| { |
| "paper_id": "2018", |
| "header": { |
| "generated_with": "S2ORC 1.0.0", |
| "date_generated": "2023-01-19T07:26:56.345680Z" |
| }, |
| "title": "Leveraging Discriminative Training and Model Combination for Semi-supervised Speech Recognition", |
| "authors": [ |
| { |
| "first": "\u7f85\u5929\u5b8f", |
| "middle": [ |
| "\uf02a" |
| ], |
| "last": "\u3001\u9673\u67cf\u7433 \uf02a", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "National Taiwan Normal University", |
| "location": {} |
| }, |
| "email": "" |
| }, |
| { |
| "first": "Tien-Hong", |
| "middle": [], |
| "last": "Lo", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "National Taiwan Normal University", |
| "location": {} |
| }, |
| "email": "teinhonglo@ntnu.edu.tw" |
| }, |
| { |
| "first": "Berlin", |
| "middle": [], |
| "last": "Chen", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "National Taiwan Normal University", |
| "location": {} |
| }, |
| "email": "berlin@ntnu.edu.tw" |
| } |
| ], |
| "year": "", |
| "venue": null, |
| "identifiers": {}, |
| "abstract": "\u8fd1\u5e74\u4f86\u9451\u5225\u5f0f\u8a13\u7df4(Discriminative training)\u7684\u76ee\u6a19\u51fd\u6578 Lattice-free Maximum Mutual Information (LF-MMI)\u5728\u81ea\u52d5\u8a9e\u97f3\u8fa8\u8b58(Automatic speech recognition, ASR)\u4e0a\u53d6\u5f97\u4e86\u91cd\u5927\u7684\u7a81\u7834\u3002\u5118\u7ba1 LF-MMI \u5728\u76e3\u7763\u5f0f\u74b0\u5883\u4e0b\u65ac\u7372\u6700\u597d\u7684\u6210\u679c\uff0c \u7136\u800c\u5728\u534a\u76e3\u7763\u5f0f\u8a2d\u5b9a\u4e0b\uff0c\u7531\u65bc\u7a2e\u5b50\u6a21\u578b(Seed model)\u5e38\u56e0\u70ba\u8a9e\u6599\u6709\u9650\u800c\u6548\u679c\u4e0d \u4f73\u3002\u4e14\u7531\u65bc LF-MMI \u5c6c\u65bc\u9451\u5225\u5f0f\u8a13\u7df4\u4e4b\u6545\uff0c\u6613\u53d7\u5230\u8f49\u5beb\u6b63\u78ba\u8207\u5426\u7684\u5f71\u97ff\u3002\u672c\u8ad6 \u6587\u5229\u7528\u5169\u7a2e\u601d\u8def\u65bc\u534a\u76e3\u7763\u5f0f\u8a13\u7df4\u3002\u5176\u4e00\uff0c\u5f15\u5165\u8ca0\u689d\u4ef6\u71b5(Negative conditional entropy, NCE) \u6b0a \u91cd \u8207 \u8a5e \u5716 (Lattice) \uff0c \u524d \u8005 \u662f \u6700 \u5c0f \u5316 \u8a5e \u5716 \u8def \u5f91 \u7684 \u689d \u4ef6 \u71b5 (Conditional entropy)\uff0c\u7b49\u540c\u5c0d MMI \u7684\u53c3\u8003\u8f49\u5beb(Reference transcript)\u505a\u6b0a\u91cd\u5e73\u5747\uff0c \u6b0a\u91cd\u7684\u6539\u8b8a\u80fd\u81ea\u7136\u5730\u52a0\u5165 MMI \u8a13\u7df4\u4e2d\uff0c\u4e26\u540c\u6642\u5c0d\u4e0d\u78ba\u5b9a\u6027\u5efa\u6a21\u3002\u5176\u76ee\u7684\u5e0c\u671b \u7121\u4fe1\u5fc3\u904e\u6ffe\u5668(Confidence-based filter)\u4e5f\u53ef\u8a13\u7df4\u6a21\u578b\u3002\u5f8c\u8005\u52a0\u5165\u8a5e\u5716\uff0c\u6bd4\u8d77\u904e\u5f80 \u7684\u53ea\u4f7f\u7528\u6700\u4f73\u8fa8\u8b58\u7d50\u679c\uff0c\u53ef\u4fdd\u7559\u66f4\u591a\u5047\u8aaa\u7a7a\u9593\uff0c\u9032\u800c\u63d0\u5347\u627e\u5230\u53c3\u8003\u8f49\u5beb (Reference transcript)\u7684\u53ef\u80fd\u6027\uff1b\u5176\u4e8c\uff0c\u6211\u5011\u501f\u9452\u6574\u9ad4\u5b78\u7fd2(Ensemble learning) \u7684\u6982\u5ff5\uff0c\u4f7f\u7528\u5f31\u5b78\u7fd2\u5668(Weak learner)\u4fee\u6b63\u5f7c\u6b64\u7684\u932f\u8aa4\uff0c\u5206\u70ba\u5047\u8aaa\u5c64\u7d1a\u5408\u4f75 (Hypothesis-level combination)\u548c\u97f3\u6846\u5c64\u7d1a\u5408\u4f75(Frame-level combination)\u3002\u5be6\u9a57 \u7d50\u679c\u986f\u793a\uff0c\u52a0\u5165 NCE \u8207\u8a5e\u5716\u7686\u80fd\u964d\u4f4e\u8a5e\u932f\u8aa4\u7387(Word error rate, WER)\uff0c\u800c\u6a21\u578b \u5408\u4f75(Model combination)\u5247\u80fd\u5728\u5404\u500b\u968e\u6bb5\u986f\u8457\u63d0\u5347\u6548\u80fd\uff0c\u4e14\u5169\u8005\u7d50\u5408\u53ef\u4f7f\u8a5e\u4fee \u5fa9\u7387(WER recovery rate, WRR)\u9054\u5230 60.8%\u3002", |
| "pdf_parse": { |
| "paper_id": "2018", |
| "_pdf_hash": "", |
| "abstract": [ |
| { |
| "text": "\u8fd1\u5e74\u4f86\u9451\u5225\u5f0f\u8a13\u7df4(Discriminative training)\u7684\u76ee\u6a19\u51fd\u6578 Lattice-free Maximum Mutual Information (LF-MMI)\u5728\u81ea\u52d5\u8a9e\u97f3\u8fa8\u8b58(Automatic speech recognition, ASR)\u4e0a\u53d6\u5f97\u4e86\u91cd\u5927\u7684\u7a81\u7834\u3002\u5118\u7ba1 LF-MMI \u5728\u76e3\u7763\u5f0f\u74b0\u5883\u4e0b\u65ac\u7372\u6700\u597d\u7684\u6210\u679c\uff0c \u7136\u800c\u5728\u534a\u76e3\u7763\u5f0f\u8a2d\u5b9a\u4e0b\uff0c\u7531\u65bc\u7a2e\u5b50\u6a21\u578b(Seed model)\u5e38\u56e0\u70ba\u8a9e\u6599\u6709\u9650\u800c\u6548\u679c\u4e0d \u4f73\u3002\u4e14\u7531\u65bc LF-MMI \u5c6c\u65bc\u9451\u5225\u5f0f\u8a13\u7df4\u4e4b\u6545\uff0c\u6613\u53d7\u5230\u8f49\u5beb\u6b63\u78ba\u8207\u5426\u7684\u5f71\u97ff\u3002\u672c\u8ad6 \u6587\u5229\u7528\u5169\u7a2e\u601d\u8def\u65bc\u534a\u76e3\u7763\u5f0f\u8a13\u7df4\u3002\u5176\u4e00\uff0c\u5f15\u5165\u8ca0\u689d\u4ef6\u71b5(Negative conditional entropy, NCE) \u6b0a \u91cd \u8207 \u8a5e \u5716 (Lattice) \uff0c \u524d \u8005 \u662f \u6700 \u5c0f \u5316 \u8a5e \u5716 \u8def \u5f91 \u7684 \u689d \u4ef6 \u71b5 (Conditional entropy)\uff0c\u7b49\u540c\u5c0d MMI \u7684\u53c3\u8003\u8f49\u5beb(Reference transcript)\u505a\u6b0a\u91cd\u5e73\u5747\uff0c \u6b0a\u91cd\u7684\u6539\u8b8a\u80fd\u81ea\u7136\u5730\u52a0\u5165 MMI \u8a13\u7df4\u4e2d\uff0c\u4e26\u540c\u6642\u5c0d\u4e0d\u78ba\u5b9a\u6027\u5efa\u6a21\u3002\u5176\u76ee\u7684\u5e0c\u671b \u7121\u4fe1\u5fc3\u904e\u6ffe\u5668(Confidence-based filter)\u4e5f\u53ef\u8a13\u7df4\u6a21\u578b\u3002\u5f8c\u8005\u52a0\u5165\u8a5e\u5716\uff0c\u6bd4\u8d77\u904e\u5f80 \u7684\u53ea\u4f7f\u7528\u6700\u4f73\u8fa8\u8b58\u7d50\u679c\uff0c\u53ef\u4fdd\u7559\u66f4\u591a\u5047\u8aaa\u7a7a\u9593\uff0c\u9032\u800c\u63d0\u5347\u627e\u5230\u53c3\u8003\u8f49\u5beb (Reference transcript)\u7684\u53ef\u80fd\u6027\uff1b\u5176\u4e8c\uff0c\u6211\u5011\u501f\u9452\u6574\u9ad4\u5b78\u7fd2(Ensemble learning) \u7684\u6982\u5ff5\uff0c\u4f7f\u7528\u5f31\u5b78\u7fd2\u5668(Weak learner)\u4fee\u6b63\u5f7c\u6b64\u7684\u932f\u8aa4\uff0c\u5206\u70ba\u5047\u8aaa\u5c64\u7d1a\u5408\u4f75 (Hypothesis-level combination)\u548c\u97f3\u6846\u5c64\u7d1a\u5408\u4f75(Frame-level combination)\u3002\u5be6\u9a57 \u7d50\u679c\u986f\u793a\uff0c\u52a0\u5165 NCE \u8207\u8a5e\u5716\u7686\u80fd\u964d\u4f4e\u8a5e\u932f\u8aa4\u7387(Word error rate, WER)\uff0c\u800c\u6a21\u578b \u5408\u4f75(Model combination)\u5247\u80fd\u5728\u5404\u500b\u968e\u6bb5\u986f\u8457\u63d0\u5347\u6548\u80fd\uff0c\u4e14\u5169\u8005\u7d50\u5408\u53ef\u4f7f\u8a5e\u4fee \u5fa9\u7387(WER recovery rate, WRR)\u9054\u5230 60.8%\u3002", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Abstract", |
| "sec_num": null |
| } |
| ], |
| "body_text": [ |
| { |
| "text": "\u8fd1 \u5e74 \u4f86 \u57fa \u65bc \u985e \u795e \u7d93 \u7db2 \u8def \u7684 \u8072 \u5b78 \u6a21 \u578b (Deep neural network-hidden Markov model, DNN-HMM)\u53d6\u5f97\u91cd\u5927\u7684\u7a81\u7834 (Seide, Li & Yu, 2011) (Dahl, Yu, Deng & Acero, 2012 )\u3002\u50b3\u7d71 \u7684 DNN-HMM \u900f\u904e\u4ea4\u4e92\u71b5\u8a13\u7df4(Cross-Entropy training, CE)\u548c\u9451\u5225\u5f0f\u8a13\u7df4(Discriminative training) (Valtchev, Odell, Woodland & Young, 1996) (Valtchev, Odell, Woodland & Young, 1997) ) \uff0c\u5169\u968e\u6bb5\u7684\u8a13\u7df4\u63d0\u5347\u8072\u5b78\u6a21\u578b\u7684\u8fa8\u8b58\u6548\u679c\u3002\u5c24\u5176\u662f\u7b2c\u4e8c \u7d50\u5408\u9451\u5225\u5f0f\u8a13\u7df4\u8207\u6a21\u578b\u5408\u4f75\u65bc\u534a\u76e3\u7763\u5f0f\u8a9e\u97f3\u8fa8\u8b58\u4e4b\u7814\u7a76 21 \u968e\u6bb5\u7684\u9451\u5225\u5f0f\u8a13\u7df4\uff0c\u7531\u65bc\u63d0\u5347\u6548\u679c\u986f\u8457\uff0c\u5438\u5f15\u4e86\u8a31\u591a\u7814\u7a76\u8005\u7684\u76ee\u5149\u3002\u904e\u5f80\u65bc\u9451\u5225\u5f0f\u8a13\u7df4 \u7684\u7814\u7a76\u4e3b\u984c\u7a2e\u985e\u7e41\u591a\uff0c\u5982 MMI (Bahl, Brown, de Souza & Mercer, 1986) , MCE (Juang, Hou & Lee, 1997) , MPE , sMBR (Kaiser, Horvat & Kacic, 2000) (Gibson & Hain, 2006 )\u548c bMMI (Povey et al, 2008) \u7b49\u3002\u6700\u8fd1\uff0c\u96a8\u8457\u8a9e\u6599\u7684\u589e\u9577\uff0c\u4e0d\u900f\u904e\u7b2c \u4e00\u968e\u6bb5\u7684 CE \u8a13\u7df4\uff0c\u5c07\u9451\u5225\u5f0f\u8a13\u7df4\u505a\u4e00\u968e\u6bb5\u8a13\u7df4\u7684\u7aef\u5c0d\u7aef\u8a13\u7df4(End-to-End)\u4e5f\u8d8a\u4f86\u8d8a\u6d41\u884c\u3002 \u76ee \u524d \u5169 \u7a2e \u4e3b \u6d41 \u7684 \u7aef \u5c0d \u7aef \u67b6 \u69cb \u7684 \u76ee \u6a19 \u51fd \u6578 \u70ba CTC (Graves, Fern\u00e1ndez, Gomez & Schmidhuber, 2006 )\u548c Lattice-free MMI (LF-MMI) (Povey et al., 2016) \u3002\u524d\u8005\u5728\u8a9e\u6599\u975e\u5e38\u5145 \u8db3(\u901a\u5e38\u5927\u65bc 500 \u5c0f\u6642)\u7684\u60c5\u6cc1\u4e0b\uff0c\u8868\u73fe\u53ef\u4ee5\u5ab2\u7f8e\u751a\u81f3\u8d85\u8d8a\u50b3\u7d71\u7684\u4e8c\u968e\u6bb5\u65b9\u6cd5\u3002\u800c\u5f8c\u8005\u8b49 \u660e\u4e86\u5728\u8a9e\u6599\u8f03\u70ba\u7f3a\u4e4f\u7684\u60c5\u6cc1\u4e0b\uff0c\u5118\u7ba1\u6548\u80fd\u6703\u4e0b\u964d\uff0c\u4f46\u4ecd\u53ef\u4ee5\u52dd\u904e\u524d\u8005\uff0c\u56e0\u6b64\u6210\u70ba\u4e86\u76ee\u524d \u6700\u5177\u9b45\u529b\u7684\u7814\u7a76\u4e3b\u984c\u3002\u5728 (Povey et al., 2016) \u7684\u5be6\u9a57\u4e2d\u5c55\u793a\uff0c\u57fa\u65bc LF-MMI \u7684\u76ee\u6a19\u51fd\u6578\u4e0b\uff0c \u53ef\u5f9e\u4e82\u6578\u521d\u59cb\u5316\u53c3\u6578\u5f8c\uff0c\u4ee5\u9451\u5225\u5f0f\u6e96\u5247\u8a13\u7df4\u985e\u795e\u7d93\u7db2\u8def\u3002\u5be6\u9a57\u7d50\u679c\u986f\u793a LF-MMI \u6548\u679c\u66f4 \u52dd\u5169\u968e\u6bb5\u8a13\u7df4\u7684 sMBR \u4e00\u7c4c\uff0c\u4e14\u9084\u53ef\u7d50\u5408 