d\u53ea\u4f7f\u7528[15]\u63d0\u51fa\u7684\u767c\u97f3\u6aa2\u6e2c\u7279\u5fb5\uff0c\u9084 \u984d\u5916\u5730\u5c07\u76ee\u6a19\u97f3\u7d20\u8207\u5176\u5b83\u97f3\u7d20\u7684\u5c0d\u6578 \u76f8\u4f3c\u5ea6\u6bd4\u503c\u52a0\u5165\u6210\u70ba\u984d\u5916\u8f38\u5165\u7684\u767c\u97f3\u6aa2\u6e2c\u7279\u5fb5\uff0c\u4e26\u4f7f\u7528\u7279\u6b8a\u7d50\u69cb\u7684\u908f\u8f2f\u8ff4\u6b78\u4f86\u9032\u884c\u932f\u8aa4 \u767c\u97f3\u6aa2\u6e2c\uff0c\u8a72\u7d50\u69cb\u900f\u904e\u5171\u4eab\u96b1\u85cf\u5c64\u4f86\u89e3\u6c7a\u90e8\u5206\u97f3\u7d20\u8cc7\u6599\u7a00\u758f(data sparse)\u7684\u554f\u984c\u3002\u4e0d\u540c\u65bc [16]\u7684\u8ca2\u737b\uff0c\u6211\u5011\u8a8d\u70ba\u85c9\u7531\u826f\u597d\u7684\u8072\u5b78\u6a21\u578b\u7522\u751f\u4e4b\u4e8b\u5f8c\u6a5f\u7387\u800c\u5f97\u7684\u5177\u9451\u5225\u6027\u767c\u97f3\u6aa2\u6e2c\u7279 \u5fb5\uff0c\u61c9 \u6709\u52a9\u65bc\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c\u7684\u6548\u679c\uff1b\u56e0\u6b64\uff0c\u672c\u8ad6\u6587\u5c07\u805a\u7126\u65bc\u524d\u7aef\u8072\u5b78\u6a21\u578b\u7684\u6bd4\u8f03\u8207\u878d\u5408\u3002 \u4e0a\u8ff0\u7684\u65b9\u6cd5\u7686\u662f\u904b\u7528\u8072\u5b78\u6a21\u578b\u6240\u64f7\u53d6\u7684\u767c\u97f3\u6aa2\u6e2c\u7279\u5fb5\u9032\u884c\u932f\u8aa4\u767c\u97f3\u7684\u6aa2\u6e2c\uff0c\u9664\u4e86\u5c07 \u97f3\u7d20\u6216\u8a9e\u53e5\u5206\u985e\u70ba\u6b63\u78ba\u767c\u97f3\u8207\u932f\u8aa4\u767c\u97f3\u5916\uff0c\u4e5f\u6709\u7814\u7a76\u8457\u91cd\u5728\u8a55\u65b7\u8a9e\u53e5\u7684\u767c\u97f3\u54c1\u8cea\u3002 Neumeyer \u7b49\u4eba[17]\u4f7f\u7528 HMM \u8a08\u7b97\u51fa\u5c0d\u6578\u76f8\u4f3c\u5ea6\u503c\u8207\u5f37\u5236\u5c0d\u4f4d(forced alignment)\u5f8c\u7684\u97f3 \u7d20\u767c\u97f3\u6301\u7e8c\u6642\u9593(duration)\u8cc7\u8a0a\uff0c\u4e26\u64da\u6b64\u5c0d\u975e\u6bcd\u8a9e\u5b78\u7fd2\u8005\u8a9e\u53e5\u5c64\u6b21\u7684\u767c\u97f3\u54c1\u8cea\u9032\u884c\u8a55\u4f30\u3002 Chen \u7b49\u4eba[18][19][20]\u63d0\u51fa\u8a5e\u5c64\u6b21\u7684\u767c\u97f3\u54c1\u8cea\u8a55\u4f30\uff0c\u5171\u5206\u6210 5 \u500b\u7b49\u7d1a\u4f86\u5340\u5206\u767c\u97f3\u7684\u54c1\u8cea\uff0c \u4e26\u4f7f\u7528\u8cc7\u8a0a\u6aa2\u7d22\u7684\u6392\u5e8f\u5b78\u7fd2\u6cd5(learning to rank)\u4f86\u7d50\u5408\u4e0d\u540c\u767c\u97f3\u6aa2\u6e2c\u7279\u5fb5\u7528\u65bc\u767c\u97f3\u54c1\u8cea \u8a55\u4f30\uff1b\u5176\u4e2d\uff0c\u5728[20]\u6bd4\u8f03\u5404\u985e\u767c\u97f3\u6aa2\u6e2c\u7279\u5fb5\u7684\u5f71\u97ff\u529b\u8207 4 \u7a2e\u97f3\u7d20\u5c64\u6b21\u8f49\u63db\u5230\u8a5e\u5c64\u6b21\u7684\u767c \u97f3\u6aa2\u6e2c\u7279\u5fb5\u8f49\u63db\u65b9\u6cd5\u3002 \u8b8a\u6027\u3002\u5176\u6b21\uff0cDNN \u5ffd\u7565\u4e86\u8f38\u5165\u7684\u62d3\u64b2(topological)\u7d50\u69cb\uff0c\u5b83\u7684\u8f38\u5165\u7279\u5fb5\u53ef\u4ee5\u4ee5\u4efb\u4f55\u9806\u5e8f \u8f38\u5165\u7db2\u8def\uff0c\u800c\u4e0d\u5f71\u97ff\u6700\u5f8c\u7684\u6548\u80fd[21]\uff1b\u7136\u800c\u8a9e\u97f3\u8a0a\u865f\u6240\u5c0d\u61c9\u7684\u983b\u8b5c\u5167\u5bb9\u8457\u5be6\u542b\u6709\u8c50\u5bcc\u7684 \u95dc\u806f\u6027\uff0c\u800c\u80fd\u5920\u5584\u7528\u983b\u8b5c\u7684\u5c40\u90e8\u76f8\u95dc\u6027\u800c\u5efa\u7acb\u6a21\u578b\u7684 CNN \u5728\u8a31\u591a\u4efb\u52d9\u4e0a\u7684\u6548\u679c\u90fd\u660e\u986f \u512a\u65bc DNN[25][26][27][28]\u3002\u56e0\u6b64\uff0c\u672c\u8ad6\u6587\u5c07\u878d\u5408\u5169\u8005\u7684\u512a\u9ede\uff0c\u4e26\u63a2\u8a0e\u5169\u7a2e\u985e\u795e\u7d93\u7db2\u8def\u6240 \u8a13\u7df4\u7684\u8072\u5b78\u6a21\u578b(DNN-HMM \u8207 CNN-HMM)\u5c0d\u65bc\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c\u7684\u6548\u679c\u3002 \u4e09\u3001 \u8072\u5b78\u6a21\u578b \u800c\u5728\u8072\u5b78\u6a21\u578b\u65b9\u9762\uff0c\u8207\u50b3\u7d71 GMM-(conventional filter)\u6cbf\u8457\u983b\u8b5c\u7684\u6642\u9593\u8207\u983b\u7387\u6383\u63cf\uff0c\u4ee5\u8f03\u5c11\u7684\u53c3\u6578\u6578\u91cf\u6355\u6349\u5230\u983b\u8b5c\u5e73\u79fb\u7684\u4e0d 3.1 \u6df1\u5c64\u985e\u795e\u7d93\u7db2\u8def</td></tr><tr><td>\u672c\u7bc7\u8ad6\u6587\u7684\u5b89\u6392\u5982\u4e0b\uff1a\u7b2c\u4e8c\u5c0f\u7bc0\u5c07\u4ecb\u7d39\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c\u76f8\u95dc\u7814\u7a76\u7684\u767c\u5c55\u8fd1\u6cc1\uff1b\u7b2c\u4e09\u5c0f</td></tr><tr><td>\u7bc0\u5247\u662f\u4ecb\u7d39\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c\u524d\u7aef\u6a21\u7d44\u7684\u8072\u5b78\u6a21\u578b\uff0c\u5206\u5225\u6709 GMM\u3001DNN \u8207 CNN \u4e09\u7a2e\u6a21\u578b</td></tr><tr><td>\u8207 HMM \u7684\u7d50\u5408\uff1b\u7b2c\u56db\u5c0f\u7bc0\u4ecb\u7d39\u4e09\u7a2e\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c\u7684\u65b9\u6cd5\uff0c\u5206\u5225\u662f\u767c\u97f3\u512a\u52a3\u7a0b\u5ea6</td></tr><tr><td>(goodness of pronunciation, GOP)\u3001\u652f\u6301\u5411\u91cf\u6a5f(support vector machine, SVM)\u8207\u908f\u8f2f\u8ff4\u6b78</td></tr><tr><td>(logistic regression, LR)\uff1b\u7b2c\u4e94\u5c0f\u7bc0\u5247\u662f\u5206\u6790\u4e0d\u540c\u8072\u5b78\u6a21\u578b(DNN-HMM \u548c CNN-HMM)\u5728</td></tr><tr><td>\u4e0d\u540c\u5206\u985e\u5668(GOP\u3001SVM \u548c LR)\u4e2d\u7684\u8868\u73fe\uff0c\u8207\u5c07\u5169\u7a2e\u8072\u5b78\u6a21\u578b\u7d93\u904e\u5206\u985e\u5668 LR \u6240\u7522\u751f\u7684</td></tr><tr><td>\u767c\u97f3\u6aa2\u6e2c\u5206\u6578\u503c\u4f5c\u7dda\u6027\u7d44\u5408\u5f8c\u7684\u7d50\u679c\uff0c\u4ee5\u53ca\u57fa\u65bc CNN \u8072\u5b78\u6a21\u578b\u5728\u4e0d\u540c\u5206\u985e\u5668\u6240\u7522\u751f\u7684</td></tr></table>", | |
| "text": "Keywords\uff1aMispronunciation detection, Automatic Speech Recognition, Deep Neural Networks, Convolutional Neural Networks \u3001\u8aaa(speaking)\u3001\u8b80(reading)\u548c\u5beb(writing)\u7b49\u56db\u985e\u5b78\u7fd2\u9762\u5411\u3002\u96a8\u8457\u7b2c\u4e8c\u5916\u8a9e\u5b78\u7fd2 \u8005(second language learner)\u7684\u4eba\u6578\u8207\u65e5\u4ff1\u589e\uff0c\u83ef\u8a9e\u5e2b\u8cc7\u7684\u9700\u6c42\u4e5f\u8d8a\u4f86\u8d8a\u5927\uff1b\u5c24\u5176\u5728\u8a9e\u8a00 \u5b78\u7fd2\u4e2d\uff0c\u8aaa\u8207\u5beb\u7684\u5c0d\u932f\u5f80\u5f80\u9700\u8981\u900f\u904e\u5c08\u696d\u7684\u8a9e\u8a00\u6559\u5e2b\u4f86\u8a55\u65b7\uff0c\u4f46\u8a9e\u8a00\u6559\u5e2b\u7684\u4eba\u6578\u9060\u9060\u4e0d \u53ca\u83ef\u8a9e\u5b78\u7fd2\u8005\u6578\u91cf\u3002\u56e0\u6b64\uff0c\u96fb\u8166\u8f14\u52a9\u8a9e\u8a00\u5b78\u7fd2(computer assisted language learning, CALL)\u7684\u7814\u7a76\u9818\u57df\u8d8a\u4f86\u8d8a\u91cd\u8981\uff0c\u672c\u7bc7\u8ad6\u6587\u5c07\u5c08\u6ce8\u6b64\u7814\u7a76\u9818\u57df\u6709\u95dc\u65bc\u96fb\u8166 \u8f14 \u52a9 \u767c \u97f3 \u8a13 \u7df4 (computer assisted pronunciation training, CAPT)-\u300c\u8aaa\u300d\u7684\u6280\u8853\u767c\u5c55\u8207\u63a2\u8a0e\u3002 \u5716\u4e00\u3001\u81ea\u52d5\u767c\u97f3\u6aa2\u6e2c\u4e4b\u6d41\u7a0b \u4e00 \u822c \u800c \u8a00 \uff0c \u96fb \u8166 \u8f14 \u52a9 \u767c \u97f3 \u8a13 \u7df4 (CAPT) \u5305 \u62ec \u5169 \u500b \u90e8 \u5206 \uff1a \u5206 \u5225 \u662f \u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c (mispronunciation detection)\u8207\u932f\u8aa4\u767c\u97f3\u8a3a\u65b7(mispronunciation diagnosis)\u3002\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c \u7cfb\u7d71\u662f\u8acb\u5b78\u7fd2\u8005\u8b80\u8aa6\u53e3\u8aaa\u6559\u6750\uff0c\u91dd\u5c0d\u5b78\u7fd2\u8005\u5ff5\u8aa6\u7684\u9304\u97f3\uff0c\u6a19\u8a18\u5b78\u7fd2\u8005\u7684\u767c\u97f3\u662f\u6b63\u78ba\u767c\u97f3 (correct pronunciation)\u6216\u932f\u8aa4\u767c\u97f3(mispronunciation)\uff0c\u6a19\u8a18\u7684\u76ee\u6a19\u53ef\u4ee5\u662f\u97f3\u7d20(phone)\u5c64\u6b21 \u6216\u8a5e(word)\u5c64\u6b21\uff1b\u932f\u8aa4\u767c\u97f3\u8a3a\u65b7\u662f\u7576\u7cfb\u7d71\u5075\u6e2c\u5230\u4f7f\u7528\u8005\u7684\u767c\u97f3\u51fa\u73fe\u932f\u8aa4\u6642\u7d66\u4e88\u6709\u5e6b\u52a9\u7684 \u56de\u994b\uff0c\u5047\u8a2d\u6559\u6750\u984c\u76ee\u70ba\u300c\u5e2b\u7bc4(shi1 fan4)\u300d \uff0c\u4f46\u5b78\u7fd2\u8005\u5ff5\u6210\u300c\u5403\u7bc4(chi1 fan4)\u300d \uff0c\u7cfb\u7d71\u9664\u4e86 \u5224\u65b7\u51fa\u5b78\u7fd2\u8005\u6709\u932f\u8aa4\u767c\u97f3\u4e4b\u5916\uff0c\u9084\u53ef\u4ee5\u56de\u994b\u5b78\u7fd2\u8005\u5ff5\u7684 \u300c\u5e2b(shi1)\u300d \u53ef\u80fd\u5ff5\u6210 \u300c\u5403(chi1)\u300d\u3002 HMM \u76f8\u6bd4\uff0cDNN-HMM \u5728\u8a9e\u97f3\u8fa8\u8b58\u6e96\u78ba\u7387\u4e0a\u5df2 \u88ab\u8b49\u5be6\u80fd\u6709\u986f\u8457\u7684\u6548\u80fd\u63d0\u5347[21][22]\uff0c\u9019 \u4e3b\u8981\u53ef\u80fd\u6b78\u529f\u65bc DNN \u80fd\u5920\u6a21\u64ec\u4efb\u610f\u7684\u51fd\u6578\uff0c\u80fd \u66ff\u8a9e\u97f3\u8a0a\u865f\u6240\u5167\u542b\u7684\u8907\u96dc\u5c0d\u61c9\u95dc\u4fc2\u5efa\u7acb\u6a21\u578b\uff0c\u8868\u9054\u80fd\u529b\u6bd4 GMM \u66f4\u5f37\u3002\u512a\u826f\u7684\u4e8b\u5f8c\u6a5f\u7387 \u860a\u85cf\u8c50\u5bcc\u7684\u767c\u97f3\u9451\u5225\u6027\u8cc7\u8a0a\uff0c\u4f7f\u5f97\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c\u7684\u6548\u679c\u66f4\u597d\uff0c\u6709\u8a31\u591a DNN-HMM \u61c9\u7528 \u5728 CAPT \u7684\u6548\u679c\u5df2\u88ab\u9a57\u8b49\u52dd\u904e\u50b3\u7d71\u7684 GMM-HMM[2][16][23]\uff0c\u56e0\u6b64\u8072\u5b78\u6a21\u578b\u5728\u8a08\u7b97\u4e8b \u5f8c\u6a5f\u7387\u7684\u4efb\u52d9\u4e2d\u626e\u6f14\u8457\u975e\u5e38\u95dc\u9375\u7684\u89d2\u8272[24]\uff0c\u800c\u57fa\u65bc\u6df1\u5c64\u985e\u795e\u7d93\u7db2\u8def\u7684\u8072\u5b78\u6a21\u578b\u8a08\u7b97\u800c \u5f97\u5c0d\u767c\u97f3\u6aa2\u6e2c\u6709\u5e6b\u52a9\u7684\u4e8b\u5f8c\u6a5f\u7387\u53ef\u4f7f GOP \u8207\u5176\u5b83\u5206\u985e\u5668\u9054\u5230\u6700\u4f73\u7684\u6aa2\u6e2c\u6548\u679c\u3002\u76f8\u8f03\u65bc DNN\uff0cCNN \u88ab\u8996\u70ba\u662f\u53e6\u4e00\u7a2e\u66f4\u6709\u6548\u7387\u7684\u6df1\u5c64\u985e\u795e\u7d93\u7db2\u8def\uff0c\u53ef\u7528\u65bc\u64f7\u53d6\u8a9e\u97f3\u8a0a\u865f\u4e2d\u7684\u983b \u8b5c\u8b8a\u5316\u7684\u4f4d\u79fb\u4e0d\u8b8a\u6027\u4e26\u4e14\u80fd\u91dd\u5c0d\u983b\u8b5c\u7684\u76f8\u95dc\u6027\u5efa\u7acb\u6a21\u578b[6][7] \u3002CNN \u8207 DNN \u4e0d\u540c\u5728\u65bc\uff1a \u795e\u7d93\u5143\u9593\u7684\u9023\u63a5\u4e0d\u662f\u5168\u9023\u63a5\u7684(fully-connected)\u4ee5\u53ca\u540c\u4e00\u5c64\u7684\u67d0\u4e9b\u795e\u7d93\u5143\u9593\u6703\u5171\u4eab\u9023\u63a5 \u7684\u6b0a\u91cd(weight sharing)\u3002Sainath \u7b49\u4eba[7]\u63d0\u51fa CNN \u4f5c\u70ba\u8072\u5b78\u6a21\u578b\u66f4\u52dd\u65bc DNN \u7684\u539f\u56e0\u662f \u56e0\u70ba\u4ed6\u5011\u8a8d\u70ba DNN \u6709\u5169\u9805\u7f3a\u9ede\u3002\u9996\u5148\uff0cDNN \u7684\u67b6\u69cb\u4e2d\u6c92\u6709\u660e\u78ba\u5730\u8655\u7406\u8a9e\u97f3\u8a0a\u865f\u4e2d\u7684\u4e0d \u8b8a\u7279\u5fb5\u7684\u529f\u80fd\uff0c\u4f8b\u5982\u4e0d\u540c\u8a9e\u8005\u8aaa\u8a71\u65b9\u5f0f\u4e0d\u540c\uff0c\u5728\u983b\u8b5c\u4e0a\u6703\u6709\u7d30\u5fae\u7684\u4f4d\u79fb\u3002DNN \u9700\u8981\u904b \u7528\u5404\u7a2e\u8a9e\u8005\u8abf\u9069(speaker adaptation)\u6280\u8853\u4f86\u964d\u4f4e\u7279\u5fb5\u7684\u8b8a\u5316\uff0cDNN \u540c\u6642\u9700\u8981\u5de8\u5927\u7684\u7db2\u8def \u898f \u6a21 \u53ca \u5927 \u91cf \u7684 \u8a13 \u7df4 \u6a23 \u672c (training sample) \u4f86 \u9054 \u5230 \u9019\u4ef6\u4e8b\uff1b \u4f46 CNN \u80fd\u900f\u904e\u647a\u7a4d\u6838" | |
| }, | |
| "TABREF1": { | |
| "num": null, | |
| "type_str": "table", | |
| "html": null, | |
| "content": "<table><tr><td>\u2206W \u2113 = \u2022 ( \u2113\u22121 ) \u2032 \u2113 = log (O| ) ( ) \u2211 (O| ) ( ) =1 \u63a5\u8457\u5c31\u53ef\u4ee5\u5efa\u7acb\u97f3\u7d20\u5c64\u6b21\u7684\u97f3\u7d20\u767c\u97f3\u6aa2\u6e2c\u7279\u5fb5\uff0c\u6211\u5011\u5b9a\u7fa9\u76ee\u6a19\u767c\u97f3\u7684\u97f3\u7d20 \u6240\u5c0d\u61c9\u7684 (4) \u5c0d\u61c9\u7684\u8f38\u51fa\u4e4b\u6a5f\u7387\u4e0d\u6703\u70ba 1\uff0c\u4e26\u5b9a\u7fa9\u51fd\u6578 \u70ba\u6700\u5c0f\u5316\u4ea4\u53c9\u71b5\u76ee\u6a19\u51fd\u6578\u5982\u5f0f(19)\uff0c\u63a5\u8457\u4f7f\u7528 \u97f3\u8207\u932f\u8aa4\u767c\u97f3\uff0c\u96d9\u97f3\u7bc0\u6bcf\u500b\u8a9e\u53e5\u7531 2 \u500b\u4e2d\u6587\u5b57\u7d44\u6210\uff0c\u610f\u5373\u6bcf\u500b\u8a9e\u53e5\u53ef\u62c6\u89e3\u6210 4 \u500b\u97f3\u7d20\uff0c \u56db\u7684\u5be6\u9a57\u4e2d\u89c0\u5bdf\u5230\uff0c\u82e5\u80fd\u7d66\u4e88\u5206\u985e\u5668\u66f4\u591a\u7684\u4e8b\u5f8c\u6a5f\u7387\u505a\u70ba\u7279\u5fb5\uff0c\u5c07\u53ef\u4ee5\u5f97\u5230\u66f4\u597d\u7684\u932f\u8aa4 \u6578\u210e(. )\u53ef\u8868\u793a\u70ba\uff1a \u8ddf\u96d9\u97f3\u7bc0\u7686\u56e0\u70ba\u6a21\u578b\u7684\u7d50\u5408\u4f7f\u5f97 EER \u8207 AUC \u7684\u8868\u73fe\u90fd\u6709\u6240\u63d0\u5347\u3002\u96d6\u7136 DNN-LR \u8207 (10) \u2245 log (O| ) max =1,2,\u2026, , \u2260 (O| ) (11) \u767c\u97f3\u6aa2\u6e2c\u7279\u5fb5 \u53ef\u4ee5\u8868\u793a\u70ba\u5f0f(16)\uff1a = [LPP( 1 , O ), LPP( 2 , O ), \u2026 , LPP( , O ), (16) \u96a8\u6a5f\u68af\u5ea6\u4e0b\u964d\u6cd5\u4f86\u6700\u5c0f\u5316\u76ee\u6a19\u51fd\u6578 \uff0c\u5982\u5f0f(20)\uff1a = \u2211 \u2211 ( ( | ) \u2212 ) \u2022 (20) \u4f46\u662f\u4e0d\u4ee3\u8868\u6bcf\u500b\u8a9e\u53e5\u7684\u97f3\u7d20\u90fd\u662f\u5ff5\u932f\uff0c\u56e0\u6b64\u8a9e\u53e5\u5c64\u6b21\u7684\u932f\u8aa4\u6a23\u672c\u61c9\u8a72\u8981\u53c3\u8003\u97f3\u7d20\u5c64\u6b21\u90a3 \u6b04\uff0c\u540c\u6a23\u7684\u9053\u7406\u4e5f\u5957\u7528\u5728\u55ae\u97f3\u7bc0\u8a9e\u6599\u5eab\u3002\u55ae\u97f3\u7bc0\u8a9e\u6599\u5eab\u4e2d\uff0c\u7537\u5973\u8a9e\u6599\u7684\u6bd4\u4f8b\u70ba 21:34\uff0c \u6bcd\u8a9e\u70ba\u83ef\u8a9e(L1)\u7684\u8a9e\u6599\u7686\u70ba\u53f0\u7063\u4eba\u53e3\u97f3\u6240\u9304\u88fd\uff0c\u975e\u6bcd\u8a9e\u7684\u83ef\u8a9e\u5b78\u7fd2\u8005(L2)\u6536\u9304\u7684\u53e3\u97f3\u5305 \u767c\u97f3\u6aa2\u6e2c\u7d50\u679c\u3002 \u6574\u9ad4\u800c\u8a00\uff0c\u96d9\u97f3\u7bc0\u7684\u8868\u73fe\u7686\u4e0d\u5982\u55ae\u97f3\u7bc0\uff0c\u539f\u56e0\u6709\u5169\u9ede\uff1a\u9996\u5148\uff0c\u9032\u884c\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c\u524d\uff0c \u5fc5\u9808\u5148\u900f\u904e\u8072\u5b78\u6a21\u578b\u4f86\u64f7\u53d6\u4e8b\u5f8c\u6a5f\u7387\u505a\u70ba\u6aa2\u6e2c\u7528\u7684\u7279\u5fb5\uff1b\u800c\u8072\u5b78\u6a21\u578b\u7686\u662f\u7528\u767c\u97f3\u6b63\u78ba\u7684 \u210e( ) = 2 \u2022 iRank ( ( )) \u2022 iRank ( ( CNN-LR \u5404\u81ea\u4f7f\u7528\u7684\u7d50\u679c\u4e26\u7121\u660e\u986f\u7684\u5dee\u7570\uff0c\u4f46\u7d50\u5408\u6642\u7684\u6548\u679c\u537b\u51fa\u4e4e\u610f\u6599\uff0c\u9019\u8868\u793a\u4e0d\u540c\u7684 )) iRank ( ( )) + iRank ( ( )) \u8072\u5b78\u6a21\u578b\u7522\u751f\u7684\u767c\u97f3\u6aa2\u6e2c\u7279\u5fb5\u53ef\u80fd\u5177\u6709\u4e92\u88dc\u6027\u3002\u5e0c\u671b\u5728\u672a\u4f86\u7684\u7814\u7a76\u4e2d\u53ef\u4ee5\u4f7f\u7528\u66f4\u597d\u7684\u8072 (25) \u5b78\u6a21\u578b\u7279\u5fb5(\u5982\u9451\u5225\u5f0f\u8a13\u7df4\u5f8c\u7684\u8072\u5b78\u6a21\u578b\u6240\u7522\u751f\u7684\u7279\u5fb5)\uff0c\u9664\u4e86\u8072\u5b78\u6a21\u578b\u6240\u63d0\u4f9b\u7684\u76f8\u4f3c\u5ea6 LPR( 1 , , O ), LPR( 2 , , O ), \u2026 , LPR( , , O )] =1 =1 \u62ec\u7f8e\u570b\u3001\u74dc\u5730\u99ac\u62c9\u3001\u8d8a\u5357\u3001\u97d3\u570b\u3001\u65e5\u672c\u3001\u897f\u73ed\u7259\u3001\u963f\u6839\u5ef7\u7b49 23 \u570b\u7684\u5b78\u7fd2\u8005\u53e3\u97f3\uff0c\u55ae\u97f3 \u8a9e\u53e5\u8a13\u7df4\u800c\u6210\uff0c\u4f46\u662f\u5f37\u5236\u5c0d\u4f4d\u7684\u97f3\u7d20\u908a\u754c(boundary)\u662f\u6839\u64da\u6b63\u78ba\u8a9e\u53e5\u6240\u8a13\u7df4\u7684\u8072\u5b78\u6a21\u578b \u503c\u7279\u5fb5\u5916\uff0c\u672a\u4f86\u5617\u8a66\u52a0\u5165\u97fb\u5f8b(prosodic)\u7279\u5fb5\u4e26\u63a2\u8a0e\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c\u7d50\u679c\u7684\u5f71\u97ff\uff1b\u53e6\u4e00\u65b9\u9762</td></tr><tr><td>random initial)\u4f86 \u7576\u4f5c\u7db2\u8def\u521d\u59cb\u7684\u6b0a\u91cdW\u3002\u8fd1\u5e74\u4f86\u6709\u5b78\u8005\u63d0\u51fa\u900f\u904e\u9650\u5236\u6027\u6ce2\u8332\u66fc\u6a5f(restricted boltzmann machine, RBM)\u7684\u975e\u76e3\u7763\u5f0f\u9810\u8a13\u7df4(unsupervised pre-training)[31][32][31][33]\uff0c\u9010\u5c64\u5f80\u4e0a\u9810 \u8a13\u7df4(pre-training)DNN \u7684\u53c3\u6578\u3002\u5f85\u9810\u8a13\u7df4\u5b8c\u7562\u5f8c\uff0c\u57fa\u65bc\u9810\u8a13\u7df4\u53c3\u6578\u518d\u9032\u884c\u76e3\u7763\u5f0f\u8a13\u7df4\uff0c \u53d6\u4ee3\u50b3\u7d71\u96a8\u6a5f\u521d\u59cb\u5316\u53c3\u6578\u7684\u65b9\u6cd5\u4f86\u6539\u5584\u8a9e\u97f3\u8fa8\u8b58\u7684\u6b63\u78ba\u7387[34][35][36]\uff0c\u6211\u5011\u6bcf\u5c64 DNN \u53c3\u6578\u7686\u4f7f\u7528 RBM \u4f86\u9810\u8a13\u7df4\u6b0a\u91cd\u7684\u521d\u59cb\u503c\uff0c \u2113\u22121 \u70ba\u7b2c\u2113-1\u5c64\u7684\u8f38\u51fa\u5411\u91cf\uff0c \u2113 \u70ba\u7b2c\u2113\u5c64\u7684 \u504f\u79fb\u91cf\u5411\u91cf\u3002 0 \uff1d \u2208 \u211d 0 \u00d71 \u8868\u793a\u70ba\u8f38\u5165\u8a9e\u97f3\u97f3\u6846\u5c0d\u61c9\u4e4b\u8a9e\u97f3\u7279\u5fb5\u6216\u8207\u76f8\u9130\u97f3\u6846\u5c0d\u61c9\u4e4b \u8a9e\u97f3\u7279\u5fb5\u6240\u4e32\u63a5\u800c\u6210\u7684\u7279\u5fb5\uff0c 0 \u70ba\u7279\u5fb5\u7684\u7dad\u5ea6\u3002\u5f0f(2)\u4e2d\uff0c ( )\u70ba sigmoid \u51fd\u6578\uff0c\u5176\u503c\u57df \u7bc4\u570d\u5728 0 \u5230 1 \u4e4b\u9593\u3002 DNN \u904b\u7528\u65bc\u985e\u5225(\u5982\u97f3\u7d20\u72c0\u614b\u6216\u66f4\u5c0f\u55ae\u4f4d)\u4e8b\u5f8c\u6a5f\u7387\u9810\u6e2c\u554f\u984c\u4e0a\u6642\uff0c\u6bcf\u4e00\u500b\u8f38\u51fa\u795e \u7d93\u5143\u90fd\u8868\u793a\u4e00\u7a2e\u985e\u5225\uff0c\u7e3d\u5171\u53ef\u5206\u70ba\u210b\u985e\uff0c\u8868\u793a\u70ba \u2208 {1, \u2026 , \u210b}\uff0c\u5247\u7b2c \u500b\u8f38\u51fa\u795e\u7d93\u5143\u7684\u503c \u8868\u793a\u8f38\u5165\u8a9e\u97f3\u97f3\u6846\u5c0d\u61c9\u8a9e\u97f3\u7279\u5fb5 \u5c0d\u61c9\u5230\u985e\u5225 \u7684\u6a5f\u7387 ( | )\uff0c\u5047\u8a2d\u8f38\u51fa\u5411\u91cf \u6eff\u8db3 \u591a\u9805\u5f0f\u5206\u4f48(multinomial distribution)\uff0c\u90a3\u9ebc \u9700\u8981\u6eff\u8db3 \u2265 0\u53ca\u2211 \u210b =1 = 1\uff0c\u53ef\u4ee5\u900f\u904e\u8edf \u5f0f\u6700\u5927\u5316(softmax)\u505a\u5230\uff0c\u5982\uff1a = softmax( , ) = exp( ) \u2211 exp( ) \u210b =1 (3) \u5728\u8a13\u7df4\u968e\u6bb5\uff0c\u9996\u5148\u5c0d\u6bcf\u500b\u8f38\u5165\u8a9e\u97f3\u97f3\u6846\u5c0d\u61c9\u7684\u7279\u5fb5\u505a\u5f37\u5236\u5c0d\u9f4a\uff0c\u7522\u751f\u72c0\u614b\u6a19\u7c64(state label) \u7684\u5e8f\u5217\uff0c\u9019\u4e9b\u6a19\u7c64\u7528\u65bc\u76e3\u7763\u5f0f\u8a13\u7df4\u4f86\u6700\u5c0f\u5316\u4ea4\u53c9\u71b5(cross entropy)\u76ee\u6a19\u51fd\u6578\u2212 \u2211 log \uff0c \u64ad\u6f14\u7b97\u6cd5(back-propagation)[24]\u4f7f\u7528\u96a8\u6a5f\u68af\u5ea6\u4e0b\u964d(stochastic gradient descent algorithm) \u4f86\u6700\u5c0f\u5316\u76ee\u6a19\u51fd\u6578\uff0c\u5247\u6bcf\u500b\u6b0a\u91cd\u77e9\u9663W\u7684\u66f4\u65b0\u53ef\u900f\u904e\u5f0f(4)\uff1a = log (O| ) ( ) (O) (9) LPR( , , O ) = LPP( , O ) \u2212 LPP( , O ) (15) = \u2212 ( ) (19) \u5176\u4e2d\u5f0f(18)\u7684 = {0, 1}\uff0c0 \u8868\u793a\u932f\u8aa4\u767c\u97f3\uff0c1 \u8868\u793a\u767c\u97f3\u6b63\u78ba\uff0c \u4f7f\u5f97\u767c\u97f3\u6aa2\u6e2c\u7279\u5fb5 \u96d9\u97f3\u7bc0\u8a9e\u6599\u5eab\u53ca\u55ae\u97f3\u7bc0\u8a9e\u6599\u5eab\u5169\u90e8\u5206\uff0c\u5982\u8868\u4e00\u6240\u793a\u3002\u96d9\u97f3\u7bc0\u8a9e\u6599\u5eab\u4e2d\uff0c\u7537\u5973\u8a9e\u6599\u7684\u6bd4\u4f8b \u767c\u97f3\uff0c\u800c\u975e\u6bcd\u8a9e\u7684\u83ef\u8a9e\u5b78\u7fd2\u8005(L2)\u7684\u8a9e\u6599\u5305\u542b\u65e5\u672c\u53ca\u97d3\u570b\u5169\u7a2e\u5916\u570b\u53e3\u97f3\uff0c\u6536\u9304\u4e86\u6b63\u78ba\u767c SVM \u8f38\u51fa\u5206\u6578\u7531\u4f4e\u5230\u9ad8\u6392\u540d\uff0c\u4e5f\u5c31\u662f\u5f9e\u932f\u8aa4\u767c\u97f3\u6392\u5230\u767c\u97f3\u6b63\u78ba\uff0c\u56e0\u6b64\u5b9a\u7fa9\u8abf\u548c\u5e73\u5747\u51fd \u70ba 2:3\uff0c\u6bcd\u8a9e\u70ba\u83ef\u8a9e(L1)\u7684\u8a9e\u6599\u5168\u662f\u53f0\u7063\u8a9e\u8005\u6240\u9304\u88fd\uff0c\u53ea\u6536\u9304\u6b63\u78ba\u7684\u767c\u97f3\uff0c\u6c92\u6709\u932f\u8aa4\u7684 \u6578(harmonic mean)\uff0c\u6211\u5011\u5b9a\u7fa9iRank ( ( )) \u8868\u793a\u6210\u7279\u5fb5 \u5728\u6e2c\u8a66\u96c6\u7684\u5206\u985e\u5668 \u610f\u7fa9\u662f\u8981\u6700\u5c0f\u5316 DNN \u9810\u6e2c\u7684 softmax \u8f38\u51fa\u8207\u5176\u5c0d\u61c9\u7684\u53c3\u8003\u6a19\u7c64 \u7684\u5dee\u7570\u3002\u5047\u8a2d\u53cd\u5411\u50b3 \u5716\u4e09\u3001\u647a\u7a4d\u985e\u795e\u7d93\u7db2\u8def\u4e4b\u7279\u5fb5\u67b6\u69cb 3.2 \u647a\u7a4d\u795e\u7d93\u7db2\u8def CNN \u7531\u6578\u7d44\u7684\u647a\u7a4d\u5c64(convolution layers)\u548c\u6c60\u5316\u5c64(pooling layers)\u6240\u7d44\u6210\uff0c\u647a\u7a4d\u5c64\u548c \u6c60\u5316\u5c64\u7684\u904b\u7b97\u5206\u5225\u7a31\u70ba\u647a\u7a4d(convolution)\u53ca\u6c60\u5316(pooling)\u3002\u647a\u7a4d\u5c64\u900f\u904e\u647a\u7a4d\u6838\u6383\u63cf\u8f38\u5165 \u7684\u7279\u5fb5\u5716\uff0c\u647a\u7a4d\u6838\u5c31\u50cf\u662f\u751f\u7269\u8996\u89ba\u795e\u7d93\u7684\u611f\u53d7\u5340[37]\uff0c\u6bcf\u4e00\u500b\u647a\u7a4d\u6838\u80fd\u5920\u7372\u53d6\u8f38\u5165\u7279\u5fb5 \u7684\u5c40\u90e8\u7279\u5fb5\uff1b\u800c\u6c60\u5316\u76ee\u6a19\u662f\u5c07\u647a\u7a4d\u5c64\u7684\u7279\u5fb5\u505a\u964d\u7dad\u3002\u5df2\u77e5\u8f38\u5165\u7684\u8a9e\u97f3\u7279\u5fb5\u5e8f\u5217\uff0c\u7576\u8a08\u7b97 \u97f3\u6846 \u6642\uff0c\u9700\u5de6\u53f3\u5404\u53d6 \u500b\u97f3\u6846\uff0c\u6240\u7d44\u6210\u7684\u7279\u5fb5\u5716(feature maps)\u77e9\u9663\u8868\u793a\u70ba \uff0c\u647a\u7a4d\u904b\u7b97 \u5f8c\u7684\u985e\u5225\u7279\u5fb5\u5716\u8868\u793a\u70baQ ( = 1 ,2 , \u2026 , )\uff0c\u7531 \u500b\u647a\u7a4d\u7279\u5fb5\u5716\u6240\u7d44\u6210\uff0c\u5247\u647a\u7a4d\u904b\u7b97\u53ef\u4ee5\u8996 \u70ba\u900f\u904e\u6b0a\u91cd\u77e9\u9663W , ( = 1 , \u2026 , ; = 1 , \u2026 , )\uff0c\u5c07\u8f38\u5165\u7279\u5fb5 \u6620\u5c04\u5230\u647a\u7a4d\u7279\u5fb5Q \u7684\u77e9 \u9663\u4e58\u6cd5\uff0c\u5982\u5f0f(5)\u8868\u793a\uff1a = [ \u2212C , \u2026 , \u22121 , , \u2026 , +1 , +C ] Q = ( * W , + ), ( = 1,2, \u2026 , ) (5) \u5176\u4e2d * \u8868\u793a\u70ba\u647a\u7a4d\u904b\u7b97\uff0cW , \u70ba\u5c07\u7b2c \u500b\u8f38\u5165\u7279\u5fb5\u6620\u5c04\u5230\u7b2c \u500b\u647a\u7a4d\u7279\u5fb5\u7684\u5340\u57df\u6b0a\u91cd\u77e9\u9663\uff0c \u70ba\u504f\u79fb\u91cf\u3002\u66f4\u591a\u7684\u7d30\u7bc0\u8acb\u53c3\u8003[26]\u3002\u647a\u7a4d\u5c64\u4e2d\u7684\u6b0a\u91cd\u540c\u6a23\u80fd\u900f\u904e\u53cd\u5411\u50b3\u64ad\u4f86\u5b78\u7fd2[38]\u3002 \u647a\u7a4d\u5c64\u8207\u5168\u9023\u63a5\u96b1\u85cf\u5c64\u7684\u5dee\u5225\u6709\u5169\u9ede\uff1a1)\u647a\u7a4d\u5c64\u53ea\u5f9e\u5c40\u90e8\u611f\u53d7\u91ce\u63a5\u6536\u5340\u57df\u7684\u8f38\u5165\u7279\u5fb5\uff0c \u63db\u53e5\u8a71\u8aaa\uff0c\u647a\u7a4d\u5c64\u7684\u6bcf\u500b\u5143\u7d20\u90fd\u8868\u793a\u8f38\u5165\u7684\u5340\u57df\u7279\u5fb5\u30022)\u647a\u7a4d\u5c64\u4e2d\u7684\u6bcf\u500b\u647a\u7a4d\u7279\u5fb5\u53ef\u4ee5 \u8996\u70ba\u7279\u5fb5\u5716\uff0c\u5716\u4e2d\u7684\u6bcf\u500b\u5143\u7d20\u90fd\u5171\u4eab\u76f8\u540c\u7684\u6b0a\u91cd\uff0c\u4f46\u5b83\u5011\u5404\u81ea\u662f\u6fc3\u7e2e\u81ea\u524d\u4e00\u5c64\u4e4b\u4e0d\u540c\u5340 \u57df\u7684\u7279\u5fb5\u800c\u4f86\u3002\u63a5\u4e0b\u4f86\u662f\u6c60\u5316\u7684\u90e8\u5206\uff0c\u6c60\u5316\u5c64\u662f\u5f9e\u647a\u7a4d\u5c64\u7522\u751f\u5c0d\u61c9\u7684\u6c60\u5316\u5c64\uff0c\u6bcf\u4e00\u500b\u6c60 \u5316\u7279\u5fb5\u5716\u90fd\u662f\u7531\u524d\u4e00\u5c64\u647a\u7a4d\u5c64\u7684\u647a\u7a4d\u7279\u5fb5\u5716\u505a\u6c60\u5316\u904b\u7b97\u800c\u4f86\uff0c\u56e0\u6b64\u6c60\u5316\u7279\u5fb5\u5716\u7684\u6578\u91cf\u4e5f \u6703\u8207\u647a\u7a4d\u7279\u5fb5\u5716\u7684\u6578\u91cf\u76f8\u540c\uff0c\u4e5f\u5177\u5099\u647a\u7a4d\u7279\u5fb5\u6240\u5305\u542b\u7684\u7684\u5340\u57df\u4e0d\u8b8a\u6027(local invariance)\u7684 \u7279\u6027\uff0c\u6c60\u5316\u904b\u7b97\u5206\u6210\u6700\u5927\u6c60\u5316(max-pooling)\u53ca\u5e73\u5747\u6c60\u5316(average-pooling)\u5169\u7a2e\uff0c\u4ee5\u6700\u5927\u6c60 \u5316\u6700\u591a\u4eba\u4f7f\u7528[39]\u3002\u5f71\u50cf\u8655\u7406\u4e2d\u6240\u4f7f\u7528\u7684 CNN\uff0c\u5176\u6c60\u5316\u7a97(pooling window)\u4e0d\u6703\u4e92\u76f8\u91cd \u758a\uff0c\u6c60\u5316\u7a97\u4e4b\u9593\u5f7c\u6b64\u4e26\u6392\u6c92\u6709\u7a7a\u9699\uff1b\u5728\u672c\u7bc7\u8ad6\u6587\u4e2d\uff0c\u6211\u5011\u7684\u6c60\u5316\u904b\u7b97\u4e5f\u63a1\u53d6\u9019\u6a23\u7684\u505a\u6cd5\u3002 \u56db\u3001 \u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c 4.1 \u767c\u97f3\u512a\u52a3\u7a0b\u5ea6(goodness of pronunciation, GOP) GOP \u662f\u66ff\u6bcf\u4e00\u8a5e\u5f59\u6240\u5305\u542b\u7684\u6bcf\u4e00\u500b\u97f3\u7d20\u5efa\u7acb\u4e00\u500b\u8a55\u4f30\u5206\u6578\uff0c\u4e26\u5236\u5b9a\u4e00\u500b\u9580\u6abb\u503c\u4f86 \u5340\u5206\u8a72\u97f3\u7d20\u662f\u5426\u767c\u97f3\u6b63\u78ba\u3002\u800c\u6211\u5011\u57fa\u65bc\u8a9e\u97f3\u8fa8\u8b58\u8072\u5b78\u6a21\u578b\u6240\u7d66\u4e88\u7684\u5c0d\u6578\u76f8\u4f3c\u5ea6\u503c\u4f86\u8a08\u7b97 GOP\uff0c\u82e5\u5df2\u77e5\u8a9e\u97f3\u6bb5\u843d\u7684\u8a9e\u97f3\u7279\u5fb5\u5e8f\u5217O\u5728\u5176\u76ee\u6a19(\u6b63\u78ba)\u767c\u97f3\u70ba\u97f3\u7d20 \u4e4b\u5c0d\u6578\u4e8b\u5f8c\u6a5f\u7387 log ( |O)\u5728(\u672c\u8ad6\u6587\u4e2d\u8a9e\u97f3\u7279\u5fb5\u5e8f\u5217O\u662f\u70ba\u57fa\u65bc MFCC \u6216 mel-filter bank \u8f38\u51fa\u7684\u8a9e\u97f3\u7279 \u5fb5\u6240\u69cb\u6210)\uff0c\u5247 GOP \u7684\u516c\u5f0f\u53ef\u4ee5\u5b9a\u7fa9\u6210\uff1a GOP( , O) = log ( |O) (8) \u7531\u65bc\u7121\u6cd5\u7aae\u8209\u8a9e\u53e5\u5c0d\u61c9\u7684\u6240\u6709\u8a9e\u97f3\u8a0a\u865f\uff0c\u6211\u5011\u7121\u6cd5\u5c0d\u8a9e\u97f3\u6bb5\u843d\u5c0d\u61c9\u7279\u5fb5\u5e8f\u5217O\u5efa\u7acb\u6a5f\u7387 \u6a21\u578b\uff0c\u56e0\u6b64\u5f0f(8)\u53ef\u85c9\u7531\u8c9d\u5f0f\u5b9a\u7406\u5c07\u4e8b\u5f8c\u6a5f\u7387\u8f49\u63db\u6210\u76f8\u4f3c\u5ea6\u503c (O| )\u4e58\u4e0a\u4e8b\u524d\u6a5f\u7387 ( ) \u9664\u4ee5\u7279\u5fb5\u5e8f\u5217O\u7684\u6a5f\u7387\uff0c\u5982 \u5f0f(9)\u6240\u793a\u3002\u800c\u5f0f(9)\u7684\u4e8b\u524d\u6a5f\u7387 (O)\u53ef\u4ee5\u8f49\u63db\u6210\u5c07\u6240\u6709\u97f3\u7d20\u7684 \u5c0d\u6578\u76f8\u4f3c\u5ea6\u503c\u52a0\u7e3d\u3002\u5982\u5f0f(10)\u7684\u5206\u6bcd\u9805\uff0c\u5e38\u6578 \u8868\u793a\u76ee\u6a19\u8a9e\u8a00\u4e2d\u97f3\u7d20\u7684\u7e3d\u6578\u91cf\uff0c\u5728\u932f\u8aa4 \u767c\u97f3\u6aa2\u6e2c\u7684\u4efb\u52d9\u4e2d\u4e0d\u61c9\u53d7\u5230\u97f3\u7d20\u672c\u8eab\u5728\u8a13\u7df4\u8cc7\u6599\u4e2d\u7684\u6578\u91cf\u5f71\u97ff\uff0c\u6240\u4ee5\u6211\u5011\u5047\u8a2d\u6240\u6709\u97f3\u7d20 \u7684\u4e8b\u524d\u6a5f\u7387\u7686\u76f8\u7b49( ( ) = ( ))\uff0c\u4e14\u5f0f(10)\u7684\u5206\u6bcd\u9805\u7d04\u7b49\u65bc\u97f3\u7d20 \u7684\u76f8\u4f3c\u5ea6\u503c\u53d6\u6700\u5927\u503c\uff0c \u56e0\u6b64\u5f0f(10)\u53ef\u4ee5\u88ab\u7c21\u5316\u6210\u5f0f(11)\u3002\u63a5\u8457\u5728\u5b9a\u7fa9\u9580\u6abb\u503c \u4f86\u9810\u6e2c\u767c\u97f3\u662f\u5426\u6b63\u78ba\uff1a GOP( , O) > { (12) \u5176\u4e2d\u5f0f(11)\u7684\u76f8\u4f3c\u5ea6\u503c (O| )\u5728\u8a9e\u53e5\u4e2d\u90fd\u6703\u6a6b\u8de8\u6578\u500b\u97f3\u6846\uff0c\u56e0\u6b64\u6211\u5011\u5c07\u97f3\u7d20 \u7684\u8d77\u59cb\u6642 \u9593 \u5230\u7d50\u675f\u6642\u9593 \u53d6\u5e73\u5747\uff0c\u56e0\u6b64\u97f3\u7d20 \u7684\u76f8\u4f3c\u5ea6\u503c\u53ef\u4ee5\u5beb\u6210\uff1a log (O| ) = 1 \u2212 + 1 \u2211 log ( | ) = (13) \u5728 GOP \u767c\u97f3\u7684\u8a55\u4f30\u65b9\u6cd5\u4e2d\uff0c\u53ef\u4ee5\u57fa\u65bc\u8072\u5b78\u6a21\u578b\u7684\u4e8b\u5f8c\u6a5f\u7387(\u53ef\u8996\u70ba\u4e00\u7a2e\u767c\u97f3\u6aa2\u6e2c\u7279\u5fb5) \u4f86\u9032\u884c\u8a08\u7b97\uff0c\u4e26\u900f\u904e\u9580\u6abb\u503c \u4f86\u5206\u8fa8\u767c\u97f3\u6b63\u78ba\u8207\u5426\u3002\u56e0\u6b64\uff0c\u6211\u5011\u53ef\u76f4\u89ba\u5730\u5c07 GOP \u770b\u6210\u662f \u4e00\u7a2e\u5206\u985e\u5668\uff0c\u4f46\u56e0\u70ba GOP \u53ea\u6709\u89c0\u6e2c\u76ee\u6a19\u767c\u97f3(\u6b63\u78ba)\u7684\u97f3\u7d20 \u7684\u4e8b\u5f8c\u6a5f\u7387\uff0c\u4e0b\u4e00\u5c0f\u7bc0\u5c07\u900f \u904e\u89c0\u6e2c\u5176\u5b83\u975e\u76ee\u6a19\u97f3\u7d20\u7684\u4e8b\u5f8c\u6a5f\u7387\u4e26\u4f7f\u7528\u4e0d\u540c\u7684\u5206\u985e\u6280\u8853\u4f86\u6539\u5584 GOP \u7684\u4e0d\u8db3\u3002 4.2 \u5206\u985e\u5668(Classifier) \u6b64\u5c0f\u7bc0\u5c07\u8a0e\u8ad6\u5169\u7a2e\u5206\u985e\u5668(SVM \u8207 LR)\u88ab\u5be6\u969b\u904b\u7528\u65bc\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c\u7684\u4f5c\u6cd5\u3002\u7121\u8ad6\u662f SVM \u6216\u662f LR \u5206\u985e\u5668\uff0c\u90fd\u9700\u8981\u8f38\u5165\u767c\u97f3\u6aa2\u6e2c\u7279\u5fb5 \u8207\u5c0d\u61c9\u7684 2 \u7a2e\u8f38\u51fa\u7d50\u679c{ , \u2133}\u4f5c\u70ba \u8a13\u7df4\u7684\u6a23\u672c\uff0c\u5176\u4e2d \u4ee3\u8868\u6b63\u78ba\u767c\u97f3\uff0c\u2133\u4ee3\u8868\u932f\u8aa4\u767c\u97f3\uff0c \u8868\u793a\u7b2c \u500b\u8a9e\u53e5\u7684\u7b2c \u500b\u97f3\u7d20 \u7684\u767c\u97f3\u6aa2\u6e2c\u7279\u5fb5\uff0c \u8868\u793a\u8a72\u7279\u5fb5\u5c0d\u61c9\u7684\u76ee\u6a19\u767c\u97f3\u7684(\u6b63\u78ba)\u97f3\u7d20\u3002\u8f38\u5165\u767c\u97f3\u6aa2\u6e2c\u7279\u5fb5 \u7531\u5c0d\u6578\u97f3\u7d20\u4e8b\u5f8c\u6a5f\u7387(log phone posterior, LPP)[11][17]\u8207\u5c0d\u6578\u4e8b\u5f8c\u6a5f\u7387\u6bd4\u503c(log posterior ratio, LPR)[16]\u6240\u7d44\u5408\u800c\u6210\uff0c\u6211\u5011\u63a5\u7e8c 4.1 \u5c0f\u7bc0\u6240\u63d0\u53ca\u7684\u4e8b\u5f8c\u6a5f\u7387\u8a08\u7b97\u5f0f(13)\uff0c\u5c0d\u65bc\u4efb\u610f \u97f3\u7d20 \u6211\u5011\u5c07 LPP \u5b9a\u7fa9\u6210\uff1a LPP( , O ) = log ( |O ) (14) \u9664\u6b64\u4e4b\u5916\u6211\u5011\u9084\u9700\u8981\u77e5\u9053\u76ee\u6a19\u767c\u97f3(\u6b63\u78ba)\u7684\u97f3\u7d20 \u8207\u5176\u5b83\u4efb\u610f\u97f3\u7d20 \u7684\u6bd4\u503c\uff0c\u4e5f\u5c31\u662f LPR\uff0c\u5176\u516c\u5f0f\u53ef\u4ee5\u5b9a\u7fa9\u6210\uff1a LPP( , O ) \u6703\u7b49\u65bcLPP( 1 , O ), LPP( 2 , O ), \u2026 , LPP( , O ) \u7684\u5176\u4e2d\u4e00\u9805\uff0c\u4e14\u97f3\u7d20 \u2206 = \u2022 \u7bc0 L1 \u53ca L2 \u7686\u6536\u9304\u4e86\u6b63\u78ba\u8207\u932f\u8aa4\u7684\u767c\u97f3\uff0c\u55ae\u97f3\u7bc0\u4e2d\u6bcf\u500b\u8a9e\u53e5\u90fd\u662f\u4e00\u500b\u4e2d\u6587\u5b57\uff0c\u6bcf\u500b\u4e2d \u800c\u5f97\uff0c\u56e0\u6b64\u932f\u8aa4\u767c\u97f3\u7684\u5f37\u5236\u5c0d\u4f4d\u7d50\u679c\u5c07\u7121\u6cd5\u9810\u671f\uff1b\u9019\u6a23\u7684\u60c5\u6cc1\u5728\u55ae\u97f3\u7bc0\u4e2d\u4e5f\u6703\u767c\u751f\uff0c\u4e14 \u5e0c\u671b\u63a2\u7a76\u4e0d\u540c\u7d50\u5408\u65b9\u5f0f\u8207\u5404\u5f0f\u5206\u985e\u6280\u8853\u5728\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c\u7684\u8868\u73fe\uff0c\u4e26 \u4e14\u66f4\u8a73\u7d30\u8207\u5ee3\u6cdb\u5730\u63a2 (21) \u6587\u5b57\u53ef\u62c6\u89e3\u6210 2 \u500b\u97f3\u7d20\u3002\u5169\u7a2e\u8a9e\u6599\u5eab\u5728\u8a13\u7df4\u8072\u5b78\u6a21\u578b\u6642\uff0c\u90fd\u53ea\u4f7f\u7528\u8a9e\u53e5\u5b8c\u5168\u6b63\u78ba\u7684\u6a23\u672c \u5728\u96d9\u97f3\u7bc0\u6216\u591a\u97f3\u7bc0\u7684\u8a9e\u53e5\u4e2d\u5c07\u6703\u66f4\u56b4\u91cd\u3002\u7b2c\u4e8c\u500b\u53ef\u80fd\u7684\u539f\u56e0\u5247\u662f\u96d9\u97f3\u7bc0\u7684\u8cc7\u6599\u91cf\u76f8\u8f03\u65bc \u8a0e\u5404\u7a2e\u8072\u5b78\u6a21\u578b\u6240\u64f7\u53d6\u7684\u767c\u97f3\u7279\u5fb5\u4e4b\u512a\u7f3a\u9ede\u3002 ( = 1, 2, 3, \u2026 , )\u7684\u5176\u4e2d\u4e00\u9805\u7b49\u65bc\u97f3\u7d20 \u6642\uff0cLPR( , , O )\u6703\u70ba 0\u3002\u800c 4.1 \u7bc0\u63d0\u5230\u7684 GOP \u8a55\u4f30\u503c\u7b49\u540c\u65bc\u767c\u97f3\u6aa2\u6e2c\u7279\u5fb5 \u5f0f(21)\u4e2d\u7684\u53c3\u6578 \u70ba\u6b0a\u91cd \u66f4\u65b0\u6642\u7684\u5b78\u7fd2\u7387\uff0c\u5b78\u7fd2\u7387\u5c07\u96a8\u8457\u66f4\u65b0\u7684\u6b21\u6578\u9032\u884c\u8abf\u6574\uff0c\u7d93\u904e \u4f86\u8a13\u7df4\u8072\u5b78\u6a21\u578b\uff0c\u800c\u5728\u8a13\u7df4\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c\u6a21\u578b\u6642\u5247\u6703\u4f7f\u7528\u932f\u8aa4\u767c\u97f3\u8207\u6b63\u78ba\u7684\u8a9e\u53e5\uff0c\u4ee5\u97f3 \u55ae\u97f3\u7bc0\u9084\u8981\u5c11\u8a31\u591a\uff0c\u56e0\u6b64\u4e00\u4e9b\u8f03\u7279\u5225\u7684\u932f\u8aa4\u767c\u97f3\u4e26\u672a\u5728\u8a13\u7df4\u8cc7\u6599\u4e2d\u51fa\u73fe\u3002 \u5f8c\u534a\u90e8\u7684\u5176\u4e2d\u4e00\u500b\u7dad\u5ea6\u4e4b\u5012\u6578\uff1b\u56e0\u6b64\u3002\u5229\u7528\u7279\u5fb5 \u8a13\u7df4\u51fa\u7684\u5206\u985e\u5668\u5c07\u6703\u6bd4 GOP \u64c1\u6709\u66f4\u591a\u95dc\u65bc\u767c\u97f3\u7684\u8a0a\u606f\u3002\u63a5\u8457\u5c07\u4ecb\u7d39\u672c\u8ad6\u6587\u5617\u8a66\u6bd4 \u8f03\u7684\u5169\u7a2e\u5206\u985e\u5668\u3002 LR \u88ab\u5ee3\u6cdb\u5229\u7528\u5728\u4e8c\u985e\u5206\u985e\u554f\u984c\u7684\u4efb\u52d9\u4e2d[16][18]\uff0c\u5229\u7528 sigmoid \u7684\u7279\u6027\u4f86\u8868\u793a\u8cc7\u6599 \u7684\u5206\u4f48\uff0c\u4f46\u5728\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c\u7684\u4efb\u52d9\u4e2d\uff0c\u4e0d\u540c\u97f3\u7d20\u61c9\u8a72\u4f7f\u7528\u4e0d\u540c\u7684 LR \u5206\u985e\u5668\uff0c\u82e5\u5c07\u6240\u6709 \u97f3\u7d20\u6df7\u5728\u4e00\u8d77\u9032\u884c\u8ff4\u6b78\u5206\u6790\u53ef\u80fd\u5c0e\u81f4\u904e\u5ea6\u6df7\u6dc6\u3002\u4ee5\u4e0b\u5148\u4ecb\u7d39\u5206\u985e\u5668 LR \u5c0d\u6b63\u78ba\u767c\u97f3\u3001\u932f \u8aa4\u767c\u97f3\u6a23\u672c\u7684\u6a5f\u7387\u8868\u793a\u5982\u5f0f(17)\uff1a ( | ) = ( ) (\u2133| ) = 1 \u2212 ( | ) (17) (. )\u70ba sigmoid \u51fd\u6578\uff0c ( | )\u70ba\u5df2\u77e5\u6709\u767c\u97f3\u6aa2\u6e2c\u7279\u5fb5 \u4e0b\u767c\u751f \u7684\u6a5f\u7387\uff0c (\u2133| ) \u70ba\u5df2\u77e5\u767c\u97f3\u6aa2\u6e2c\u7279\u5fb5 \u767c\u751f\u2133\u7684\u6a5f\u7387\uff0c \u5247\u662f\u900f\u904e\u5b78\u7fd2\u4f86\u66f4\u65b0\u7684\u6b0a\u91cd(weight)\uff0c\u4e0d\u540c \u8a9e\u53e5\u4e2d\u76f8\u540c\u7684\u97f3\u7d20\u4e5f\u6703\u4f7f\u7528\u76f8\u540c\u7684\u6b0a\u91cd\uff0c\u63a5\u8457\u5b9a\u7fa9\u76f8\u4f3c\u5ea6\u503c\u51fd\u6578 \uff1a = \u220f \u220f ( | ) (\u2133| ) 1\u2212 =1 =1 (18) \u8868\u4e00\u3001\u55ae\u97f3\u7bc0\u8a9e\u6599\u5eab\u8207\u96d9\u97f3\u7bc0\u8a9e\u6599\u5eab\u4e4b\u5167\u5bb9 -\u55ae\u97f3\u7bc0 \u96d9\u97f3\u7bc0 \u6bcd\u8a9e(L1) \u7b2c\u4e8c\u5916\u8a9e(L2) L1 L2 L1 L2 \u4eba\u6578(\u4eba) 62 63 115 40 \u97f3\u7d20\u5c64\u6b21 \u6b63\u78ba\u767c\u97f3(T) \u97f3\u7d20\u5c64\u6b21 \u6b63\u78ba\u767c\u97f3(F) T F T F T F T F \u6642\u9593(\u5c0f\u6642) 9.32 1.04 13.79 9.03 3.97 0 0.89 0.94 \u8a9e\u53e5\u6578(\u53e5) 37,976 4,827 50,856 32,726 10,384 0 1,994 2,003 \u97f3\u7d20\u6578\u91cf(\u500b) 76,638 4,976 119,512 36,862 38,939 0 12,449 2,539 \u6578\u6b21\u66f4\u65b0\u5f8c\u76f4\u5230\u6b0a\u91cd \u7684\u6539\u8b8a\u904e\u5c0f\u5247\u6536\u6582\uff0c\u63a5\u8457\u7576\u8f38\u5165\u767c\u97f3\u6aa2\u6e2c\u7279\u5fb5\u70ba \u6642\uff0c\u8a72\u6bb5 \u767c\u97f3\u70ba\u6b63\u78ba\u767c\u97f3\u7684\u6a5f\u7387\u5247\u70ba ( | ) = ( )\u3002 SVM[15]\u662f\u4e00\u7a2e\u6548\u80fd\u8868\u73fe\u826f\u597d\u7684\u5206\u985e\u5668\uff0c\u4ed6\u53ef\u4ee5\u900f\u904e\u5c07\u7279\u5fb5\u8f49\u63db\u5230\u66f4\u9ad8\u7dad\u5ea6\u7684\u7a7a\u9593 \u4f86\u89e3\u6c7a\u8cc7\u6599\u7dda\u6027\u4e0d\u53ef\u5206\u7684\u554f\u984c\uff0c\u6211\u5011\u5b9a\u7fa9\u51fd\u6578s(. )\u7528\u4f86\u8868\u793a SVM \u7d66\u4e88\u7279\u5fb5\u7684 \u6c7a\u7b56 \u503c\uff0c\u4e26\u5c07s( )\u4ee3\u5165 sigmoid \u51fd\u6578 (. )\u7528\u4ee5\u8868\u793a\u6b63\u78ba\u767c\u97f3\u7684\u6a5f\u7387 ( | ) = (s( ))\u3002 \u672c\u7bc7\u8ad6\u6587\u4f7f\u7528 python \u7684\u73fe\u6709\u6a21\u7d44\"scikit-learn[40]\"\u6240\u63d0\u4f9b\u7684 SVM \u8207 LR \u5de5\u5177\uff0c\u6838\u5fc3\u51fd\u6578 \u70ba\u5f91\u5411\u57fa\u51fd\u6578\u6838(radial basis function kernel)\u3002 \u4e94\u3001 \u5be6\u9a57 \u6bcf\u7a2e\u8a9e\u8a00\u7684\u932f\u8aa4\u53ef\u5206\u6210\u4e09\u7a2e\uff1a\u66ff\u63db(substitution)\u3001\u63d2\u5165(insertion)\u3001\u522a\u9664(deletion)[41]\u3002 \u5c0d\u83ef\u8a9e\u4f86\u8aaa\uff0c\u6bcf\u500b\u5b57(character)\u90fd\u5c6c\u65bc\u4e00\u500b\u97f3\u7bc0\uff0c\u800c\u6bcf\u500b\u97f3\u7bc0\u53c8\u53ef\u62c6\u6210\u4e09\u500b\u90e8\u5206\uff1a\u8072\u6bcd (initial)\u3001\u97fb\u6bcd(final)\u3001\u8072\u8abf(tone)\u3002\u5c0d\u65bc\u6709\u83ef\u8a9e\u57fa\u790e\u77e5\u8b58\u7684\u5b78\u7fd2\u8005\u800c\u8a00\u4e26\u4e0d\u6613\u767c\u751f\u63d2\u5165 \u53ca\u522a\u9664\u7684\u932f\u8aa4\uff0c\u4f46\u83ef\u8a9e\u662f\u4e00\u7a2e\u8072\u8abf\u8a9e\u8a00(tonal language)\uff0c\u8072\u8abf\u7684\u767c\u97f3\u76f8\u8f03\u65bc\u8072\u6bcd\u3001\u97fb\u6bcd \u5247\u66f4\u5bb9\u6613\u5ff5\u932f\u3002[8][42][43] \u7684\u7814\u7a76\u4e0d\u63a2\u7a76\u8072\u8abf\u7684\u5f71\u97ff\uff0c\u800c\u672c\u8ad6\u6587\u5c07\u8072\u8abf\u4f9d\u9644\u5728\u97fb\u6bcd\u4e4b \u5f8c\uff0c\u4e5f\u5c31\u662f\u4e00\u500b\u97f3\u7bc0\u53ef\u62c6\u6210\u8072\u6bcd\u53ca\u8072\u8abf\u97fb\u6bcd(tonal final)\u5169\u500b\u97f3\u7d20\u3002 5.1 \u8a9e\u6599\u5eab \u65b9\u5f0f\uff0c\u5728\u6b64\u6211\u5011\u57fa\u65bc\u6bcf\u500b\u97f3\u7d20\u5728\u4e0d\u540c\u5206\u985e\u5668\u4e4b\u7d50\u679c\u7684\u6392\u540d\uff0c\u4e26\u5c0d\u6392\u540d\u7d50\u679c\u8a08\u7b97\u8abf\u548c\u5e73\u5747 \u6211\u5011\u7684\u8a9e\u6599\u5eab\u4f7f\u7528\u81fa\u7063\u5e2b\u7bc4\u5927\u5b78\u9081\u5411\u9802\u5c16\u5927\u5b78\u8a08\u756b\u7684\u83ef\u8a9e\u5b78\u7fd2\u8005\u53e3\u8a9e\u8a9e\u6599\u5eab\uff0c\u5206\u6210 \u8868\u4e8c\u3001\u55ae\u97f3\u7bc0\u8207\u96d9\u97f3\u7bc0\u5728\u4e0d\u540c HMM \u7684\u5b57\u932f\u8aa4\u7387(character error rate, CER)\u8207\u97f3\u7d20\u932f \u8aa4\u7387(phone error rate, PER) ASR performance \u55ae\u97f3\u7bc0 (%) \u96d9\u97f3\u7bc0 (%) CER PER CER PER L1 L2 L1 L2 L1 L2 L1 L2 GMM 66.53 80.16 46.00 58.70 55.83 57.29 39.66 39.45 DNN 22.25 37.11 13.34 24.71 15.62 24.37 10.20 16.46 CNN(a) 21.17 36.32 12.76 24.23 16.06 22.61 10.37 14.95 CNN(b) 20.15 36.05 12.01 24.32 17.16 24.37 11.89 16.08 \u7d20\u5c64\u6b21\u7684\u767c\u97f3\u4f86\u8a13\u7df4\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c\u6a21\u578b\u3002 5.2 \u5be6\u9a57\u8a2d\u5b9a \u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c\u7cfb\u7d71\u7684\u512a\u52a3\u8207\u8a9e\u97f3\u8fa8\u8b58\u7cfb\u7d71\u7684\u8868\u73fe\u606f\u606f\u76f8\u95dc\uff0c\u56e0\u6b64\u6211\u5011\u5148\u5206\u6790\u8a9e\u97f3\u8fa8 \u8b58\u7cfb\u7d71\u7684\u8868\u73fe\u3002\u6211\u5011\u5c07\u8a9e\u6599\u5eab\u4e2d\u7684\u6b63\u78ba\u767c\u97f3\u5206\u6210\u8a13\u7df4\u96c6(training set) \u3001\u767c\u5c55\u96c6(development set)\u8207\u6e2c\u8a66\u96c6(test set)\uff0c\u4f7f\u7528\u9ad8\u65af\u6df7\u5408\u6a21\u578b\u91dd\u5c0d\u8a13\u7df4\u96c6\u4f86\u5b78\u7fd2\u8a9e\u97f3\u8a0a\u865f\u7684\u5206\u4f48\uff0c\u4ee5\u53ca\u57fa\u65bc GMM-\u6587\u7684\u76ee\u7684\u70ba\u97f3\u7d20\u5c64\u6b21\u7684\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c\uff0c\u56e0\u6b64\u6211\u5011\u5c07\u9078\u64c7\u5c0d\u65bc\u83ef\u8a9e\u5b78\u7fd2\u8005(L2)\u4e14\u97f3\u7d20\u932f \u8aa4\u7387\u8f03\u4f4e\u7684\u8072\u5b78\u6a21\u578b\u505a\u70ba\u7522\u751f\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c\u6240\u9700\u7684\u7279\u5fb5\u3002 5.3 \u5be6\u9a57\u7d50\u679c \u5716\u56db\u6211\u5011\u6bd4\u8f03\u4e86\u8072\u5b78\u6a21\u578b DNN\u3001CNN \u5206\u5225\u4f7f\u7528 GOP\u3001SVM\u3001LR \u7b49\u5206\u985e\u5668\u6240\u7522\u751f \u7684 6 \u7a2e\u7d50\u679c\uff0c\u6bcf\u7a2e\u7d50\u679c\u90fd\u662f\u7531\u4e0d\u540c\u5206\u985e\u5668\u6240\u7522\u751f\u7684\u8f38\u51fa\u5206\u6578\u4e26\u900f\u904e\u8abf\u6574\u9580\u6abb\u503c\u4f86\u7e6a\u88fd\u5716 \u56db\u3001\u4e94\u8207\u516d\u7684 Recall-Precision \u66f2\u7dda\uff0c\u6211\u5011\u5c07\u66f2\u7dda\u4e2d\u53ec\u56de\u7387\u8207\u7cbe\u6e96\u5ea6\u76f8\u540c\u7684\u9ede\u4f5c\u70ba\u8a55\u4f30\u6a19 \u6e96\u3002\u5176\u4e2d\u53ec\u56de\u7387\u8207\u7cbe\u6e96\u5ea6\u6240\u986f\u793a\u7684\u6578\u503c\u662f\u5c0d\u65bc\u932f\u8aa4\u767c\u97f3\u7684\u6a23\u672c\u6240\u505a\u7684\u8a08\u7b97\uff0c\u7531\u65bc\u767c\u97f3 \u6b63\u78ba\u7684\u6a23\u672c\u6578\u591a\u904e\u65bc\u932f\u8aa4\u7684\u767c\u97f3\uff0c\u56e0\u6b64\u5728\u672c\u8ad6\u6587\u7684\u5be6\u9a57\u4e2d\u5c07\u4e0d\u984d\u5916\u63a2\u8a0e\u6b63\u78ba\u767c\u97f3\u7684 Recall-Precision \u66f2\u7dda\u3002\u9996\u5148\u5206\u6790\u55ae\u97f3\u7bc0\u7684\u90e8\u5206(\u5716\u56db\u5de6)\uff0c\u5728\u5206\u985e\u5668 GOP\u3001SVM\u3001LR \u4f7f \u8072\u5b78\u6a21\u578b DNN \u8207 CNN \u5728\u5206\u985e\u5668 LR \u4e0b\u5404\u81ea\u7684\u8868\u73fe\u8207\u7d50\u5408\u5f8c\u7684\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c\u4e4b\u8868 \u73fe\u5982\u5716\u4e94\uff0c\u6211\u5011\u5c07 CNN-LR \u7684\u5206\u985e\u6a5f\u7387\u51fd\u6578\u5b9a\u7fa9\u6210 (. )\uff0cDNN-LR \u7684\u5206\u985e\u6a5f\u7387\u51fd\u6578 \u5b9a\u7fa9\u6210 (. )\uff0c\u5ef6\u7e8c 4.2 \u5c0f\u7bc0\u7684\u7279\u5fb5 \uff0c\u5176\u4e2d ( )\u3001 ( )\u53ef\u4ee5\u8868\u793a\u6210\uff1a ( ) = (( ) ) (22) ( ) = (( ) ) (23) \u6b0a\u91cd \u6703\u56e0\u70ba\u8072\u5b78\u6a21\u578b\u7684\u4e0d\u540c\u800c\u4f7f\u7528\u4e0d\u540c\u7684\u6b0a\u91cd( \u8207 )\uff0c \u8868\u793a\u5c0d\u61c9\u97f3\u7d20 \u7684\u6b0a\u91cd\uff0c\u5728 4.2 \u5c0f\u7bc0\u6709\u8aaa\u660e \u7684\u8a13\u7df4\u65b9\u5f0f\u4ee5\u53ca\u63d0\u5230\u6bcf\u500b\u97f3\u7d20\u61c9\u5206\u958b\u8a13\u7df4\uff0c\u56e0\u70ba\u5404\u97f3\u7d20 \u7684\u5c0d\u932f\u60c5\u6cc1\u5404\u6709\u4e0d\u540c\uff0c\u61c9\u907f\u514d\u5728\u540c\u4e00\u5206\u985e\u5668\u4e2d\u7522\u751f\u4e0d\u5fc5\u8981\u7684\u6df7\u6dc6\u3002\u63a5\u8457\u6211\u5011\u5728\u5b9a\u7fa9\u4e00\u500b \u53c3\u6578\u03bb\uff0c\u5176\u503c\u57df\u70ba0 \u2264 \u2264 1\uff0c\u8a72\u53c3\u6578\u7528\u4f86\u7dda\u6027\u7d50\u5408 \u8207 \u7684\u7d50\u679c\uff1a ( ) = \u2022 ( ) + (1 \u2212 ) \u2022 ( ) (24) ( )\u5247\u70ba\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c\u6a21\u578b DNN-LR \u8207 CNN-LR \u8f38\u51fa\u5206\u6578\u7684\u7d50\u5408\uff0c\u5982\u540c 4.1 \u5c0f\u7bc0\u7684 \u5f0f(8)\u5b9a\u7fa9\u9580\u6abb\u503c \u4f86\u6c7a\u5b9a\u767c\u97f3\u70ba\u6b63\u78ba\u6216\u932f\u8aa4\uff0c\u5728\u5716\u4e94\u7684\u5be6\u9a57\u4e2d\u6211\u5011\u5c07\u03bb\u8a2d\u5b9a\u70ba 0.5\uff0c\u4e26\u8abf \u6574\u9580\u6abb\u503c \u756b\u51fa\u5716\u4e94\u7684\u66f2\u7dda\uff0c\u5f9e\u5716\u4e94(\u5de6)\u55ae\u97f3\u7bc0\u5be6\u9a57\u4e2d\u53ef\u4ee5\u89c0\u5bdf\u5230\u7dda\u6027\u7d50\u5408\u5169\u7a2e\u7279\u5fb5\u6240 \u7522\u751f\u7684\u6a5f\u7387\u503c\u5c07\u53ef\u4ee5\u5f97\u5230\u4e0d\u932f\u7684\u6210\u6548\uff0c\u7531 DNN-LR \u7684 63.81%\u9032\u6b65\u81f3\u7dda\u6027\u7d50\u5408\u5f8c\u7684 67.23%\u7d04\u6709 3.42%\u7684\u9032\u6b65\uff0c\u800c\u96d9\u97f3\u7bc0(\u5716\u4e94\u53f3)\u7d93\u904e\u7dda\u6027\u7d50\u5408\u5f8c\u5f9e 54.47%\u5230 55.98%\u5f97\u5230 1.51%\u7684\u9032\u6b65\u3002 \u63a5\u8457\u5728\u5716\u516d\u4e2d\u6211\u5011\u5c07\u63a2\u8a0e\u4f7f\u7528 CNN \u8072\u5b78\u6a21\u578b\u6240\u64f7\u53d6\u7684\u7279\u5fb5\u5728\u4e0d\u540c\u5206\u985e\u5668(SVM\u3001 LR)\u7d50\u5408\u7684\u6548\u679c\uff0c\u7531\u65bc\u4e0d\u540c\u5206\u985e\u5668\u7684\u8f38\u51fa\u503c\u57df\u4e26\u4e0d\u4e00\u81f4\uff0c\u6240\u4ee5\u6211\u5011\u4e0d\u4f7f\u7528\u5f0f(24)\u7684\u7d50\u5408 \u56e0\u6b64\u51fd\u6578\u210e(. )\u7684\u8f38\u51fa\u5206\u6578\u5982\u540c\u51fd\u6578 (. )\u548c (. )\uff0c\u8d8a\u9ad8\u8868\u793a\u6b63\u78ba\u767c\u97f3\u3001\u8d8a\u4f4e\u8868\u793a\u932f\u8aa4\u767c\u97f3\uff0c \u51fd\u6578\u210e(. )\u8207iRank(. )\u7684\u503c\u57df\u70ba1~ \uff0c\u5e38\u6578 \u8868\u793a\u6e2c\u8a66\u96c6\u7684\u97f3\u7d20\u5c64\u6b21\u6a23\u672c\u6578\u3002\u5f9e\u5716\u516d(\u5de6)\u53ef \u4ee5\u767c\u73fe\u8072\u5b78\u6a21\u578b CNN \u4e4b\u7279\u5fb5\u7528\u65bc\u5206\u985e\u5668 SVM \u8207 LR \u4e4b\u7d50\u5408\u5728\u55ae\u97f3\u7bc0\u7684\u8868\u73fe(65.21%)\u4e0d \u5982\u5716\u4e94(\u5de6)(27)\uff1a TPR = TP (TP + FN) (26) FPR = FP (FP + TN) (27) \u66f2\u7dda\u4e0b\u9762\u7a4d(AUC)\u4e4b\u6bd4\u8f03 -TP TN FP FN AUC (%) DNN 921 185 61 301 80.90 CNN 925 186 60 297 81.18 C+D 940 189 57 282 82.58 \u5716\u516b\u3001\u96d9\u97f3\u7bc0\u5728\u5206\u985e\u5668 LR \u5728\u4e0d\u540c\u8072\u5b78 \u6a21\u578b\u7684 ROC \u66f2\u7dda CNN 9,569 2,880 821 2,698 84.02 C+D 9,821 2,967 734 2,446 86.70 \u5716\u4e03\u3001\u55ae\u97f3\u7bc0\u5728\u5206\u985e\u5668 LR \u5728\u4e0d\u540c\u8072\u5b78 \u6a21\u578b\u7684 ROC \u66f2\u7dda \u7528\u4e0d\u540c\u8072\u5b78\u6a21\u578b(CNN \u8207 \u5716\u4e94\u3001\u6bd4\u8f03\u96d9\u97f3\u7bc0(\u5de6\u5074)\u8207\u55ae\u97f3\u7bc0(\u53f3\u5074)\u5f9e\u4e0d\u540c HMM \u64f7\u53d6\u51fa\u7684\u767c\u97f3\u6aa2\u6e2c\u7279\u5fb5 \u4f7f\u7528 LR \u5206\u985e\u5668\u8207\u4e0d\u540c HMM \u767c\u97f3\u6aa2\u6e2c\u4e4b LR \u5206\u985e\u5668\u8f38\u51fa\u5206\u6578\u7684\u7dda\u6027\u7d44\u5408(M/DS-C+D-LR)\u4e4b Recall-Precision \u66f2\u7dda \u7121\u8ad6\u662f\u55ae\u97f3\u7bc0\u6216\u96d9\u97f3\u7bc0\u4e2d\uff0c\u53ef\u4ee5\u89c0\u5bdf\u5230\u5206\u985e\u5668 LR \u7684\u8868\u73fe\u7686\u512a\u65bc SVM\uff1b\u6211\u5011\u4f7f\u7528 \u8072\u5b78\u6a21\u578b DNN \u7522\u751f\u7684\u6aa2\u6e2c\u7279\u5fb5\u6240\u8a13\u7df4\u7684\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c\u6a21\u578b\u5728\u55ae\u97f3\u7bc0\u8a13\u7df4\u8cc7\u6599\u4e2d\u9032\u884c\u6e2c \u8868\u4e09\u3001\u55ae\u97f3\u7bc0\u5728\u5206\u985e\u5668 LR \u5728\u4e0d\u540c\u8072\u5b78\u6a21 \u81f4\u8b1d \u578b\u7684 ROC \u7a7a\u9593\u503c(TP\u3001TN\u3001FP \u8207 FN)\u8207 \u8a66(test on train set)\uff0c\u6703\u767c\u73fe\u5206\u985e\u5668 SVM \u7684\u6a21\u578b\u5c0d\u65bc\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c\u7684 Recall-Precision \u76f8 \u540c\u6642\u53ef\u4ee5\u9054\u5230 99.49%\uff0c\u800c\u5206\u985e\u5668 LR \u5247\u662f 86.73%\uff0c\u4f46\u662f\u63db\u5230\u6e2c\u8a66\u8cc7\u6599\u6642\u5247\u662f LR \u8868\u73fe \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(104-2911-I-003-301) \u66f2\u7dda\u4e0b\u9762\u7a4d(AUC)\u4e4b\u6bd4\u8f03 \u8207\u884c\u653f\u9662\u79d1\u6280\u90e8\u7814\u7a76\u8a08\u756b(MOST 104-2221-E-003-018-MY3, MOST 103-2221-E-003-016-\u52dd\u904e SVM\u3002\u6211\u5011\u5c0d\u65bc SVM \u6548\u679c\u4e0d\u5982 LR \u5206\u985e\u5668\u7684\u73fe\u8c61\u6709\u5169\u7a2e\u89e3\u91cb\uff1a1)\u7531\u4e0a\u8ff0\u73fe\u8c61\u53ef\u89c0 MY2, NSC 103-2911-I-003-301)\u4e4b\u7d93\u8cbb\u652f\u6301\uff0c\u8b39\u6b64\u81f4\u8b1d\u3002 \u5bdf\u5230 SVM \u767c\u751f\u904e\u5ea6\u64ec\u5408\u7684\u73fe\u8c61\uff0c\u4f7f\u5f97\u8f49\u63db\u5230\u6e2c\u8a66\u8cc7\u6599\u9032\u884c\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c\u6642\u7684\u8868\u73fe\u4e0d\u5982 -TP TN FP AUC FN (%) \u9810\u671f\uff1b2)\u6211\u5011\u4f7f\u7528\u7684 SVM \u6838\u5fc3\u51fd\u6578\u6703\u5c07\u7279\u5fb5\u8f49\u63db\u5230\u8f03\u9ad8\u7684\u7dad\u5ea6\u4ee5\u4fbf\u9032\u884c\u7dda\u6027\u8ff4\u6b78\u5206\u6790\uff0c \u53ef\u80fd\u8f49\u63db\u7684\u65b9\u6cd5\u4e26\u4e0d\u5b8c\u5168\u9069\u7528\u65bc\u6e2c\u8a66\u8cc7\u6599\u3002\u56e0\u6b64\uff0c\u6211\u5011\u5728\u63a5\u4e0b\u4f86\u7684\u5be6\u9a57\u5c07\u63a2\u8a0e\u5206\u985e\u5668 LR \u5728\u4e0d\u540c\u8072\u5b78\u6a21\u578b\u4ee5\u53ca\u4e0d\u540c\u6aa2\u6e2c\u6a21\u578b\u4e4b\u8f38\u51fa\u5206\u6578\u5728\u7dda\u6027\u7d44\u5408\u4e0a\u7684\u8868\u73fe\u3002 DNN 9,609 2,896 805 2,658 84.42 \u4e03\u3001 \u53c3\u8003\u6587\u737b</td></tr></table>", | |
| "text": "\u5176\u4e2d \u70ba\u5b78\u7fd2\u7387(learning rate)\uff0c \u2113 \u70ba\u7b2c\u2113\u5c64\u7684\u932f\u8aa4\u8a0a\u865f(error signal)\u3002 HMM \u6240\u6821\u6e96\u7684\u6587\u5b57\u8207\u97f3\u6846\u4e4b\u5c0d\u61c9\u95dc\u4fc2\u4f86\u85c9\u7531 DNN-HMM \u8207 CNN-DNN-HMM \u5b78\u7fd2\u8a9e\u97f3\u8a0a\u865f\u7684\u5206\u4f48\uff0c\u70ba\u4e86\u63cf\u8ff0\u65b9\u4fbf\uff0c\u5c07\u8072\u5b78\u6a21\u578b GMM-HMM\u3001DNN-HMM \u8207 CNN-DNN-HMM \u7c21\u7a31\u70ba GMM\u3001DNN \u8207 CNN\u3002\u767c\u5c55\u96c6\u76ee\u7684\u5728\u65bc\u985e\u795e\u7d93\u7db2\u8def\u7b49\u76f8\u95dc\u7684\u6a21\u578b\u5728 \u8a13\u7df4\u6642\u5bb9\u6613\u767c\u751f\u904e\u5ea6\u64ec\u5408(over fitting)\uff0c\u56e0\u6b64\u6211\u5011\u5207\u51fa\u4e00\u584a\u767c\u5c55\u96c6\u4f86\u5f15\u5c0e\u6a21\u578b\u5728\u8a13\u7df4\u6642 \u4e0d\u8981\u904e\u5ea6\u50be\u5411\u8a13\u7df4\u96c6\u3002\u63a5\u8457\u518d\u4f7f\u7528 GMM\u3001DNN \u8207 CNN \u6240\u8a13\u7df4\u51fa\u7684\u8a9e\u97f3\u8fa8\u8b58\u5668\u5c0d\u6e2c\u8a66 \u96c6\u9032\u884c\u8fa8\u8b58\uff0c\u8fa8\u8b58\u7d50\u679c\u5982\u8868\u4e8c\u3002 \u6211\u5011\u57fa\u65bc Kaldi \u8a9e\u97f3\u8fa8\u8b58\u5de5\u5177[44]\uff0c\u5c07\u83ef\u8a9e\u5b78\u7fd2\u8005\u767c\u97f3\u7684\u8a9e\u97f3\u8a0a\u865f\u5207\u6210\u97f3\u6846\u5f8c\uff0c\u9130 \u8fd1\u7684\u6578\u500b\u97f3\u6846\u6574\u5408\u6210\u9130\u8fd1\u97f3\u7a97(context window)\uff0c\u901a\u5e38\u63a1\u7528\u524d\u5f8c\u5404 5 \u500b\u97f3\u6846\uff0c\u7e3d\u5171 11 \u500b \u97f3\u6846\u4f86\u7576\u4f5c\u4e00\u500b\u9130\u8fd1\u97f3\u7a97\u7684\u5927\u5c0f\uff0c\u5f9e\u97f3\u7a97\u7684\u8a9e\u97f3\u8a0a\u865f\u62bd\u53d6\u51fa 13 \u7dad\u7684 MFCC \u7279\u5fb5\u52a0\u4e0a 3 \u7dad\u5ea6\u7684\u97f3\u8abf(pitch)\uff0c\u4e26\u5c0d 16 \u7dad\u8a9e\u97f3\u7279\u5fb5\u53d6\u76f8\u5c0d\u7684\u4e00\u7686\u5dee\u91cf\u4fc2\u6578(delta coefficient)\u548c\u4e8c\u7686 \u5dee\u91cf\u4fc2\u6578(acceleration coefficient)\u7576\u4f5c DNN \u7684\u8f38\u5165\u7279\u5fb5\u3002\u5148\u900f\u904e GMM \u5c0d\u8a9e\u97f3\u7279\u5fb5\u8a13\u7df4 \u55ae\u9023\u97f3\u7d20(monophone)\u7684\u8072\u5b78\u6a21\u578b\uff0c\u55ae\u97f3\u7bc0\u8207\u96d9\u97f3\u7bc0\u8a9e\u6599\u5eab\u4e2d\u7686\u6709 183 \u500b\u55ae\u9023\u97f3\u7d20(\u8072\u6bcd \u6709 24 \u500b\uff0c\u8072\u8abf\u97fb\u6bcd\u6709 159 \u500b)\uff0c\u63a5\u8457\u4fdd\u7559 GMM \u8a08\u7b97\u51fa\u4f86\u7684\u521d\u59cb\u6a5f\u7387\u3001\u8f49\u79fb\u6a5f\u7387\u8207\u5f37\u5236 \u5c0d\u9f4a\u7684\u8cc7\u8a0a\uff0c\u53d6\u4ee3 GMM\uff0c\u8a13\u7df4\u7522\u751f\u6bcf\u500b\u97f3\u6846\u6240\u5c0d\u61c9 HMM \u72c0\u614b\u7684\u6a5f\u7387\uff0c\u518d\u6839\u64da\u5f37 \u5716\u56db\u3001\u6bd4\u8f03\u55ae\u97f3\u7bc0(\u5de6\u5074)\u8207\u96d9\u97f3\u7bc0(\u53f3\u5074)\u5f9e\u4e0d\u540c HMM \u8403\u53d6\u51fa\u7684\u7279\u5fb5\u4f7f\u7528\u4e0d\u540c \u5206\u985e\u5668\u4e4b Recall-Precision \u66f2\u7dda \u5236\u5c0d\u9f4a\u7684\u8cc7\u8a0a\u53d6\u5f97\u6bcf\u500b\u97f3\u7d20\u6240\u5c0d\u61c9\u5230\u97f3\u6846\u6578\u91cf\uff0c\u4f86\u8a08\u7b97\u6bcf\u500b\u97f3\u7d20\u7684\u4e8b\u5f8c\u6a5f\u7387\uff0c\u7576\u4f5c HMM \u7684\u89c0\u6e2c\u6a5f\u7387\u3002\u5728 CNN \u7684\u8f38\u5165\u7279\u5fb5\u8a2d\u5b9a\u65b9\u9762\uff0c\u6211\u5011\u4f7f\u7528\u5f9e\u6885\u723e\u983b\u8b5c\u4fc2\u6578(mel-scale frequency spectral coefficients, MFSC)\u53d6\u5f97\u7684\u5c0d\u6578\u80fd\u91cf\u7279\u5fb5\u4e26\u900f\u904e\u6ffe\u6ce2\u5668\u7d44(filter banks) \u6240\u7522\u751f\u7684 40 \u7dad\u8f38\u51fa\u4f5c\u70ba CNN \u7684\u8f38\u5165\u8a9e\u97f3\u7279\u5fb5\uff0c\u9130\u8fd1\u97f3\u7a97\u6211\u5011\u63a1\u7528\u524d\u5f8c\u5404 5 \u500b\u97f3\u6846\uff0c \u5171\u542b 11 \u500b\u97f3\u6846\uff0c\u6bcf\u500b\u97f3\u6846\u7686\u70ba 40 \u7dad\u7684 filter banks \u8f38\u51fa\u52a0\u4e0a 3 \u7dad\u5ea6\u97f3\u8abf\u7279\u5fb5\uff0c\u4e26\u5c0d 43 \u7dad\u8a9e\u97f3\u7279\u5fb5\u53d6\u76f8\u5c0d\u7684\u4e00\u7686\u5dee\u91cf\u4fc2\u6578(delta coefficient)\u548c\u4e8c\u7686\u5dee\u91cf\u4fc2\u6578(acceleration coefficient)\uff0c\u5247\u8f38\u5165\u7684\u8a9e\u97f3\u7279\u5fb5\u5c31\u6703\u5f97\u5230 11 \u500b 129 \u7dad\u7684\u7279\u5fb5\u5411\u91cf\u3002\u6211\u5011\u8b93 CNN \u6cbf\u8457\u7279\u5fb5\u983b \u7387\u8ef8\u505a\u647a\u7a4d\uff0c\u4e26\u4f7f\u7528 