sMBR \u9032\u4e00\u6b65\u63d0\u5347\u8fa8\u8b58\u7d50\u679c\u3002\u7136\u800c\uff0c\u5728\u9019\u6a23\u7684 \u8a13\u7df4\u6e96\u5247\u4e0b\uff0c\u4ecd\u53d7\u9650\u65bc\u9700\u5927\u91cf\u8a13\u7df4\u8a9e\u6599\u7684\u554f\u984c(Data hungry)\u3002\u9032\u4e00\u6b65\u4f86\u8aaa\uff0c\u4fbf\u662f\u5728\u5c0f\u8a9e \u6599\u5eab\u4e0a\u7684\u8868\u73fe(\u901a\u5e38\u5c0f\u65bc 100 \u5c0f\u6642)\u4ecd\u7121\u6cd5\u52dd\u904e\u5728\u5927\u8a9e\u6599\u5eab\u7684\u512a\u7570\u7d50\u679c (Pundak & Sainath, 2016 )\u3002 \u5728\u73fe\u5be6\u751f\u6d3b\u4e2d\uff0c\u76f8\u5c0d\u65bc\u9ad8\u6210\u672c\u7684\u4eba\u5de5\u8f49\u5beb\u8a9e\u6599\uff0c\u672a\u8f49\u5beb\u8a9e\u6599\u5341\u5206\u5bb9\u6613\u53d6\u5f97\u3002\u7576\u6211\u5011 \u6c92\u8fa6\u6cd5\u53d6\u5f97\u5927\u91cf\u7684\u8f49\u5beb\u8a9e\u6599\u6642\uff0c\u5c31\u5fc5\u9808\u66f4\u6709\u6548\u5730\u5229\u7528\u5927\u91cf\u7684\u672a\u8f49\u5beb\u8a9e\u6599\u8a13\u7df4\u6a21\u578b\u3002\u63db\u53e5 \u8a71\u8aaa\uff0c\u63a2\u7d22\u5b58\u5728\u65bc\u672a\u8f49\u5beb\u8a9e\u6599\u7684\u7dda\u7d22\uff0c\u4e26\u52a0\u5165\u534a\u76e3\u7763\u5f0f\u8a13\u7df4\u7684\u8072\u5b78\u6a21\u578b\u5c31\u66f4\u986f\u91cd\u8981\u3002\u53e6 \u4e00\u65b9\u9762\uff0c\u534a\u76e3\u7763\u5f0f\u8a13\u7df4\u7528\u9014\u591a\u5143\uff0c\u4e0d\u50c5\u53ef\u7528\u5728\u81ea\u52d5\u8a9e\u97f3\u8fa8\u8b58\u7684\u8a13\u7df4\uff0c\u4e5f\u540c\u6a23\u9069\u7528\u65bc\u81ea\u52d5 \u8f49\u5beb(Automatic labeling)\u53ca\u9077\u79fb\u5b78\u7fd2(Transfer learning)\u3001\u8a9e\u8005\u8abf\u9069(Speaker adaptation)\u3002\u904e \u5f80\u7684\u7814\u7a76\u65bc\u534a\u76e3\u7763\u5f0f\u8072\u5b78\u6a21\u578b (Zavaliagkos, Siu, Colthurst & Billa, 1998) \uff0c\u6700\u5e38\u898b\u7684\u8a13\u7df4\u65b9 \u6cd5\u662f\u81ea\u6211\u8a13\u7df4(Self-training) (Vesely, Hannemann & Burget, 2013) (Grezl & Karafi\u00e1t, 2013) (Zhang, Liu & Hain, 2014) \u3002\u81ea\u6211\u8a13\u7df4\u7684\u67b6\u69cb\u4e3b\u8981\u5206\u6210\u5169\u968e\u6bb5\uff0c\u7b2c\u4e00\u968e\u6bb5\u70ba\u5229\u7528\u8f49\u5beb\u8a9e\u6599 \u8a13\u7df4\u7a2e\u5b50\u6a21\u578b\u76f4\u5230\u7a69\u5b9a\uff0c\u7b2c\u4e8c\u968e\u6bb5\u5247\u662f\u5229\u7528\u7a2e\u5b50\u6a21\u578b\u8fa8\u8b58\u672a\u8f49\u5beb\u8a9e\u6599\uff0c\u4e26\u4ee5\u6b64\u70ba\u7b54\u6848\u91cd \u65b0\u8a13\u7df4\u6a21\u578b\u3002\u5728\u7b2c\u4e8c\u968e\u6bb5\u7684\u8fa8\u8b58\u7d50\u679c\u8207\u771f\u5be6\u7b54\u6848\u96e3\u514d\u6703\u6709\u8aa4\u5dee\uff0c\u56e0\u6b64\u6703\u518d\u52a0\u5165\u4fe1\u5fc3\u904e\u6ffe \u5668(Confidence-based filter) (Lamel, Gauvain & Adda, 2002 ) (Chan & Woodland, 2004 (Liu, Chu, Lin & Chen, 2007) \u6311\u9078\u8a13\u7df4\u8a9e\u6599\uff0c\u8a72\u52d5\u4f5c\u53ef\u5728\u4e0d\u540c\u5c64\u7d1a\u4e0a\u9032\u884c\uff0c\u5206\u70ba\u97f3\u6846\u5c64\u7d1a (Vesely, Hannemann & Burget, 2013) \u3001\u8a5e\u5c64\u7d1a (Thomas, Seltzer, Church & Hermansky, 2013) \u4ee5\u53ca\u8a9e\u53e5\u5c64\u7d1a (Grezl & Karafi\u00e1t, 2013) (Vesely et al., 2013) (Zhang et al., 2014) (Mathias, Yegnanarayanan & Fritsch, 2005) (Yu, Gales, Wang & Woodland, 2010) (Cui, Huang & Chien, 2011 )\uff0c\u800c\u5c6c\u65bc\u9451\u5225\u5f0f\u8a13\u7df4\u7684 LF-MMI \u4e5f\u540c\u6a23\u5c0d\u65bc\u6b63\u78ba\u6027\u5341\u5206\u654f\u611f\u3002\u7136\u800c\uff0c\u5728\u534a\u76e3 \u7763\u5f0f\u8a13\u7df4\u904e\u7a0b\u4e2d\uff0c\u7531\u65bc\u7b2c\u4e8c\u968e\u6bb5\u8a13\u7df4\u6642\u7121\u6cd5\u4fdd\u8b49\u8a9e\u53e5\u7684\u6b63\u78ba\u6027\uff0c\u56e0\u6b64\u5728\u904e\u5f80\u7814\u7a76\u5e38\u8457\u91cd \u65bc\u4e8c\u968e\u6bb5\u9451\u5225\u5f0f\u8a13\u7df4\u524d\u7684\u4fe1\u5fc3\u904e\u6ffe\u5668\uff0c\u5982 (Liu et al., 2007) (Mathias et al., 2005) \u5c07\u97f3\u6846\u5c64 \u7f85\u5929\u5b8f\u8207\u9673\u67cf\u7433 \u7d1a\u7684\u4fe1\u5fc3\u904e\u6ffe\u5668\u52a0\u5165\u9451\u5225\u5f0f\u8a13\u7df4\u3002\u800c\u5728 (Walker, Pedersen, Orife & Flaks, 2017 ) \u52a0\u5165\u8a9e\u53e5 \u5c64\u7d1a\u7684\u4fe1\u5fc3\u904e\u6ffe\u5668\u4ee5\u53ca\u5f8c\u8655\u7406\u6700\u4f73\u8fa8\u8b58\u7d50\u679c(One-best result)\u3002\u672c\u8ad6\u6587\u8207 (Manohar, Hadian, Povey & Khudanpur, 2018) (Manohar, Povey & Khudanpur, 2015) (Fiscus, 1997) (Evermann & Woodland, 2000) (Deng & Platt, 2014) (Xu, Povey, Mangu & Zhu, 2011 (Zavaliagkos et al., 1998) (Vesely et al., 2013) (Grezl & Karafi\u00e1t, 2013) (Zhang et al., 2014) \u3002\u81ea\u6211\u8a13\u7df4\u7684\u6b65\u9a5f\u5206\u70ba\u5169\u968e\u6bb5\uff0c\u9996\u5148\u4f7f\u7528\u8f49\u5beb\u8a9e\u6599\u8a13\u7df4\u7a2e\u5b50\u6a21\u578b\u76f4\u5230 \u7a69\u5b9a(\u901a\u5e38\u70ba CE \u8a13\u7df4\uff0c\u4f46\u4e5f\u53ef\u52a0\u5165\u9451\u5225\u5f0f\u8a13\u7df4)\uff0c\u7b2c\u4e8c\u968e\u6bb5\u5247\u5229\u7528\u7a2e\u5b50\u6a21\u578b\u8fa8\u8b58\u672a\u8f49\u5beb\u8a9e \u6599\uff0c\u52a0\u5165\u4fe1\u5fc3\u904e\u6ffe\u5668 (Lamel et al., 2002 ) (Chan & Woodland, 2004 (Liu et al., 2007) \u7be9\u9078 \u8a13\u7df4\u8a9e\u6599\uff0c\u904e\u6ffe\u53ef\u80fd\u6703\u5f71\u97ff\u8a13\u7df4\u7684\u8a9e\u6599\uff0c\u518d\u91cd\u65b0\u8a13\u7df4\u6a21\u578b\u3002\u800c\u4fe1\u5fc3\u904e\u6ffe\u5668(Confidence filter) \u53ef\u5728\u97f3\u6846\u5c64\u7d1a (Vesely et al., 2013) \u3001\u8a5e\u5c64\u7d1a (Thomas et al., 2013) \u3001\u8a9e\u53e5\u5c64\u7d1a (Grezl & Karafi\u00e1t, 2013) (Vesely et al., 2013) (Thomas et al., 2013 )\u591a\u7a2e\u5c64\u7d1a\u4e0a\u9032\u884c\u3002\u7531\u65bc\u9451\u5225\u5f0f\u8a13\u7df4\u5c0d\u65bc\u8a13 \u7df4\u8a9e\u53e5\u7684\u6b63\u78ba\u6027\u5341\u5206\u654f\u611f (Mathias et al., 2005) (Yu et al., 2010 ) (Cui et al., 2011 \uff0c\u56e0\u6b64\u904e \u5f80\u7684\u7814\u7a76\u8457\u91cd\u65bc\u4fe1\u5fc3\u904e\u6ffe\u5668\u7684\u9078\u64c7\u3002\u5728 (Liu et al., 2007) (Mathias et al., 2005) (Nadas, 1983) ", |
| "cite_spans": [ |
| { |
| "start": 88, |
| "end": 110, |
| "text": "(Seide, Li & Yu, 2011)", |
| "ref_id": "BIBREF27" |
| }, |
| { |
| "start": 111, |
| "end": 140, |
| "text": "(Dahl, Yu, Deng & Acero, 2012", |
| "ref_id": "BIBREF0" |
| }, |
| { |
| "start": 223, |
| "end": 264, |
| "text": "(Valtchev, Odell, Woodland & Young, 1996)", |
| "ref_id": "BIBREF30" |
| }, |
| { |
| "start": 265, |
| "end": 306, |
| "text": "(Valtchev, Odell, Woodland & Young, 1997)", |
| "ref_id": "BIBREF31" |
| }, |
| { |
| "start": 417, |
| "end": 455, |
| "text": "(Bahl, Brown, de Souza & Mercer, 1986)", |
| "ref_id": null |
| }, |
| { |
| "start": 462, |
| "end": 486, |
| "text": "(Juang, Hou & Lee, 1997)", |
| "ref_id": "BIBREF11" |
| }, |
| { |
| "start": 500, |
| "end": 530, |
| "text": "(Kaiser, Horvat & Kacic, 2000)", |
| "ref_id": "BIBREF12" |
| }, |
| { |
| "start": 531, |
| "end": 551, |
| "text": "(Gibson & Hain, 2006", |
| "ref_id": "BIBREF6" |
| }, |
| { |
| "start": 560, |
| "end": 579, |
| "text": "(Povey et al, 2008)", |
| "ref_id": "BIBREF23" |
| }, |
| { |
| "start": 687, |
| "end": 732, |
| "text": "(Graves, Fern\u00e1ndez, Gomez & Schmidhuber, 2006", |
| "ref_id": "BIBREF8" |
| }, |
| { |
| "start": 762, |
| "end": 782, |
| "text": "(Povey et al., 2016)", |
| "ref_id": "BIBREF25" |
| }, |
| { |
| "start": 887, |
| "end": 907, |
| "text": "(Povey et al., 2016)", |
| "ref_id": "BIBREF25" |
| }, |
| { |
| "start": 1102, |
| "end": 1125, |
| "text": "(Pundak & Sainath, 2016", |
| "ref_id": "BIBREF26" |
| }, |
| { |
| "start": 1371, |
| "end": 1414, |
| "text": "(Zavaliagkos, Siu, Colthurst & Billa, 1998)", |
| "ref_id": "BIBREF37" |
| }, |
| { |
| "start": 1446, |
| "end": 1480, |
| "text": "(Vesely, Hannemann & Burget, 2013)", |
| "ref_id": "BIBREF32" |
| }, |
| { |
| "start": 1481, |
| "end": 1505, |
| "text": "(Grezl & Karafi\u00e1t, 2013)", |
| "ref_id": "BIBREF9" |
| }, |
| { |
| "start": 1506, |
| "end": 1531, |
| "text": "(Zhang, Liu & Hain, 2014)", |
| "ref_id": "BIBREF38" |
| }, |
| { |
| "start": 1665, |
| "end": 1693, |
| "text": "(Lamel, Gauvain & Adda, 2002", |
| "ref_id": "BIBREF13" |
| }, |
| { |
| "start": 1694, |
| "end": 1718, |
| "text": ") (Chan & Woodland, 2004", |
| "ref_id": null |
| }, |
| { |
| "start": 1719, |
| "end": 1747, |
| "text": "(Liu, Chu, Lin & Chen, 2007)", |
| "ref_id": "BIBREF14" |
| }, |
| { |
| "start": 1775, |
| "end": 1809, |
| "text": "(Vesely, Hannemann & Burget, 2013)", |
| "ref_id": "BIBREF32" |
| }, |
| { |
| "start": 1815, |
| "end": 1858, |
| "text": "(Thomas, Seltzer, Church & Hermansky, 2013)", |
| "ref_id": "BIBREF29" |
| }, |
| { |
| "start": 1866, |
| "end": 1890, |
| "text": "(Grezl & Karafi\u00e1t, 2013)", |
| "ref_id": "BIBREF9" |
| }, |
| { |
| "start": 1891, |
| "end": 1912, |
| "text": "(Vesely et al., 2013)", |
| "ref_id": "BIBREF32" |
| }, |
| { |
| "start": 1913, |
| "end": 1933, |
| "text": "(Zhang et al., 2014)", |
| "ref_id": "BIBREF38" |
| }, |
| { |
| "start": 1934, |
| "end": 1975, |
| "text": "(Mathias, Yegnanarayanan & Fritsch, 2005)", |
| "ref_id": "BIBREF17" |
| }, |
| { |
| "start": 1976, |
| "end": 2010, |
| "text": "(Yu, Gales, Wang & Woodland, 2010)", |
| "ref_id": "BIBREF36" |
| }, |
| { |
| "start": 2011, |
| "end": 2036, |
| "text": "(Cui, Huang & Chien, 2011", |
| "ref_id": null |
| }, |
| { |
| "start": 2134, |
| "end": 2152, |
| "text": "(Liu et al., 2007)", |
| "ref_id": "BIBREF14" |
| }, |
| { |
| "start": 2153, |
| "end": 2175, |
| "text": "(Mathias et al., 2005)", |
| "ref_id": "BIBREF17" |
| }, |
| { |
| "start": 2207, |
| "end": 2245, |
| "text": "(Walker, Pedersen, Orife & Flaks, 2017", |
| "ref_id": "BIBREF33" |
| }, |
| { |
| "start": 2295, |
| "end": 2337, |
| "text": "(Manohar, Hadian, Povey & Khudanpur, 2018)", |
| "ref_id": "BIBREF15" |
| }, |
| { |
| "start": 2338, |
| "end": 2372, |
| "text": "(Manohar, Povey & Khudanpur, 2015)", |
| "ref_id": "BIBREF16" |
| }, |
| { |
| "start": 2373, |
| "end": 2387, |
| "text": "(Fiscus, 1997)", |
| "ref_id": "BIBREF5" |
| }, |
| { |
| "start": 2388, |
| "end": 2415, |
| "text": "(Evermann & Woodland, 2000)", |
| "ref_id": "BIBREF4" |
| }, |
| { |
| "start": 2416, |
| "end": 2436, |
| "text": "(Deng & Platt, 2014)", |
| "ref_id": "BIBREF1" |
| }, |
| { |
| "start": 2437, |
| "end": 2466, |
| "text": "(Xu, Povey, Mangu & Zhu, 2011", |
| "ref_id": "BIBREF35" |
| }, |
| { |
| "start": 2467, |
| "end": 2493, |
| "text": "(Zavaliagkos et al., 1998)", |
| "ref_id": "BIBREF37" |
| }, |
| { |
| "start": 2494, |
| "end": 2515, |
| "text": "(Vesely et al., 2013)", |
| "ref_id": "BIBREF32" |
| }, |
| { |
| "start": 2516, |
| "end": 2540, |
| "text": "(Grezl & Karafi\u00e1t, 2013)", |
| "ref_id": "BIBREF9" |
| }, |
| { |
| "start": 2541, |
| "end": 2561, |
| "text": "(Zhang et al., 2014)", |
| "ref_id": "BIBREF38" |
| }, |
| { |
| "start": 2646, |
| "end": 2665, |
| "text": "(Lamel et al., 2002", |
| "ref_id": "BIBREF13" |
| }, |
| { |
| "start": 2666, |
| "end": 2690, |
| "text": ") (Chan & Woodland, 2004", |
| "ref_id": null |
| }, |
| { |
| "start": 2691, |
| "end": 2709, |
| "text": "(Liu et al., 2007)", |
| "ref_id": "BIBREF14" |
| }, |
| { |
| "start": 2772, |
| "end": 2793, |
| "text": "(Vesely et al., 2013)", |
| "ref_id": "BIBREF32" |
| }, |
| { |
| "start": 2799, |
| "end": 2820, |
| "text": "(Thomas et al., 2013)", |
| "ref_id": "BIBREF29" |
| }, |
| { |
| "start": 2827, |
| "end": 2851, |
| "text": "(Grezl & Karafi\u00e1t, 2013)", |
| "ref_id": "BIBREF9" |
| }, |
| { |
| "start": 2852, |
| "end": 2873, |
| "text": "(Vesely et al., 2013)", |
| "ref_id": "BIBREF32" |
| }, |
| { |
| "start": 2874, |
| "end": 2894, |
| "text": "(Thomas et al., 2013", |
| "ref_id": "BIBREF29" |
| }, |
| { |
| "start": 2927, |
| "end": 2949, |
| "text": "(Mathias et al., 2005)", |
| "ref_id": "BIBREF17" |
| }, |
| { |
| "start": 2950, |
| "end": 2966, |
| "text": "(Yu et al., 2010", |
| "ref_id": "BIBREF36" |
| }, |
| { |
| "start": 2967, |
| "end": 2986, |
| "text": ") (Cui et al., 2011", |
| "ref_id": null |
| }, |
| { |
| "start": 3010, |
| "end": 3028, |
| "text": "(Liu et al., 2007)", |
| "ref_id": "BIBREF14" |
| }, |
| { |
| "start": 3029, |
| "end": 3051, |
| "text": "(Mathias et al., 2005)", |
| "ref_id": "BIBREF17" |
| }, |
| { |
| "start": 3052, |
| "end": 3065, |
| "text": "(Nadas, 1983)", |
| "ref_id": "BIBREF19" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "\u7dd2\u8ad6 (INTRODUCTION)", |
| "sec_num": "1." |
| }, |
| { |
| "text": "\u3002 \u6700\u8fd1 LF-MMI \u8a13\u7df4\u65b9\u6cd5\u5728 ASR \u53d6\u5f97\u4e86\u91cd\u5927\u7684\u7a81\u7834\u3002\u6709\u5225\u65bc\u50b3\u7d71\u7684\u4e8c\u968e\u6bb5\u8a13\u7df4\uff0c LF-MMI \u63d0\u4f9b\u66f4\u5feb\u7684\u8a13\u7df4\u8207\u89e3\u78bc\uff0c\u540c\u6642\u5728\u6a21\u578b\u6e96\u5ea6\u4e0a\u53d6\u5f97\u76ee\u524d\u6700\u512a\u7570\u7684\u8868\u73fe\u3002\u5118\u7ba1 LF-MMI \u5728\u76e3\u7763\u5f0f\u74b0\u5883\u4e0b\u7372\u5f97\u6700\u597d\u7684\u6210\u679c\uff0c\u4f46\u5728\u534a\u76e3\u7763\u5f0f\u74b0\u5883\u4e0b\u7684\u7814\u7a76\u6210\u679c\u4ecd\u7136\u6709\u9650\u3002 \u5728\u904e\u5f80\u7684\u7814\u7a76\u4e2d\uff0c\u9451\u5225\u5f0f\u8a13\u7df4\u7684\u597d\u58de\u5f88\u5927\u5c64\u5ea6\u5730\u4ef0\u8cf4\u65bc\u8a13\u7df4\u8a9e\u53e5\u7684\u6b63\u78ba\u6027", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "\u7dd2\u8ad6 (INTRODUCTION)", |
| "sec_num": "1." |
| }, |
| { |
| "text": "\u76f8\u540c\uff0c\u662f\u5c07\u8a5e\u5716\u7684\u4e0d\u78ba\u5b9a \u6027\u4ee5\u689d\u4ef6\u71b5(Conditional entropy)\u7684\u5f62\u5f0f\u52a0\u5165\uff0c\u4fdd\u7559\u6574\u500b\u8a5e\u5716\u4f86\u505a\u4e8c\u968e\u6bb5\u7684\u8a13\u7df4\u3002\u672c\u8ad6\u6587 \u8207\u5176\u4e0d\u540c\u7684\u662f\uff0c\u6211\u5011\u5c07\u9019\u6a23\u7684\u65b9\u6cd5\u505a\u5728\u66f4\u53e3\u8a9e\u5316\u7684\u6703\u8b70\u8a9e\u6599\uff0c\u4ee5\u53ca\u57fa\u65bc\u9019\u500b\u65b9\u6cd5\u4e4b\u4e0a\uff0c \u5229\u7528\u6574\u9ad4\u5b78\u7fd2\u7684\u89c0\u5ff5\uff0c\u9032\u4e00\u6b65\u5730\u63a2\u8a0e\u6a21\u578b\u5408\u4f75\u5e36\u4f86\u7684\u6210\u6548\u3002 \u6574 \u9ad4 \u7684 \u6a21 \u578b \u5408 \u4f75 \u5728 \u81ea \u52d5 \u8a9e \u97f3 \u8fa8 \u8b58 \u4e0a \u80fd \u53d6 \u5f97 \u512a \u65bc \u55ae \u4e00 \u6a21 \u578b \u7684 \u6210 \u679c", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "\u7dd2\u8ad6 (INTRODUCTION)", |
| "sec_num": "1." |
| }, |
| { |
| "text": "EQUATION", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [ |
| { |
| "start": 0, |
| "end": 8, |
| "text": "EQUATION", |
| "ref_id": "EQREF", |
| "raw_str": "\u7684\u76ee\u6a19\u51fd\u6578\u662f\u5728 \u7d66\u4e88\u8072\u5b78\u7279\u5fb5 O \u548c\u6a21\u578b\u53c3\u6578\u4e0b\uff0c\u4f30\u6e2c\u8f49\u5beb(Transcript)\u7684\u5c0d\u6578\u53ef\u80fd\u6027\u3002\u5206\u5b50\u70ba\u6b63\u78ba\u8f49\u5beb (Reference transcript)\u7684\u6a5f\u7387\uff0c\u800c\u5206\u6bcd\u70ba\u6240\u6709\u53ef\u80fd\u7b54\u6848\u7684\u6a5f\u7387\u3002\u56e0\u70ba\u4e00\u4e9b\u6b77\u53f2\u7684\u539f\u56e0\uff0cCML \u6210\u70ba\u4e86\u6211\u5011\u76ee\u524d\u719f\u77e5\u7684 MMI (Bahl et al., 1986)\uff0c\u5f0f\u5b50\u5982\u4e0b\uff1a \u2211 log | ,", |
| "eq_num": "(1)" |
| } |
| ], |
| "section": "\u7dd2\u8ad6 (INTRODUCTION)", |
| "sec_num": "1." |
| }, |
| { |
| "text": "EQUATION", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [ |
| { |
| "start": 0, |
| "end": 8, |
| "text": "EQUATION", |
| "ref_id": "EQREF", |
| "raw_str": "\u5f0f(1)\u7684 u \u70ba\u8a9e\u53e5\uff0c \u70ba\u8a9e\u53e5 u \u7684\u6b63\u78ba\u72c0\u614b\u5e8f\u5217(Reference state sequences)\uff0cO \u70ba\u8a9e\u53e5 u \u7684 \u8072\u5b78\u7279\u5fb5\uff0c\u03bb\u70ba\u6a21\u578b\u53c3\u6578\u3002MMI \u7684\u76ee\u6a19\u4fbf\u662f\u6700\u5927\u5316\u4e0a\u8ff0\u7684\u5f0f\u5b50\u3002\u8a73\u7d30\u7684\u8a08\u7b97\u53ef\u900f\u904e\u8c9d\u5f0f \u5b9a\u7406(Bayes' theorem)\u62c6\u89e3\u6210\u5f0f(2)\uff1a \u2211 log | , \u2211 | ,", |
| "eq_num": "(2)" |
| } |
| ], |
| "section": "\u7dd2\u8ad6 (INTRODUCTION)", |
| "sec_num": "1." |
| }, |
| { |
| "text": "EQUATION", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [ |
| { |
| "start": 0, |
| "end": 8, |
| "text": "EQUATION", |
| "ref_id": "EQREF", |
| "raw_str": "\u5728\u5f0f(2)\u4e2d\uff0c \u70ba\u8a9e\u53e5 u \u7684\u7af6\u722d\u72c0\u614b\u5e8f\u5217(Competing state sequence)\u3002\u53ef\u900f\u904e\u9451\u5225\u5f0f\u8a13\u7df4\uff0c \u5c07\u6a21\u578b\u76ee\u6a19\u51fd\u6578\u5b9a\u7fa9\u6210\u63a5\u8fd1\u6b63\u78ba\u72c0\u614b\u5e8f\u5217\u548c\u9060\u96e2\u7af6\u722d\u72c0\u614b\u5e8f\u5217\u3002\u82e5\u8981\u8a08\u7b97\u8a9e\u53e5 u \u5728\u6642\u9593 t\uff0c\u800c\u8f38\u51fa\u5c64\u70ba , \uff0c\u5247\u53ef\u504f\u5fae\u5206\u5f0f(1)\uff1a , : , ,", |
| "eq_num": "(3)" |
| } |
| ], |
| "section": "\u7dd2\u8ad6 (INTRODUCTION)", |
| "sec_num": "1." |
| }, |
| { |
| "text": "EQUATION", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [ |
| { |
| "start": 0, |
| "end": 8, |
| "text": "EQUATION", |
| "ref_id": "EQREF", |
| "raw_str": "\u5f0f(3)\u4e2d\u7684 : , \u70ba\u6307\u793a\u51fd\u6578(Indicator function)\uff0c\u7576 , \u7684\u8f38\u51fa\u5c6c\u65bc\u6b63\u78ba\u72c0\u614b\u5e8f\u5217 \u6642\u70ba 1\uff0c\u53cd\u4e4b\u5247\u70ba 0\u3002 , \u5247\u8868\u793a , \u70ba\u6b63\u78ba\u72c0\u614b\u5e8f\u5217\u7684\u4e8b\u5f8c\u6a5f\u7387(Posterior)\uff0c\u53ef\u8868\u793a \u5982\u4e0b\uff1a , \u2211 : , |", |
| "eq_num": ", \u2211 : , | , \u2211" |
| } |
| ], |
| "section": "\u7dd2\u8ad6 (INTRODUCTION)", |
| "sec_num": "1." |
| }, |
| { |
| "text": "EQUATION", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [ |
| { |
| "start": 0, |
| "end": 8, |
| "text": "EQUATION", |
| "ref_id": "EQREF", |
| "raw_str": "\u2211 | ,", |
| "eq_num": "(7)" |
| } |
| ], |
| "section": "\u7dd2\u8ad6 (INTRODUCTION)", |
| "sec_num": "1." |
| }, |
| { |
| "text": "\u5f0f(7)\u4fbf\u662f NCE (Grandvalet & Bengio, 2005) Bahl, L., Brown, P., de Souza, P., & Mercer, R. (1986) . Maximum mutual information estimation of hidden markov model parameters for speech recognition. In Proceedings of ICASSP 1986 . doi: 10.1109 /ICASSP.1986 .1169179 Chan, H. Y., & Woodland, P. (2004 . Improving broadcast news transcription by lightly supervised discriminative training. In Proceedings of ICASSP 2004 . doi: 10.1109 /ICASSP.2004 Cui, X., Huang, J., & Chien, J.-T. (2011) . Multi-view and multiobjective semi-supervised learning for large vocabulary continuous speech recognition. In Proceedings of ICASSP 2011. doi: 10.1109/ICASSP.2011.5947396", |
| "cite_spans": [ |
| { |
| "start": 11, |
| "end": 38, |
| "text": "(Grandvalet & Bengio, 2005)", |
| "ref_id": "BIBREF7" |
| }, |
| { |
| "start": 60, |
| "end": 93, |
| "text": "de Souza, P., & Mercer, R. (1986)", |
| "ref_id": null |
| }, |
| { |
| "start": 207, |
| "end": 221, |
| "text": "of ICASSP 1986", |
| "ref_id": null |
| }, |
| { |
| "start": 222, |
| "end": 236, |
| "text": ". doi: 10.1109", |
| "ref_id": null |