2 \u5c64\u7684 CNN\uff0c\u53d6\u4ee3 DNN \u4f5c\u70ba\u7279\u5fb5\u62bd\u53d6\u7684\u5de5\u5177\uff0c\u4f7f\u7d93\u904e\u7db2\u8def\u5f97\u5230\u7684 \u4e8b\u5f8c\u6a5f\u7387\u5bcc\u542b\u767c\u97f3\u9451\u5225\u529b\u7684\u8cc7\u8a0a\u3002 CNN(a)\u548c(b)\u4f7f\u7528 40 \u7dad\u5ea6 filter banks \u7279\u5fb5\u52a0\u4e0a 3 \u7dad\u5ea6\u97f3\u8abf\u7279\u5fb5\u3002\u5728 DNN \u8207 CNN \u7684\u96b1\u85cf\u5c64\u6578\u91cf\u8207\u5404\u5c64\u795e\u7d93\u5143\u6578\u91cf\u7684\u9078\u64c7\u4e2d\uff0cDNN \u4f7f\u7528\u57fa\u672c\u7684 4 \u5c64\u96b1\u85cf\u5c64\uff0c\u5404\u5c64\u6709 1024 \u500b\u795e\u7d93\u5143\uff1bCNN(a)\u4f7f\u7528 2 \u7d44 CNN \u5c64\u52a0\u4e0a 2 \u5c64\u5404\u6709 512 \u500b\u795e\u7d93\u5143\u7684 DNN \u96b1\u85cf \u5c64\uff1bCNN(b)\u4f7f\u7528 2 \u7d44 CNN \u5c64\u52a0\u4e0a 2 \u5c64\u5404\u6709 1024 \u500b\u795e\u7d93\u5143\u7684 DNN \u96b1\u85cf\u5c64\u3002\u7531\u65bc\u672c\u8ad6 DNN)\u6240\u7522\u751f\u7684\u767c\u97f3\u6aa2\u6e2c\u7279\u5fb5\u4e0a\u7684\u8868\u73fe\u5341\u5206\u63a5\u8fd1\u3002\u82e5\u6bd4\u8f03\u5728 DNN \u8072\u5b78\u6a21\u578b\u4e2d\u4e0d\u540c\u5206\u985e\u5668\u7684\u6539\u5584\uff0cLR \u5247\u662f\u52dd\u904e GOP \u7d04 12.60%\u7684\u5927\u5e45\u5ea6\u6539\u9032\u3002\u5728\u96d9 \u97f3\u7bc0\u4e2d\u6574\u9ad4\u8868\u73fe\u4e0d\u5982\u55ae\u97f3\u7bc0\u4f86\u7684\u512a\u79c0\uff0cGOP \u7684\u66f2\u7dda\u5728 DNN \u8207 CNN \u4e2d\u53ea\u5f97\u5230 36.03%\u8207 35.15%\uff0c\u662f\u975e\u5e38\u4e0d\u53ef\u9760\u7684\u5206\u985e\u5668\uff0c\u4f46\u662f\u96d9\u97f3\u7bc0(\u5716\u56db\u53f3)\u7684 DNN \u8072\u5b78\u6a21\u578b\u5728\u5206\u985e\u5668 LR \u4e2d \u7684\u8868\u73fe\u76f8\u8f03\u65bc GOP \u63d0\u5347\u7d04 18.44%\uff0c\u9032\u6b65\u7684\u5e45\u5ea6\u6bd4\u55ae\u97f3\u7bc0\u66f4\u70ba\u5287\u70c8\uff0c\u56e0\u6b64\u6211\u5011\u53ef\u4ee5\u5f9e\u5716 \u7684\u4e0d\u540c\u7279\u5fb5\u7528\u65bc\u5206\u985e\u5668 LR \u4e4b\u7d50\u5408(67.23%)\u3002\u96d9\u97f3\u7bc0\u7684\u8868\u73fe\u5247\u662f\u76f8\u53cd\uff0c\u5728\u5206 \u985e\u5668 SVM \u8207 LR \u7d50\u5408\u7684\u6210\u6548\u52dd\u904e\u6a21\u578b DNN \u8207 CNN \u7684\u7d50\u5408\uff0c\u7d50\u679c\u5206\u5225\u70ba 58.54%(\u5982\u5716 \u516d\u53f3)\u8207 55.98%(\u5982\u5716\u4e94\u53f3)\u3002 \u9664\u4e86\u63a2\u8a0e\u5728 Recall-Precision \u66f2\u7dda\u7684\u8868\u73fe\u5916\uff0c\u6211\u5011\u4e5f\u5728\u5716\u4e03\u8207\u5716\u516b\u500b\u5225\u5217\u51fa\u55ae\u97f3\u7bc0 \u8207\u96d9\u97f3\u7bc0\u5728 ROC \u66f2\u7dda\u4e0a\u7684\u8868\u73fe\uff0c\u800c ROC \u7a7a\u9593\u503c\u500b\u6578\u53ef\u5206\u70ba\u56db\u7a2e\uff1a\u771f\u967d\u6027(true positive, TP)\uff1a\u7cfb\u7d71\u63a8\u6e2c\u70ba\u6b63\u78ba\u767c\u97f3\uff0c\u5be6\u969b\u4e0a\u4e5f\u662f\u6b63\u78ba\u767c\u97f3\uff1b\u771f\u9670\u6027(true negative, TN)\uff1a\u7cfb\u7d71\u63a8\u6e2c \u70ba\u932f\u8aa4\u767c\u97f3\uff0c\u5be6\u969b\u4e0a\u4e5f\u662f\u932f\u8aa4\u767c\u97f3\uff1b\u507d\u967d\u6027(false positive, FP)\uff1a\u7cfb\u7d71\u63a8\u6e2c\u70ba\u6b63\u78ba\u767c\u97f3\uff0c \u5be6\u969b\u4e0a\u70ba\u932f\u8aa4\u767c\u97f3\uff1b\u507d\u9670\u6027(false negative, FN)\uff1a\u7cfb\u7d71\u63a8\u6e2c\u70ba\u932f\u8aa4\u767c\u97f3\uff0c\u5be6\u969b\u4e0a\u70ba\u6b63\u78ba\u767c \u97f3\uff0c\u800c\u5716\u4e03\u8207\u5716\u516b\u7686\u662f\u85c9\u7531\u8abf\u6574\u9580\u6abb\u503c\u5f97\u5230\u4e0d\u540c\u7684\u771f\u967d\u6027\u7387(true positive rate, TPR)\u8207\u507d \u967d\u6027\u7387(false positive rate, FPR)\u6240\u7e6a\u88fd\u800c\u6210\u7684\u66f2\u7dda\uff0cTPR \u8207 FPR \u7684\u8a08\u7b97\u65b9\u5f0f\u5982\u5f0f(26)\u8207\u5f0f \u55ae\u97f3\u7bc0\u66f2\u7dda\u4e0b\u9762\u7a4d(area under the curve of ROC, AUC)\u986f\u793a\u65bc\u8868\u4e09\uff0c\u5be6\u9a57\u4e2d\u7684 AUC \u662f\u4f7f \u7528\u68af\u5f62\u6cd5(trapezoid method)\u6c42\u5f97\uff0c\u5f9e DNN \u7d50\u5408 CNN \u7684\u6a21\u578b(\u5982\u8868\u4e09\u7684 C+D)\u4e4b\u8868\u73fe\u4f86 \u770b\uff0cTP \u8207 TN \u90fd\u6709\u6240\u63d0\u5347\uff0c\u800c FP \u8207 FN \u4e5f\u6709\u660e\u986f\u7684\u4e0b\u964d\uff0cAUC \u7684\u90e8\u5206\u76f8\u8f03 DNN \u7684 84.42%\u5247\u9032\u6b65\u4e86 2.28%\u9054\u5230\u7d04 86.70%\uff0c\u5f9e\u5716\u4e03\u4e2d\u53ef\u4ee5\u6e05\u695a\u770b\u5230 DNN \u8207 CNN \u7d50\u5408\u4e4b\u8072 \u5b78\u6a21\u578b\u7684\u8868\u73fe\u512a\u65bc CNN \u8207 DNN \u5404\u81ea\u4f7f\u7528\uff1b\u6211\u5011\u5c07\u5716\u4e03\u5de6\u4e0a\u8207\u53f3\u4e0b\u4e4b\u5c0d\u89d2\u7dda\u76f8\u9023\u6c42\u51fa \u5716\u516d\u3001\u6bd4\u8f03\u96d9\u97f3\u7bc0(\u5de6\u5074)\u8207\u55ae\u97f3\u7bc0(\u53f3\u5074)\u5f9e CNN-DNN \u8403\u53d6\u51fa\u7684\u7279\u5fb5\u4f7f\u7528\u4e0d\u540c \u5206\u985e\u5668(SVM \u8207 LR)\u8f38\u51fa\u5206\u6578\u7684\u7dda\u6027\u7d44\u5408(M/DS-CNN-SVM+LR)\u4e4b Recall-Precision \u66f2\u7dda TPR \u8207\u5047 FPR \u76f8\u52a0\u8da8\u8fd1\u65bc 1 \u7684\u9ede\uff0c\u8a72\u9ede\u6240\u8868\u793a\u7684 FPR \u7a31\u4f5c\u76f8\u7b49\u932f\u8aa4\u7387(equal error rate, EER)\uff0c\u5728\u8868\u4e09\u7684 ROC \u7a7a\u9593\u503c\u500b\u6578(TP\u3001TN\u3001FP\u3001FN)\u662f\u5229\u7528\u8a72\u9ede\u6c42\u5f97\uff1b\u5716\u4e03\u53ef\u4ee5\u767c\u73fe DNN \u7684 EER \u70ba 21.67%\uff0c\u800c\u7d93\u904e\u7d50\u5408\u7684\u6a21\u578b(\u5982\u5716\u4e03\u7684 C+D)\u53ef\u964d\u4f4e\u81f3 19.94%\u3002\u5728\u8868\u56db \u8207\u5716\u516b\u5247\u662f\u986f\u793a\u96d9\u97f3\u7bc0\u7684 ROC \u7a7a\u9593\u503c\u500b\u6578\u3001ROC \u66f2\u7dda\u8207 EER\uff0c\u5716\u516b\u53ef\u4ee5\u767c\u73fe DNN \u7684 EER \u70ba 24.30%\uff0c\u800c\u7d93\u904e\u7d50\u5408\u7684\u6a21\u578b(\u5982\u5716\u516b\u7684 C+D)\u53ef\u964d\u4f4e\u81f3 23.08%\uff1b\u8868\u56db\u5728 AUC \u7684 \u90e8\u5206\u53ef\u4ee5\u89c0\u5bdf\u5230\u6a21\u578b\u7d50\u5408\u5f8c\u5f9e CNN \u7684 80.90%\u9032\u6b65\u5230 82.58%\u3002 \u516d\u3001 \u7d50\u8ad6\u8207\u672a\u4f86\u7814\u7a76\u5c55\u671b \u672c\u8ad6\u6587\u63a2\u8a0e\u5169\u7a2e\u8072\u5b78\u6a21\u578b(DNN \u548c CNN)\u4ee5\u53ca\u5b83\u5011\u7684\u7d50\u5408\u5c0d\u65bc\u767c\u97f3\u6aa2\u6e2c\u6548\u80fd\u7684\u5f71\u97ff\u3002 \u53e6\u4e00\u65b9\u9762\uff0c\u5f9e\u5be6\u9a57\u7d50\u679c\u53ef\u4ee5\u767c\u73fe\uff0c\u672c\u8ad6\u6587\u6240\u4f7f\u7528\u7684\u4e09\u7a2e\u5206\u985e\u65b9\u6cd5(GOP\u3001SVM \u548c LR)\u4e2d \u7121\u8ad6\u662f\u55ae\u97f3\u7bc0\u6216\u96d9\u97f3\u7bc0\u7686\u4ee5 LR \u8868\u73fe\u6700\u4f73\u3002\u96d6\u7136 DNN-LR \u8207 CNN-LR \u5169\u7a2e\u932f\u8aa4\u767c\u97f3\u6aa2 \u6e2c\u6a21\u578b\u4e4b\u8868\u73fe\u5341\u5206\u76f8\u8fd1\uff0c\u4f46\u7d93\u904e\u7c21\u55ae\u7684\u7dda\u6027\u7d44\u5408\u5f8c\u4f9d\u7136\u53ef\u4ee5\u5728\u55ae\u97f3\u7bc0\u932f\u8aa4\u767c\u97f3\u6aa2\u6e2c\u7684\u53ec \u56de\u7387\u8207\u7cbe\u6e96\u5ea6\u76f8\u540c\u6642\u5f97\u5230 3.42%\u7684\u9032\u6b65\u4e26\u9054\u5230 67.23%\u7684\u8868\u73fe\uff1b\u540c\u6642\uff0c\u5728\u96d9\u97f3\u7bc0\u932f\u8aa4\u767c\u97f3 \u6aa2\u6e2c\u4e0a\uff0c\u7d93\u904e\u7dda\u6027\u7d44\u5408\u5f8c\u4e5f\u5f97\u5230 1.51%\u7684\u9032\u6b65\u4e26\u63d0\u5347\u81f3 55.98%\u3002\u800c ROC \u66f2\u7dda\u5728\u55ae\u97f3\u7bc0 \u8868\u56db\u3001\u96d9\u97f3\u7bc0\u5728\u5206\u985e\u5668 LR \u5728\u4e0d\u540c\u8072\u5b78\u6a21 \u578b\u7684 ROC \u7a7a\u9593\u503c(TP\u3001TN\u3001FP \u8207 FN)\u8207" | |
| } | |
| } | |
| } | |
| } |