| }, |
| { |
| "start": 237, |
| "end": 249, |
| "text": "/ICASSP.1986", |
| "ref_id": null |
| }, |
| { |
| "start": 250, |
| "end": 292, |
| "text": ".1169179 Chan, H. Y., & Woodland, P. (2004", |
| "ref_id": null |
| }, |
| { |
| "start": 399, |
| "end": 410, |
| "text": "ICASSP 2004", |
| "ref_id": null |
| }, |
| { |
| "start": 411, |
| "end": 425, |
| "text": ". doi: 10.1109", |
| "ref_id": null |
| }, |
| { |
| "start": 426, |
| "end": 438, |
| "text": "/ICASSP.2004", |
| "ref_id": null |
| }, |
| { |
| "start": 448, |
| "end": 480, |
| "text": "Huang, J., & Chien, J.-T. (2011)", |
| "ref_id": null |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "\u7dd2\u8ad6 (INTRODUCTION)", |
| "sec_num": "1." |
| } |
| ], |
| "back_matter": [], |
| "bib_entries": { |
| "BIBREF0": { |
| "ref_id": "b0", |
| "title": "Context-dependent pre-trained deep neural networks for large-vocabulary speech recognition", |
| "authors": [ |
| { |
| "first": "G", |
| "middle": [ |
| "E" |
| ], |
| "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 Transactions on audio, speech, and language processing", |
| "volume": "20", |
| "issue": "1", |
| "pages": "30--42", |
| "other_ids": { |
| "DOI": [ |
| "10.1109/TASL.2011.2134090" |
| ] |
| }, |
| "num": null, |
| "urls": [], |
| "raw_text": "Dahl, G. E., Yu, D., Deng, L., & Acero, A. (2012). Context-dependent pre-trained deep neural networks for large-vocabulary speech recognition. IEEE Transactions on audio, speech, and language processing, 20(1), 30-42. doi: 10.1109/TASL.2011.2134090", |
| "links": null |
| }, |
| "BIBREF1": { |
| "ref_id": "b1", |
| "title": "Ensemble deep learning for speech recognition", |
| "authors": [ |
| { |
| "first": "L", |
| "middle": [], |
| "last": "Deng", |
| "suffix": "" |
| }, |
| { |
| "first": "J", |
| "middle": [ |
| "C" |
| ], |
| "last": "Platt", |
| "suffix": "" |
| } |
| ], |
| "year": 2014, |
| "venue": "Proceedings of Interspeech", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Deng, L., & Platt, J. C. (2014). Ensemble deep learning for speech recognition. In Proceedings of Interspeech 2014.", |
| "links": null |
| }, |
| "BIBREF2": { |
| "ref_id": "b2", |
| "title": "Ensemble methods in machine learning", |
| "authors": [ |
| { |
| "first": "T", |
| "middle": [ |
| "G" |
| ], |
| "last": "Dietterich", |
| "suffix": "" |
| } |
| ], |
| "year": 2000, |
| "venue": "Proceedings of International workshop on MCS 2000", |
| "volume": "", |
| "issue": "", |
| "pages": "1--15", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Dietterich, T. G. (2000). Ensemble methods in machine learning. In Proceedings of International workshop on MCS 2000, 1-15.", |
| "links": null |
| }, |
| "BIBREF3": { |
| "ref_id": "b3", |
| "title": "An experimental comparison of three methods for constructing ensembles of decision trees: bagging, boosting, and randomization", |
| "authors": [ |
| { |
| "first": "T", |
| "middle": [ |
| "G" |
| ], |
| "last": "Dietterich", |
| "suffix": "" |
| } |
| ], |
| "year": 2000, |
| "venue": "Machine Learning", |
| "volume": "40", |
| "issue": "", |
| "pages": "139--157", |
| "other_ids": { |
| "DOI": [ |
| "10.1023/A:1007607513941" |
| ] |
| }, |
| "num": null, |
| "urls": [], |
| "raw_text": "Dietterich, T. G. (2000). An experimental comparison of three methods for constructing ensembles of decision trees: bagging, boosting, and randomization. Machine Learning, 40(2), 139-157. doi: 10.1023/A:1007607513941", |
| "links": null |
| }, |
| "BIBREF4": { |
| "ref_id": "b4", |
| "title": "Posterior probability decoding, confidence estimation and system combination", |
| "authors": [ |
| { |
| "first": "G", |
| "middle": [], |
| "last": "Evermann", |
| "suffix": "" |
| }, |
| { |
| "first": "P", |
| "middle": [ |
| "C" |
| ], |
| "last": "Woodland", |
| "suffix": "" |
| } |
| ], |
| "year": 2000, |
| "venue": "Proceedings of STW", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Evermann, G., & Woodland, P. C. (2000). Posterior probability decoding, confidence estimation and system combination. In Proceedings of STW 2000.", |
| "links": null |
| }, |
| "BIBREF5": { |
| "ref_id": "b5", |
| "title": "A post-processing system to yield reduced word error rates: recognizer output voting error reduction (ROVER)", |
| "authors": [ |
| { |
| "first": "J", |
| "middle": [ |
| "G" |
| ], |
| "last": "Fiscus", |
| "suffix": "" |
| } |
| ], |
| "year": 1997, |
| "venue": "Proceedings of ASRU 1997", |
| "volume": "", |
| "issue": "", |
| "pages": "347--352", |
| "other_ids": { |
| "DOI": [ |
| "10.1109/ASRU.1997.659110" |
| ] |
| }, |
| "num": null, |
| "urls": [], |
| "raw_text": "Fiscus, J. G. (1997). A post-processing system to yield reduced word error rates: recognizer output voting error reduction (ROVER). In Proceedings of ASRU 1997, 347-352. doi: 10.1109/ASRU.1997.659110", |
| "links": null |
| }, |
| "BIBREF6": { |
| "ref_id": "b6", |
| "title": "Hypothesis spaces for minimum bayes risk training in large vocabulary speech recognition", |
| "authors": [ |
| { |
| "first": "M", |
| "middle": [], |
| "last": "Gibson", |
| "suffix": "" |
| }, |
| { |
| "first": "T", |
| "middle": [], |
| "last": "Hain", |
| "suffix": "" |
| } |
| ], |
| "year": 2006, |
| "venue": "Proceedings of Interspeech", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Gibson, M., & Hain, T. (2006). Hypothesis spaces for minimum bayes risk training in large vocabulary speech recognition. In Proceedings of Interspeech 2006.", |
| "links": null |
| }, |
| "BIBREF7": { |
| "ref_id": "b7", |
| "title": "Semi-supervised learning by entropy minimization", |
| "authors": [ |
| { |
| "first": "Y", |
| "middle": [], |
| "last": "Grandvalet", |
| "suffix": "" |
| }, |
| { |
| "first": "Y", |
| "middle": [], |
| "last": "Bengio", |
| "suffix": "" |
| } |
| ], |
| "year": 2005, |
| "venue": "Proceedings of NIPS 2005", |
| "volume": "", |
| "issue": "", |
| "pages": "529--536", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Grandvalet, Y., & Bengio , Y. (2005). Semi-supervised learning by entropy minimization. In Proceedings of NIPS 2005, 529-536.", |
| "links": null |
| }, |
| "BIBREF8": { |
| "ref_id": "b8", |
| "title": "Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks", |
| "authors": [ |
| { |
| "first": "A", |
| "middle": [], |
| "last": "Graves", |
| "suffix": "" |
| }, |
| { |
| "first": "S", |
| "middle": [], |
| "last": "Fern\u00e1ndez", |
| "suffix": "" |
| }, |
| { |
| "first": "F", |
| "middle": [], |
| "last": "Gomez", |
| "suffix": "" |
| }, |
| { |
| "first": "J", |
| "middle": [], |
| "last": "Schmidhuber", |
| "suffix": "" |
| } |
| ], |
| "year": 2006, |
| "venue": "Proceedings of ICML 2006", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": { |
| "DOI": [ |
| "10.1145/1143844.1143891" |
| ] |
| }, |
| "num": null, |
| "urls": [], |
| "raw_text": "Graves, A., Fern\u00e1ndez, S., Gomez, F., & Schmidhuber, J. (2006). Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In Proceedings of ICML 2006. doi: 10.1145/1143844.1143891", |
| "links": null |
| }, |
| "BIBREF9": { |
| "ref_id": "b9", |
| "title": "Semi-supervised bootstrapping approach for neural network feature extractor training", |
| "authors": [ |
| { |
| "first": "F", |
| "middle": [], |
| "last": "Grezl", |
| "suffix": "" |
| }, |
| { |
| "first": "M", |
| "middle": [], |
| "last": "Karafi\u00e1t", |
| "suffix": "" |
| } |
| ], |
| "year": 2013, |
| "venue": "Proceedingd of ASRU 2013", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": { |
| "DOI": [ |
| "10.1109/ASRU.2013.6707775" |
| ] |
| }, |
| "num": null, |
| "urls": [], |
| "raw_text": "Grezl, F., & Karafi\u00e1t, M. (2013). Semi-supervised bootstrapping approach for neural network feature extractor training. In Proceedingd of ASRU 2013. doi: 10.1109/ASRU.2013.6707775", |
| "links": null |
| }, |
| "BIBREF10": { |
| "ref_id": "b10", |
| "title": "Semi-supervised training of gaussian mixture models by conditional entropy minimization", |
| "authors": [ |
| { |
| "first": "J.-T", |
| "middle": [], |
| "last": "Huang", |
| "suffix": "" |
| }, |
| { |
| "first": "M", |
| "middle": [], |
| "last": "Hasegawa-Johnson", |
| "suffix": "" |
| } |
| ], |
| "year": 2010, |
| "venue": "Proceedings of INTERSPEECH 2010", |
| "volume": "", |
| "issue": "", |
| "pages": "1353--1356", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Huang, J.-T., & Hasegawa-Johnson, M. (2010). Semi-supervised training of gaussian mixture models by conditional entropy minimization. In Proceedings of INTERSPEECH 2010, 1353-1356.", |
| "links": null |
| }, |
| "BIBREF11": { |
| "ref_id": "b11", |
| "title": "Minimum classification error rate methods for speech recognition", |
| "authors": [ |
| { |
| "first": "B.-H", |
| "middle": [], |
| "last": "Juang", |
| "suffix": "" |
| }, |
| { |
| "first": "W", |
| "middle": [], |
| "last": "Hou", |
| "suffix": "" |
| }, |
| { |
| "first": "C.-H", |
| "middle": [], |
| "last": "Lee", |
| "suffix": "" |
| } |
| ], |
| "year": 1997, |
| "venue": "IEEE Transactions on Speech and Audio Processing", |
| "volume": "5", |
| "issue": "3", |
| "pages": "257--265", |
| "other_ids": { |
| "DOI": [ |
| "10.1109/89.568732" |
| ] |
| }, |
| "num": null, |
| "urls": [], |
| "raw_text": "Juang, B.-H., Hou, W., & Lee, C.-H. (1997). Minimum classification error rate methods for speech recognition. IEEE Transactions on Speech and Audio Processing, 5(3), 257-265. doi: 10.1109/89.568732", |
| "links": null |
| }, |
| "BIBREF12": { |
| "ref_id": "b12", |
| "title": "A novel loss function for the overall risk criterion based discriminative training of HMM models", |
| "authors": [ |
| { |
| "first": "J", |
| "middle": [], |
| "last": "Kaiser", |
| "suffix": "" |
| }, |
| { |
| "first": "B", |
| "middle": [], |
| "last": "Horvat", |
| "suffix": "" |
| }, |
| { |
| "first": "Z", |
| "middle": [], |
| "last": "Kacic", |
| "suffix": "" |
| } |
| ], |
| "year": 2000, |
| "venue": "Proceedings of ICSLP 2000", |
| "volume": "", |
| "issue": "", |
| "pages": "887--890", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Kaiser, J., Horvat, B., & Kacic, Z. (2000). A novel loss function for the overall risk criterion based discriminative training of HMM models. In Proceedings of ICSLP 2000, 887-890.", |
| "links": null |
| }, |
| "BIBREF13": { |
| "ref_id": "b13", |
| "title": "Lightly supervised and unsupervised acoustic model training", |
| "authors": [ |
| { |
| "first": "L", |
| "middle": [], |
| "last": "Lamel", |
| "suffix": "" |
| }, |
| { |
| "first": "J.-L", |
| "middle": [], |
| "last": "Gauvain", |
| "suffix": "" |
| }, |
| { |
| "first": "G", |
| "middle": [], |
| "last": "Adda", |
| "suffix": "" |
| } |
| ], |
| "year": 2002, |
| "venue": "Computer Speech & Language", |
| "volume": "16", |
| "issue": "1", |
| "pages": "115--129", |
| "other_ids": { |
| "DOI": [ |
| "10.1006/csla.2001.0186" |
| ] |
| }, |
| "num": null, |
| "urls": [], |
| "raw_text": "Lamel, L., Gauvain, J.-L., & Adda, G. (2002). Lightly supervised and unsupervised acoustic model training. Computer Speech & Language, 16(1), 115-129. doi: 10.1006/csla.2001.0186", |
| "links": null |
| }, |
| "BIBREF14": { |
| "ref_id": "b14", |
| "title": "Investigating data selection for minimum phone error training of acoustic models", |
| "authors": [ |
| { |
| "first": "S.-H", |
| "middle": [], |
| "last": "Liu", |
| "suffix": "" |
| }, |
| { |
| "first": "F.-H", |
| "middle": [], |
| "last": "Chu", |
| "suffix": "" |
| }, |
| { |
| "first": "S.-H", |
| "middle": [], |
| "last": "Lin", |
| "suffix": "" |
| }, |
| { |
| "first": "B", |
| "middle": [], |
| "last": "Chen", |
| "suffix": "" |
| } |
| ], |
| "year": 2007, |
| "venue": "Proceedings of ICME 2007", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": { |
| "DOI": [ |
| "10.1109/ICME.2007.4284658" |
| ] |
| }, |
| "num": null, |
| "urls": [], |
| "raw_text": "Liu, S.-H., Chu, F.-H., Lin, S.-H., & Chen, B. (2007). Investigating data selection for minimum phone error training of acoustic models. In Proceedings of ICME 2007. doi: 10.1109/ICME.2007.4284658", |
| "links": null |
| }, |
| "BIBREF15": { |
| "ref_id": "b15", |
| "title": "Semi-supervised training of acoustic models using lattice-free MMI", |
| "authors": [ |
| { |
| "first": "V", |
| "middle": [], |
| "last": "Manohar", |
| "suffix": "" |
| }, |
| { |
| "first": "H", |
| "middle": [], |
| "last": "Hadian", |
| "suffix": "" |
| }, |
| { |
| "first": "D", |
| "middle": [], |
| "last": "Povey", |
| "suffix": "" |
| }, |
| { |
| "first": "S", |
| "middle": [], |
| "last": "Khudanpur", |
| "suffix": "" |
| } |
| ], |
| "year": 2018, |
| "venue": "Proceedings of ICASSP 2018", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": { |
| "DOI": [ |
| "10.1109/ICASSP.2018.8462331" |
| ] |
| }, |
| "num": null, |
| "urls": [], |
| "raw_text": "Manohar, V., Hadian, H., Povey, D., & Khudanpur, S. (2018). Semi-supervised training of acoustic models using lattice-free MMI. In Proceedings of ICASSP 2018. doi: 10.1109/ICASSP.2018.8462331", |
| "links": null |
| }, |
| "BIBREF16": { |
| "ref_id": "b16", |
| "title": "Semi-supervised maximum mutual information training of deep neural network acoustic models", |
| "authors": [ |
| { |
| "first": "V", |
| "middle": [], |
| "last": "Manohar", |
| "suffix": "" |
| }, |
| { |
| "first": "D", |
| "middle": [], |
| "last": "Povey", |
| "suffix": "" |
| }, |
| { |
| "first": "S", |
| "middle": [], |
| "last": "Khudanpur", |
| "suffix": "" |
| } |
| ], |
| "year": 2015, |
| "venue": "Proceedings of Interspeech", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Manohar, V., Povey, D., & Khudanpur, S. (2015). Semi-supervised maximum mutual information training of deep neural network acoustic models. In Proceedings of Interspeech 2015.", |
| "links": null |
| }, |
| "BIBREF17": { |
| "ref_id": "b17", |
| "title": "Discriminative training of acoustic models applied to domains with unreliable transcripts", |
| "authors": [ |
| { |
| "first": "L", |
| "middle": [], |
| "last": "Mathias", |
| "suffix": "" |
| }, |
| { |
| "first": "G", |
| "middle": [], |
| "last": "Yegnanarayanan", |
| "suffix": "" |
| }, |
| { |
| "first": "J", |
| "middle": [], |
| "last": "Fritsch", |
| "suffix": "" |
| } |
| ], |
| "year": 2005, |
| "venue": "Proceedings of ICASSP 2005", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": { |
| "DOI": [ |
| "10.1109/ICASSP.2005.1415062" |
| ] |
| }, |
| "num": null, |
| "urls": [], |
| "raw_text": "Mathias, L., Yegnanarayanan, G., & Fritsch, J. (2005). Discriminative training of acoustic models applied to domains with unreliable transcripts. In Proceedings of ICASSP 2005. doi: 10.1109/ICASSP.2005.1415062", |
| "links": null |
| }, |
| "BIBREF18": { |
| "ref_id": "b18", |
| "title": "The ami meeting corpus", |
| "authors": [ |
| { |
| "first": "I", |
| "middle": [], |
| "last": "Mccowan", |
| "suffix": "" |
| }, |
| { |
| "first": "G", |
| "middle": [], |
| "last": "Lathoud", |
| "suffix": "" |
| }, |
| { |
| "first": "M", |
| "middle": [], |
| "last": "Lincoln", |
| "suffix": "" |
| }, |
| { |
| "first": "A", |
| "middle": [], |
| "last": "Lisowska", |
| "suffix": "" |
| }, |
| { |
| "first": "W", |
| "middle": [], |
| "last": "Post", |
| "suffix": "" |
| }, |
| { |
| "first": "D", |
| "middle": [], |
| "last": "Reidsma", |
| "suffix": "" |
| }, |
| { |
| "first": "P", |
| "middle": [], |
| "last": "\u2026wellner", |
| "suffix": "" |
| } |
| ], |
| "year": 2005, |
| "venue": "Proceedings of ICMTBR 2005", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "McCowan, I., Lathoud, G., Lincoln, M., Lisowska, A., Post, W., Reidsma, D., \u2026Wellner, P. (2005). The ami meeting corpus. In Proceedings of ICMTBR 2005.", |
| "links": null |
| }, |
| "BIBREF19": { |
| "ref_id": "b19", |
| "title": "A decision theoretic formulation of a training problem in speech recognition and a comparison of training by unconditional versus conditional maximum likelihood", |
| "authors": [ |
| { |
| "first": "A", |
| "middle": [], |
| "last": "Nadas", |
| "suffix": "" |
| } |
| ], |
| "year": 1983, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Nadas, A. (1983). A decision theoretic formulation of a training problem in speech recognition and a comparison of training by unconditional versus conditional maximum likelihood.", |
| "links": null |
| }, |
| "BIBREF21": { |
| "ref_id": "b21", |
| "title": "A time delay neural network architecture for efficient modeling of long temporal contexts", |
| "authors": [ |
| { |
| "first": "V", |
| "middle": [], |
| "last": "Peddinti", |
| "suffix": "" |
| }, |
| { |
| "first": "D", |
| "middle": [], |
| "last": "Povey", |
| "suffix": "" |
| }, |
| { |
| "first": "S", |
| "middle": [], |
| "last": "Khudanpur", |
| "suffix": "" |
| } |
| ], |
| "year": 2015, |
| "venue": "Proceedings of Interspeech", |
| "volume": "", |
| "issue": "", |
| "pages": "3214--3218", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Peddinti, V., Povey, D., & Khudanpur, S. (2015). A time delay neural network architecture for efficient modeling of long temporal contexts. In Proceedings of Interspeech 2015, 3214-3218.", |
| "links": null |
| }, |
| "BIBREF22": { |
| "ref_id": "b22", |
| "title": "Minimum phone error and i-smoothing for improved discriminative training", |
| "authors": [ |
| { |
| "first": "D", |
| "middle": [], |
| "last": "Povey", |
| "suffix": "" |
| }, |
| { |
| "first": "P", |
| "middle": [ |
| "C" |
| ], |
| "last": "Woodland", |
| "suffix": "" |
| } |
| ], |
| "year": 2002, |
| "venue": "Proceedings of ICASSP 2002", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": { |
| "DOI": [ |
| "10.1109/ICASSP.2002.5743665" |
| ] |
| }, |
| "num": null, |
| "urls": [], |
| "raw_text": "Povey, D., & Woodland, P. C. (2002). Minimum phone error and i-smoothing for improved discriminative training. In Proceedings of ICASSP 2002. doi: 10.1109/ICASSP.2002.5743665", |
| "links": null |
| }, |
| "BIBREF23": { |
| "ref_id": "b23", |
| "title": "Boosted MMI for feature and model space discriminative training", |
| "authors": [ |
| { |
| "first": "D", |
| "middle": [], |
| "last": "Povey", |
| "suffix": "" |
| }, |
| { |
| "first": "D", |
| "middle": [], |
| "last": "Kanevsky", |
| "suffix": "" |
| }, |
| { |
| "first": "B", |
| "middle": [], |
| "last": "Kingsbury", |
| "suffix": "" |
| }, |
| { |
| "first": "B", |
| "middle": [], |
| "last": "Ramabhadran", |
| "suffix": "" |
| }, |
| { |
| "first": "G", |
| "middle": [], |
| "last": "Saon", |
| "suffix": "" |
| }, |
| { |
| "first": "K", |
| "middle": [], |
| "last": "Visweswariah", |
| "suffix": "" |
| } |
| ], |
| "year": 2008, |
| "venue": "Proceedings of ICASSP 2008", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": { |
| "DOI": [ |
| "10.1109/ICASSP.2008.4518545" |
| ] |
| }, |
| "num": null, |
| "urls": [], |
| "raw_text": "Povey, D., Kanevsky, D., Kingsbury, B., Ramabhadran, B., Saon, G., & Visweswariah, K. (2008). Boosted MMI for feature and model space discriminative training. In Proceedings of ICASSP 2008. doi: 10.1109/ICASSP.2008.4518545", |
| "links": null |
| }, |
| "BIBREF24": { |
| "ref_id": "b24", |
| "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": "N", |
| "middle": [], |
| "last": "Goel", |
| "suffix": "" |
| }, |
| { |
| "first": "M", |
| "middle": [], |
| "last": "Hannemann", |
| "suffix": "" |
| }, |
| { |
| "first": "Y", |
| "middle": [], |
| "last": "Qian", |
| "suffix": "" |
| }, |
| { |
| "first": "G", |
| "middle": [], |
| "last": "\u2026stemmer", |
| "suffix": "" |
| } |
| ], |
| "year": 2011, |
| "venue": "Proceedings of ASRU", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Povey, D., Ghoshal, A., Boulianne, G., Goel, N., Hannemann, M., Qian, Y., \u2026Stemmer, G. (2011). The Kaldi speech recognition toolkit. In Proceedings of ASRU 2011.", |
| "links": null |
| }, |
| "BIBREF25": { |
| "ref_id": "b25", |
| "title": "Purely sequence-trained neural networks for ASR Based on Lattice-Free MMI", |
| "authors": [ |
| { |
| "first": "D", |
| "middle": [], |
| "last": "Povey", |
| "suffix": "" |
| }, |
| { |
| "first": "V", |
| "middle": [], |
| "last": "Peddinti", |
| "suffix": "" |
| }, |
| { |
| "first": "D", |
| "middle": [], |
| "last": "Galvez", |
| "suffix": "" |
| }, |
| { |
| "first": "P", |
| "middle": [], |
| "last": "Ghahremani", |
| "suffix": "" |
| }, |
| { |
| "first": "V", |
| "middle": [], |
| "last": "Manohar", |
| "suffix": "" |
| }, |
| { |
| "first": "X", |
| "middle": [], |
| "last": "Na", |
| "suffix": "" |
| }, |
| { |
| "first": "S", |
| "middle": [], |
| "last": "\u2026khudanpur", |
| "suffix": "" |
| } |
| ], |
| "year": 2016, |
| "venue": "Proceedings of Interspeech", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": { |
| "DOI": [ |
| "10.21437/Interspeech.2016-595" |
| ] |
| }, |
| "num": null, |
| "urls": [], |
| "raw_text": "Povey, D., Peddinti, V., Galvez, D., Ghahremani, P., Manohar, V., Na, X., \u2026Khudanpur, S. (2016). Purely sequence-trained neural networks for ASR Based on Lattice-Free MMI. In Proceedings of Interspeech 2016. doi: 10.21437/Interspeech.2016-595", |
| "links": null |
| }, |
| "BIBREF26": { |
| "ref_id": "b26", |
| "title": "Lower frame rate neural network acoustic models", |
| "authors": [ |
| { |
| "first": "G", |
| "middle": [], |
| "last": "Pundak", |
| "suffix": "" |
| }, |
| { |
| "first": "T", |
| "middle": [ |
| "N" |
| ], |
| "last": "Sainath", |
| "suffix": "" |
| } |
| ], |
| "year": 2016, |
| "venue": "Proceedings of Interspeech", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": { |
| "DOI": [ |
| "10.21437/Interspeech.2016-275" |
| ] |
| }, |
| "num": null, |
| "urls": [], |
| "raw_text": "Pundak, G., & Sainath, T. N. (2016). Lower frame rate neural network acoustic models. In Proceedings of Interspeech 2016. doi: 10.21437/Interspeech.2016-275", |
| "links": null |
| }, |
| "BIBREF27": { |
| "ref_id": "b27", |
| "title": "Conversational speech transcription using context-dependent deep neural networks", |
| "authors": [ |
| { |
| "first": "F", |
| "middle": [], |
| "last": "Seide", |
| "suffix": "" |
| }, |
| { |
| "first": "G", |
| "middle": [], |
| "last": "Li", |
| "suffix": "" |
| }, |
| { |
| "first": "D", |
| "middle": [], |
| "last": "Yu", |
| "suffix": "" |
| } |
| ], |
| "year": 2011, |
| "venue": "Proceedings of Interspeech", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Seide, F., Li, G., & Yu, D. (2011). Conversational speech transcription using context-dependent deep neural networks. In Proceedings of Interspeech 2011.", |
| "links": null |
| }, |
| "BIBREF28": { |
| "ref_id": "b28", |
| "title": "Acoustic modelling with CD CTC-SMBR LSTM RNNs", |
| "authors": [ |
| { |
| "first": "A", |
| "middle": [], |
| "last": "Senior", |
| "suffix": "" |
| }, |
| { |
| "first": "H", |
| "middle": [], |
| "last": "Sak", |
| "suffix": "" |
| }, |
| { |
| "first": "F", |
| "middle": [ |
| "C" |
| ], |
| "last": "Quitry", |
| "suffix": "" |
| }, |
| { |
| "first": "T", |
| "middle": [], |
| "last": "Sainath", |
| "suffix": "" |
| }, |
| { |
| "first": "K", |
| "middle": [], |
| "last": "Rao", |
| "suffix": "" |
| } |
| ], |
| "year": 2015, |
| "venue": "Proceedings of ASRU 2015", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": { |
| "DOI": [ |
| "10.1109/ASRU.2015.7404851" |
| ] |
| }, |
| "num": null, |
| "urls": [], |
| "raw_text": "Senior, A., Sak, H., Quitry, F. C., Sainath, T., & Rao, K. (2015). Acoustic modelling with CD CTC-SMBR LSTM RNNs. In Proceedings of ASRU 2015. doi: 10.1109/ASRU.2015.7404851", |
| "links": null |
| }, |
| "BIBREF29": { |
| "ref_id": "b29", |
| "title": "Deep neural network features and semisupervised training for low resource speech recognition", |
| "authors": [ |
| { |
| "first": "S", |
| "middle": [], |
| "last": "Thomas", |
| "suffix": "" |
| }, |
| { |
| "first": "M", |
| "middle": [ |
| "L" |
| ], |
| "last": "Seltzer", |
| "suffix": "" |
| }, |
| { |
| "first": "K", |
| "middle": [], |
| "last": "Church", |
| "suffix": "" |
| }, |
| { |
| "first": "H", |
| "middle": [], |
| "last": "Hermansky", |
| "suffix": "" |
| } |
| ], |
| "year": 2013, |
| "venue": "Proceedings of ICASSP 2013", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": { |
| "DOI": [ |
| "10.1109/ICASSP.2013.6638959\u7f85\u5929\u5b8f\u8207\u9673\u67cf\u7433" |
| ] |
| }, |
| "num": null, |
| "urls": [], |
| "raw_text": "Thomas, S., Seltzer, M. L., Church, K., & Hermansky, H. (2013). Deep neural network features and semisupervised training for low resource speech recognition. In Proceedings of ICASSP 2013. doi: 10.1109/ICASSP.2013.6638959 \u7f85\u5929\u5b8f\u8207\u9673\u67cf\u7433", |
| "links": null |
| }, |
| "BIBREF30": { |
| "ref_id": "b30", |
| "title": "Lattice-based discriminative training for large vocabulary speech recognition", |
| "authors": [ |
| { |
| "first": "V", |
| "middle": [], |
| "last": "Valtchev", |
| "suffix": "" |
| }, |
| { |
| "first": "J", |
| "middle": [ |
| "J" |
| ], |
| "last": "Odell", |
| "suffix": "" |
| }, |
| { |
| "first": "P", |
| "middle": [ |
| "C" |
| ], |
| "last": "Woodland", |
| "suffix": "" |
| }, |
| { |
| "first": "S", |
| "middle": [ |
| "J" |
| ], |
| "last": "Young", |
| "suffix": "" |
| } |
| ], |
| "year": 1996, |
| "venue": "Proceedings of ICASSP 1996", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": { |
| "DOI": [ |
| "10.1109/ICASSP.1996.543193" |
| ] |
| }, |
| "num": null, |
| "urls": [], |
| "raw_text": "Valtchev, V., Odell, J. J., Woodland, P. C., & Young, S. J. (1996). Lattice-based discriminative training for large vocabulary speech recognition. In Proceedings of ICASSP 1996. doi: 10.1109/ICASSP.1996.543193", |
| "links": null |
| }, |
| "BIBREF31": { |
| "ref_id": "b31", |
| "title": "MMIE training of large vocabulary recognition systems", |
| "authors": [ |
| { |
| "first": "V", |
| "middle": [], |
| "last": "Valtchev", |
| "suffix": "" |
| }, |
| { |
| "first": "J", |
| "middle": [ |
| "J" |
| ], |
| "last": "Odell", |
| "suffix": "" |
| }, |
| { |
| "first": "P", |
| "middle": [ |
| "C" |
| ], |
| "last": "Woodland", |
| "suffix": "" |
| }, |
| { |
| "first": "S", |
| "middle": [ |
| "J" |
| ], |
| "last": "Young", |
| "suffix": "" |
| } |
| ], |
| "year": 1997, |
| "venue": "Speech Communication", |
| "volume": "22", |
| "issue": "4", |
| "pages": "303--314", |
| "other_ids": { |
| "DOI": [ |
| "10.1016/S0167-6393(97)00029-0" |
| ] |
| }, |
| "num": null, |
| "urls": [], |
| "raw_text": "Valtchev, V., Odell, J. J., Woodland, P. C., & Young, S. J. (1997). MMIE training of large vocabulary recognition systems. Speech Communication, 22(4), 303-314. doi: 10.1016/S0167-6393(97)00029-0", |
| "links": null |
| }, |
| "BIBREF32": { |
| "ref_id": "b32", |
| "title": "Semi-supervised training of deep neural networks", |
| "authors": [ |
| { |
| "first": "K", |
| "middle": [], |
| "last": "Vesely", |
| "suffix": "" |
| }, |
| { |
| "first": "M", |
| "middle": [], |
| "last": "Hannemann", |
| "suffix": "" |
| }, |
| { |
| "first": "L", |
| "middle": [], |
| "last": "Burget", |
| "suffix": "" |
| } |
| ], |
| "year": 2013, |
| "venue": "Proceedings of ASRU 2013", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": { |
| "DOI": [ |
| "10.1109/ASRU.2013.6707741" |
| ] |
| }, |
| "num": null, |
| "urls": [], |
| "raw_text": "Vesely, K., Hannemann, M., & Burget, L. (2013). Semi-supervised training of deep neural networks. In Proceedings of ASRU 2013. doi: 10.1109/ASRU.2013.6707741", |
| "links": null |
| }, |
| "BIBREF33": { |
| "ref_id": "b33", |
| "title": "Semi-supervised model training for unbounded conversational speech recognition", |
| "authors": [ |
| { |
| "first": "S", |
| "middle": [], |
| "last": "Walker", |
| "suffix": "" |
| }, |
| { |
| "first": "M", |
| "middle": [], |
| "last": "Pedersen", |
| "suffix": "" |
| }, |
| { |
| "first": "I", |
| "middle": [], |
| "last": "Orife", |
| "suffix": "" |
| }, |
| { |
| "first": "J", |
| "middle": [], |
| "last": "Flaks", |
| "suffix": "" |
| } |
| ], |
| "year": 2017, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Walker, S., Pedersen, M., Orife, I., & Flaks, J. (2017). Semi-supervised model training for unbounded conversational speech recognition. Retrieved from https://arxiv.org/abs/1705.09724", |
| "links": null |
| }, |
| "BIBREF34": { |
| "ref_id": "b34", |
| "title": "Large scale discriminative training of hidden markov models for speech recognition", |
| "authors": [ |
| { |
| "first": "P", |
| "middle": [ |
| "C" |
| ], |
| "last": "Woodland", |
| "suffix": "" |
| }, |
| { |
| "first": "D", |
| "middle": [], |
| "last": "Povey", |
| "suffix": "" |
| } |
| ], |
| "year": 2002, |
| "venue": "Computer Speech & Language", |
| "volume": "16", |
| "issue": "1", |
| "pages": "25--47", |
| "other_ids": { |
| "DOI": [ |
| "10.1006/csla.2001.0182" |
| ] |
| }, |
| "num": null, |
| "urls": [], |
| "raw_text": "Woodland, P. C., & Povey, D. (2002). Large scale discriminative training of hidden markov models for speech recognition. Computer Speech & Language, 16(1), 25-47. doi: 10.1006/csla.2001.0182", |
| "links": null |
| }, |
| "BIBREF35": { |
| "ref_id": "b35", |
| "title": "Minimum bayes risk decoding and system combination based on a recursion for edit distance", |
| "authors": [ |
| { |
| "first": "H", |
| "middle": [], |
| "last": "Xu", |
| "suffix": "" |
| }, |
| { |
| "first": "D", |
| "middle": [], |
| "last": "Povey", |
| "suffix": "" |
| }, |
| { |
| "first": "L", |
| "middle": [], |
| "last": "Mangu", |
| "suffix": "" |
| }, |
| { |
| "first": "J", |
| "middle": [], |
| "last": "Zhu", |
| "suffix": "" |
| } |
| ], |
| "year": 2011, |
| "venue": "Computer Speech and Language", |
| "volume": "25", |
| "issue": "4", |
| "pages": "802--828", |
| "other_ids": { |
| "DOI": [ |
| "10.1016/j.csl.2011.03.001" |
| ] |
| }, |
| "num": null, |
| "urls": [], |
| "raw_text": "Xu, H., Povey, D., Mangu, L., & Zhu, J. (2011). Minimum bayes risk decoding and system combination based on a recursion for edit distance. Computer Speech and Language, 25(4), 802-828. doi: 10.1016/j.csl.2011.03.001", |
| "links": null |
| }, |
| "BIBREF36": { |
| "ref_id": "b36", |
| "title": "Unsupervised training and directed manual transcription for LVCSR", |
| "authors": [ |
| { |
| "first": "K", |
| "middle": [], |
| "last": "Yu", |
| "suffix": "" |
| }, |
| { |
| "first": "M", |
| "middle": [], |
| "last": "Gales", |
| "suffix": "" |
| }, |
| { |
| "first": "L", |
| "middle": [], |
| "last": "Wang", |
| "suffix": "" |
| }, |
| { |
| "first": "P", |
| "middle": [ |
| "C" |
| ], |
| "last": "Woodland", |
| "suffix": "" |
| } |
| ], |
| "year": 2010, |
| "venue": "Speech Communication", |
| "volume": "52", |
| "issue": "7-8", |
| "pages": "652--663", |
| "other_ids": { |
| "DOI": [ |
| "10.1016/j.specom.2010.02.014" |
| ] |
| }, |
| "num": null, |
| "urls": [], |
| "raw_text": "Yu, K., Gales, M., Wang, L., & Woodland, P. C. (2010). Unsupervised training and directed manual transcription for LVCSR. Speech Communication, 52(7-8), 652-663. doi: 10.1016/j.specom.2010.02.014", |
| "links": null |
| }, |
| "BIBREF37": { |
| "ref_id": "b37", |
| "title": "Using untranscribed training data to improve performance", |
| "authors": [ |
| { |
| "first": "G", |
| "middle": [], |
| "last": "Zavaliagkos", |
| "suffix": "" |
| }, |
| { |
| "first": "M", |
| "middle": [], |
| "last": "Siu", |
| "suffix": "" |
| }, |
| { |
| "first": "T", |
| "middle": [], |
| "last": "Colthurst", |
| "suffix": "" |
| }, |
| { |
| "first": "J", |
| "middle": [], |
| "last": "Billa", |
| "suffix": "" |
| } |
| ], |
| "year": 1998, |
| "venue": "Proceedings of ICSLP", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Zavaliagkos, G., Siu, M., Colthurst, T., & Billa, J. (1998). Using untranscribed training data to improve performance. In Proceedings of ICSLP 1998.", |
| "links": null |
| }, |
| "BIBREF38": { |
| "ref_id": "b38", |
| "title": "Semi-supervised DNN training in meeting recognition", |
| "authors": [ |
| { |
| "first": "P", |
| "middle": [], |
| "last": "Zhang", |
| "suffix": "" |
| }, |
| { |
| "first": "Y", |
| "middle": [], |
| "last": "Liu", |
| "suffix": "" |
| }, |
| { |
| "first": "T", |
| "middle": [], |
| "last": "Hain", |
| "suffix": "" |
| } |
| ], |
| "year": 2014, |
| "venue": "Proceedings of SLT 2014", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": { |
| "DOI": [ |
| "10.1109/SLT.2014.7078564" |
| ] |
| }, |
| "num": null, |
| "urls": [], |
| "raw_text": "Zhang, P., Liu, Y., & Hain, T. (2014). Semi-supervised DNN training in meeting recognition. In Proceedings of SLT 2014. doi: 10.1109/SLT.2014.7078564", |
| "links": null |
| } |
| }, |
| "ref_entries": { |
| "TABREF0": { |
| "type_str": "table", |
| "num": null, |
| "content": "<table><tr><td>)\uff0c\u9019\u6a23</td></tr><tr><td>\u6548\u80fd\u7684\u9032\u6b65\u6b78\u529f\u65bc\u4e0b\u5217\u5e7e\u9ede\uff0c\u5404\u5225\u6a21\u578b\u53ef\u4ee5\u4fee\u6b63\u5f7c\u6b64\u7684\u932f\u8aa4\uff1b\u6e1b\u5c11\u9078\u64c7\u5230\u8f03\u5dee\u6a21\u578b\u7684\u53ef</td></tr><tr><td>\u80fd\u6027\uff1b\u589e\u52a0\u6574\u9ad4\u6a21\u578b\u641c\u5c0b\u6642\u7684\u5047\u8aaa\u7a7a\u9593(Dietterich, 2000)\uff0c\u7528\u4ee5\u4fee\u6b63\u8a13\u7df4\u6642\u7684\u554f\u984c\u3002\u5982\u8a9e</td></tr><tr><td>\u6599\u9078\u64c7(Data selection)\u3001\u76ee\u6a19\u51fd\u6578(Objective function)\u3001\u6a21\u578b(Model)\u3002\u9019\u88e1\u6211\u5011\u671f\u5f85\u5229\u7528</td></tr><tr><td>\u6574\u9ad4\u5b78\u7fd2\u589e\u52a0\u7684\u5047\u8aaa\u7a7a\u9593\uff0c\u89e3\u6c7a\u5728\u534a\u76e3\u7763\u5f0f\u8a13\u7df4\u6642\u6709\u9650\u8a9e\u6599\u9020\u6210\u6548\u80fd\u964d\u4f4e\u7684\u554f\u984c\u3002\u8a13\u7df4</td></tr><tr><td>\u7684\u904e\u7a0b\u70ba\u5404\u5225\u8a13\u7df4\u6bcf\u500b\u6a21\u578b\uff0c\u63a5\u8457\u5728\u8a13\u7df4\u7d50\u675f\u5f8c\u7684\u968e\u6bb5\u52a0\u5165\u5408\u4f75\u6a21\u578b\u7684\u6280\u8853\uff0c\u8b93\u6a21\u578b\u4fee</td></tr><tr><td>\u6b63\u5f7c\u6b64\u7684\u932f\u8aa4\uff0c\u9032\u4e00\u6b65\u63d0\u5347\u6548\u80fd\u3002\u9019\u88e1\u6211\u5011\u63a1\u7528\u5169\u7a2e\u4e0d\u540c\u5c64\u7d1a\u7684\u5408\u4f75\u65b9\u6cd5\uff0c\u97f3\u6846\u5c64\u7d1a\u7684</td></tr><tr><td>\u5408\u4f75(Frame-level combination or score fusion) (Deng & Platt, 2014)\uff0c\u4ee5\u53ca\u5047\u8aaa\u5c64\u7d1a\u7684\u5408\u4f75</td></tr><tr><td>(Hypothesis-level combination) (Fiscus, 1997) (Xu et al., 2011)\u3002\u5728(Senior, Sak, Quitry,</td></tr><tr><td>Sainath & Rao, 2015)\u7684\u7814\u7a76\u4e2d\u97f3\u6846\u5c64\u7d1a\u5408\u4f75\u7121\u52a9\u65bc CTC \u7684\u8868\u73fe\uff0c\u800c LF-MMI \u88ab\u8996\u70ba CTC</td></tr><tr><td>\u7684\u5ef6\u4f38\uff0c\u56e0\u6b64\u63a2\u8a0e\u534a\u76e3\u7763\u5f0f LF-MMI \u7684\u5408\u4f75\u7d50\u679c\u662f\u5177\u6709\u50f9\u503c\u7684\u4e8b\u60c5\u3002</td></tr><tr><td>\u672c\u8ad6\u6587\u7684\u5be6\u4f5c\u76ee\u7684\u4fbf\u662f\u5728\u8a9e\u6599\u7f3a\u4e4f\u7684\u534a\u76e3\u7763\u5f0f\u74b0\u5883\u4e0b\uff0c\u4f7f\u7528\u8ca0\u689d\u4ef6\u71b5\u8207\u8a5e\u5716\u8f14\u52a9</td></tr><tr><td>LF-MMI \u7684\u8a13\u7df4\uff0c\u4e26\u5229\u7528\u6a21\u578b\u5408\u4f75\u6280\u8853\uff0c\u9032\u4e00\u6b65\u63d0\u5347\u6a21\u578b\u7684\u8fa8\u8b58\u7d50\u679c\u3002\u6211\u5011\u5e0c\u671b\u5373\u4f7f\u5728</td></tr><tr><td>\u534a\u76e3\u7763\u5f0f\u8072\u5b78\u6a21\u578b\u76ee\u7684\u662f\u89e3\u6c7a\u4e0b\u5217\u554f\u984c\uff1a\u4f4e\u8cc7\u6e90\u7684\u8a9e\u6599\u5eab\u3001\u5927\u91cf\u7684\u672a\u8f49\u5beb\u8a9e\u6599\u3001\u6e2c\u8a66\u8a9e</td></tr><tr><td>\u6599\u8207\u8a13\u7df4\u8a9e\u6599\u7684\u4e0d\u5339\u914d\u3002\u9996\u5148\uff0c\u5145\u8db3\u7684\u8a9e\u6599\u5eab\u662f\u8b93\u76ee\u524d\u6700\u65b0\u7a4e\u7684 ASR \u7cfb\u7d71\u53ef\u4ee5\u8868\u73fe\u512a\u7570</td></tr><tr><td>\u7684\u539f\u56e0\u4e4b\u4e00\uff0c\u4f46\u6211\u5011\u64c1\u6709\u7684\u8f49\u5beb\u8a9e\u6599\u901a\u5e38\u4e0d\u5927\uff1b\u5176\u6b21\uff0c\u5118\u7ba1\u53d6\u5f97\u8db3\u5920\u7684\u8f49\u5beb\u8a9e\u6599\u5341\u5206\u56f0</td></tr><tr><td>\u96e3\uff0c\u4f46\u53d6\u5f97\u672a\u8f49\u5beb\u8a9e\u6599\u537b\u5bb9\u6613\u5f97\u591a\uff0c\u8981\u5982\u4f55\u5229\u7528\u597d\u5927\u91cf\u7684\u672a\u8f49\u5beb\u8a9e\u6599\u4fbf\u6210\u4e86\u91cd\u8981\u7684\u554f\u984c\uff1b</td></tr><tr><td>\u6700\u5f8c\uff0c\u4e5f\u662f\u6700\u5ee3\u6cdb\u7684\u554f\u984c\uff0c\u8a13\u7df4\u8207\u6e2c\u8a66\u74b0\u5883\u7684\u4e0d\u5339\u914d\u3002\u76f8\u95dc\u7814\u7a76\u88e1\u6700\u5e38\u898b\u7684\u65b9\u6cd5\u70ba\u81ea\u6211</td></tr><tr><td>\u8a13\u7df4(Self-training)</td></tr></table>", |
| "html": null, |
| "text": "" |
| }, |
| "TABREF1": { |
| "type_str": "table", |
| "num": null, |
| "content": "<table/>", |
| "html": null, |
| "text": "\u4e2d\u5c07\u97f3\u6846\u5c64 \u7d1a\u7684\u4fe1\u5fc3\u904e\u6ffe\u5668\u52a0\u5165\u9451\u5225\u5f0f\u8a13\u7df4\u3002\u800c\u5728(Walker et al., 2017)\u5728\u9451\u5225\u5f0f\u8a13\u7df4\u4e2d\u52a0\u5165\u8a9e\u53e5\u5c64\u7d1a \u7684\u4fe1\u5fc3\u904e\u6ffe\u5668\u4ee5\u53ca\u5f8c\u8655\u7406\u6700\u4f73\u8fa8\u8b58\u7d50\u679c\u3002(Manohar et al., 2018) \u5247\u5c07\u8a5e\u5716\u52a0\u5165\u5728\u534a\u76e3\u7763\u5f0f LF-MMI \u7684\u8a13\u7df4\u3002" |
| }, |
| "TABREF2": { |
| "type_str": "table", |
| "num": null, |
| "content": "<table><tr><td/><td/><td/><td/><td/><td>\u7f85\u5929\u5b8f\u8207\u9673\u67cf\u7433 \u7f85\u5929\u5b8f\u8207\u9673\u67cf\u7433</td></tr><tr><td colspan=\"6\">\u5f0f(3)\u88e1\u6700\u7e41\u96dc\u7684\u554f\u984c\u4fbf\u5448\u73fe\u5728\u5f0f(4)\uff0c\u5f0f(4)\u70ba\u8a08\u7b97\u6240\u6709\u53ef\u80fd\u5b58\u5728\u65bc\u5047\u8aaa\u7684\u7af6\u722d\u5e8f\u5217\u3002\u5728\u8f03 \u4e86 | , \u7684\u6b0a\u91cd\u65bc\u8a5e\u5716\u4e2d\uff0c\u7528\u4ee5\u6539\u8b8a\u8a5e\u5716\u4e2d\u7684\u5206\u6578\u77e9\u9663\u3002\u5f0f(6)\u9032\u4e00\u6b65\u5316\u7c21\u6210\u4e0b\u5f0f\uff1a</td></tr><tr><td colspan=\"6\">\u65e9\u671f\u7684\u7814\u7a76\u88e1\uff0c\u5b78\u8005\u5011\u5229\u7528 CE \u4f5c\u9810\u5148\u8a13\u7df4\u9650\u5236\u5047\u8aaa\u7a7a\u9593\u7684\u5927\u5c0f\uff0c\u4f7f\u5f97 MMI \u7684\u7af6\u722d\u5e8f\u5217</td></tr><tr><td colspan=\"6\">\u53ef\u7531\u6709\u9650\u7684\u8a5e\u5716\u4e2d\u7522\u751f\u3002\u9019\u6a23\u662f\u4e8c\u968e\u6bb5\u7684\u8a13\u7df4\u53d6\u5f97\u4e86\u4e0d\u932f\u7684\u6210\u679c\uff0c\u4f46\u70ba\u4e86\u7522\u751f\u53ef\u80fd\u5e8f\u5217</td></tr><tr><td colspan=\"6\">\u7684\u8a5e\u5716\uff0c\u4e0d\u50c5\u9700\u8981\u591a\u9918\u7684 CE \u8a13\u7df4\uff0c\u4e14\u53d7\u9650\u65bc CE \u7684\u8a13\u7df4\uff0c\u7b2c\u4e8c\u968e\u6bb5\u7684 MMI \u8a13\u7df4\u50c5\u80fd\u627e</td></tr><tr><td colspan=\"6\">\u5230\u4e00\u968e\u6bb5 CE \u8a13\u7df4\u7d50\u679c\u7684\u5c40\u90e8\u6700\u4f73\u89e3\u3002LF-MMI \u4e3b\u8981\u89e3\u6c7a\u7684\u662f\u5f0f(4)\u7684\u8a08\u7b97\uff0c\u4f7f\u5f97\u4e0d\u7528\u4e00</td></tr><tr><td colspan=\"5\">\u968e\u6bb5 CE \u9810\u5148\u8a13\u7df4\u7522\u751f\u8a5e\u5716\uff0c\u5373\u53ef\u76f4\u63a5\u8a08\u7b97\u6240\u6709\u53ef\u80fd\u7684\u7af6\u722d\u8a13\u7df4\u3002</td><td/></tr><tr><td>2.2.2 LF-MMI</td><td/><td/><td/><td/><td/></tr><tr><td colspan=\"6\">\u8fd1\u5e74\u4f86\uff0c(Povey et al., 2016)\u4e2d\u63d0\u51fa LF-MMI\uff0c\u907f\u958b\u9700\u8981 CE \u8a13\u7df4\u7522\u751f\u8a5e\u5716\u7684\u5197\u9918\u6b65\u9a5f\uff0c</td></tr><tr><td colspan=\"6\">\u53ef\u8996\u70ba CTC (Graves et al., 2006)\u7684\u5ef6\u4f38\u67b6\u69cb\u3002\u4e3b\u8981\u6539\u8b8a\u6709\u56db\u7a2e\uff0c\u5229\u7528 4 \u9023\u97f3\u7d20\u8a9e\u8a00\u6a21\u578b</td></tr><tr><td colspan=\"6\">(Four-gram phone LM) \u4e14\u4e0d\u6703\u9000\u5316\u5c0f\u65bc 3 \u9023\u97f3\u7d20\u8a9e\u8a00\u6a21\u578b(Tri-gram phone LM)\uff0c\u53d6\u4ee3\u50b3 \u5728\u6709\u53c3\u8003\u8f49\u5beb\u7684\u60c5\u6cc1\u4e0b\uff0c\u50b3\u7d71 MMI \u4f30\u6e2c\u65b9\u5f0f\u70ba CML\uff0c\u8a08\u7b97\u7684\u5f0f\u5b50\u70ba\u5f0f(2)\u3002\u7136\u800c\u5728\u534a\u76e3 \u7d71\u9451\u5225\u5f0f\u8a13\u7df4\u6642\u7684\u8a5e\u5716\uff0c\u4f7f\u5f97\u641c\u5c0b\u7684\u5047\u8aaa\u7a7a\u9593\u6e1b\u5c11\uff1b\u63d0\u51fa\u591a\u7a2e\u907f\u514d\u904e\u5ea6\u64ec\u5408(Overfitting) \u7763\u5f0f\u7684\u74b0\u5883\u4e0b\uff0c\u672a\u8f49\u5beb\u8a9e\u6599\u7684\u81ea\u52d5\u8f49\u5beb(\u5206\u5b50\u9805)\u672a\u5fc5\u6b63\u78ba\u3002\u56e0\u6b64\u5728\u534a\u76e3\u7763\u5f0f\u74b0\u5883\u4e0b\uff0c\u6211 \u7684\u8a13\u7df4\u6280\u5de7\uff0c\u5982\u591a\u4efb\u52d9\u67b6\u69cb\u7684 CE \u6b63\u5247\u9805(CE-based regularization)\uff0c\u8b93\u8a13\u7df4\u80fd\u540c\u6642\u6700\u4f73\u5316 \u5011\u53ef\u5c07\u539f\u5148\u7684\u5f0f(2)\u6539\u5beb\u5982\u4e0b\uff1a LF-MMI \u548c CE\uff1b\u63a1\u7528\u985e\u4f3c CTC \u7684\u5169\u500b\u5de6\u5230\u53f3\u72c0\u614b HMM (2-state left-to-right HMM)\u7684\u62d3 \u6a38\u67b6\u69cb\uff0c\u4e14\u7b2c\u4e00\u500b\u72c0\u614b\u6c92\u6709 self-loop\uff0c\u76f8\u4f3c\u65bc CTC \u7684\u7a7a\u767d\u8f38\u51fa(Blank)\uff1b\u6700\u5f8c\u7684\u5047\u8a2d\u5247\u662f \u985e\u795e\u7d93\u7db2\u8def\u7684\u8f38\u51fa\u6c92\u6709\u8edf\u5f0f\u6700\u5927\u5316(Softmax)\uff0c\u56e0\u6b64\u4e0d\u662f\u72c0\u614b\u7684\u4e8b\u5f8c\u6a5f\u7387\uff0c\u800c\u662f\u507d\u5c0d\u6578\u53ef \u2211 log | , \u2211 | , \u2208 (5)</td></tr><tr><td colspan=\"6\">\u80fd\u6027(Pseudo log likelihood)\u3002\u524d\u5169\u8005\u7684\u6539\u8b8a\u4f7f\u5f97 MMI \u7684\u8a13\u7df4\u53ef\u5728\u4e00\u968e\u6bb5\u7684\u8072\u5b78\u6a21\u578b\u4fbf\u52a0</td></tr><tr><td colspan=\"6\">\u5165\u8a13\u7df4\uff0c\u4e14\u5f0f(4)\u4e5f\u4e0d\u662f\u8a08\u7b97\u5728\u5019\u9078\u8a5e\u5716\u4e0a\uff0c\u800c\u662f\u5b8c\u6574\u641c\u5c0b(Full search)\u6240\u6709\u7684\u53ef\u80fd\u5e8f\u5217\uff0c</td></tr><tr><td colspan=\"6\">\u6700\u7d42\u6548\u679c\u53ef\u5ab2\u7f8e\u751a\u81f3\u8d85\u8d8a\u5169\u968e\u6bb5\u7684\u8072\u5b78\u6a21\u578b\u8a13\u7df4\u3002\u5f8c\u5169\u8005\u7684\u6539\u8b8a\u5247\u662f\u6a21\u4eff CTC \u7684\u67b6\u69cb\uff0c</td></tr><tr><td colspan=\"4\">\u56e0\u6b64 LF-MMI \u4e5f\u53ef\u8996\u70ba CTC \u7684\u5ef6\u4f38\u67b6\u69cb\u3002</td><td/><td/></tr><tr><td colspan=\"6\">2.3 \u6a21\u578b\u5408\u4f75\u6280\u8853 (Model Combination) \u7576\u6211\u5011\u8a08\u7b97\u5f0f(5)\u5206\u5b50\u9805\u7684\u6b63\u78ba\u5e8f\u5217\uff0c\u8207\u904e\u5f80\u53ea\u53d6\u6700\u4f73\u8fa8\u8b58\u7d50\u679c\u7684\u8a08\u7b97\u65b9\u5f0f\u4e0d\u540c\uff0c\u800c</td></tr><tr><td colspan=\"6\">\u6574\u9ad4\u6a21\u578b\u53ef\u85c9\u7531\u591a\u500b\u6a21\u578b\u4e92\u88dc\u7684\u5047\u8aaa\u7a7a\u9593\uff0c\u7528\u4ee5\u4fee\u6b63\u55ae\u4e00\u6a21\u578b\u96e3\u4ee5\u89e3\u6c7a\u7684\u554f\u984c\u3002\u5982\u8a9e\u6599 \u662f\u5c07\u6574\u500b\u8a5e\u5716\u52a0\u5165\u8a08\u7b97\uff0c\u900f\u904e\u8a2d\u5b9a\u5149\u675f(Beam)\u4fdd\u7559\u641c\u5c0b\u6642\u7684\u6578\u91cf\u3002\u4fdd\u7559\u8d8a\u591a\u5c31\u8d8a\u53ef\u80fd\u641c</td></tr><tr><td colspan=\"6\">\u9078\u64c7(Data selection)\u3001\u76ee\u6a19\u51fd\u6578(Objective function)\u3001\u6a21\u578b(Model)\u3002\u70ba\u4e86\u5be6\u73fe\u6700\u5927\u7684\u7d44\u5408 \u5c0b\u5230\u6700\u4f73\u7b54\u6848\uff0c\u4f46\u540c\u6642\u6703\u589e\u9577\u8a08\u7b97\u8907\u96dc\u5ea6\u3002\u5176\u9918\u5be6\u9a57\u8a2d\u5b9a\u8207(Povey et al., 2016)\u4e2d\u4e00\u81f4\u3002</td></tr><tr><td colspan=\"6\">\u589e\u76ca\uff0c\u5728\u6574\u9ad4\u7cfb\u7d71\u88e1\u7684\u6a21\u578b\u5fc5\u9808\u55ae\u7368\u4e14\u6e96\u78ba(Dietterich, 2000)\u3002\u5728 DNN-HMM \u7684\u6a21\u578b\u4e2d\uff0c</td></tr><tr><td colspan=\"6\">\u53ef\u5f15\u5165\u4e94\u7a2e\u591a\u6a23\u6027\u3002\u7279\u5fb5\u591a\u6a23\u6027\uff0c\u5982\u96a8\u6a5f\u7279\u5fb5\u6295\u5f71(Random feature projection)\uff1b\u67b6\u69cb\u591a\u6a23 3.2 \u689d\u4ef6\u71b5 (Conditional Entropy) \u6027\uff0c\u5982 DNN\u3001LSTM\uff1b\u6a21\u578b\u53c3\u6578\u591a\u6a23\u6027\uff0c\u5982\u96a8\u6a5f\u521d\u59cb\u5316(Random Initialization)\uff1b\u8f38\u51fa\u76ee\u6a19 \u524d\u4e00\u6bb5\u4e2d\u63d0\u5230\u6b63\u78ba\u5e8f\u5217\u7684 \u4f86\u81ea\u65bc\u7a2e\u5b50\u6a21\u578b\u7522\u751f\u7684\u5047\u8aaa \uff0c\u6211\u5011\u4e0d\u80fd\u4fdd\u8b49\u5176\u5206\u5b50\u9805\u7684\u6b63 \u591a\u6a23\u6027\uff0c\u5982\u96a8\u6a5f\u68ee\u6797(Random forest) (Dietterich, 2000)\uff1b\u8f49\u63db\u6a21\u578b(Transition model)\u548c\u8a9e \u78ba\u6027\u3002\u56e0\u6b64\u76f4\u63a5\u52a0\u5165\u7b2c\u4e8c\u968e\u6bb5\u8a13\u7df4\u662f\u5371\u96aa\u7684\u884c\u70ba\uff0c\u751a\u81f3\u6703\u60e1\u5316\u539f\u5148\u6a21\u578b\u7684\u8868\u73fe\u3002\u5728\u904e\u5f80 \u8a00\u6a21\u578b(Language model)\u7684\u591a\u6a23\u6027\u3002\u904e\u5f80\u5728\u8a9e\u97f3\u8fa8\u8b58\u7684\u6a21\u578b\u5408\u4f75\u53ef\u5206\u70ba\u5169\u7a2e\uff0c\u5047\u8aaa\u5c64\u7d1a\u5408 \u7684\u7814\u7a76\u4e2d\u70ba\u4e86\u89e3\u6c7a\u6b64\u554f\u984c\uff0c\u6700\u5e38\u898b\u7684\u4fbf\u662f\u5728\u7b2c\u4e00\u968e\u6bb5\u548c\u7b2c\u4e8c\u968e\u6bb5\u4e2d\u9593\uff0c\u52a0\u5165\u4fe1\u5fc3\u904e\u6ffe\u5668</td></tr><tr><td colspan=\"6\">\u6392\u9664\u5206\u6578\u904e\u4f4e\u7684\u8a9e\u53e5\uff0c\u7528\u4ee5\u78ba\u4fdd\u8a13\u7df4\u8a9e\u53e5\u7684\u300c\u54c1\u8cea\u300d\uff0c\u4f46\u6311\u9078\u904e\u6ffe\u5668\u7684\u9580\u6abb\u503c\u4e26\u4e0d\u5bb9\u6613</td></tr><tr><td colspan=\"6\">\u4e14\u975e\u5e38\u6d6a\u8cbb\u8a13\u7df4\u6642\u9593\u3002\u6709\u5225\u65bc\u4ee5\u5f80\u7684\u6392\u9664\u8a13\u7df4\u8a9e\u53e5\uff0c\u6211\u5011\u5e0c\u671b\u5728\u8a13\u7df4\u6642\u4ecd\u4fdd\u7559\u5206\u6578\u8f03\u4f4e</td></tr><tr><td colspan=\"6\">\u7684\u8a9e\u53e5\uff0c\u4e26\u8207\u5206\u6578\u9ad8\u7684\u8a9e\u53e5\u4e00\u8d77\u8a13\u7df4\u3002\u9019\u88e1\u6211\u5011\u5728\u539f\u5148\u7684\u5411\u524d\u5411\u5f8c\u7b97\u6cd5(Forward-backward</td></tr><tr><td colspan=\"5\">algorithm)\u52a0\u5165\u4e86\u6b0a\u91cd\u6a5f\u5236\uff0c\u4e26\u5c07\u539f\u5148\u7684\u5f0f(1)\u6539\u5beb\u5982\u4e0b\uff1a</td><td/></tr><tr><td>\u2211</td><td>\u2208</td><td>\u2211</td><td>| , log</td><td>| ,</td><td>(6)</td></tr><tr><td colspan=\"6\">| \u4e0a\u5f0f\u70ba\u672a\u8f49\u5beb\u8a9e\u6599\u7684\u4f30\u6e2c\u65b9\u5f0f\u3002\u5f0f(6)\u8207\u5f0f(1)\u76f8\u4f3c\uff0c\u4f46\u5728\u8a08\u7b97\u53ef\u80fd\u7684\u6b63\u78ba\u5e8f\u5217 \u6642\uff0c\u52a0\u5165 (4)</td></tr></table>", |
| "html": null, |
| "text": "\u4e0a\u5f0f\u7684 u \u70ba\u8a9e\u53e5\uff0c \u70ba\u8a9e\u53e5 u \u7684\u6b63\u78ba\u72c0\u614b\u5e8f\u5217\uff0c\u4f46\u5728\u534a\u76e3\u7763\u5f0f\u74b0\u5883\u4e0b\u7684 \u4f86\u81ea\u65bc\u7a2e\u5b50\u6a21 \u578b\u7522\u751f\u7684\u5047\u8aaa \uff0c\u56e0\u6b64\u4e0d\u80fd\u4fdd\u8b49\u5176\u6b63\u78ba\u6027\u3002 \u70ba\u8a9e\u53e5 u \u7684\u8072\u5b78\u7279\u5fb5\u3002 \u70ba\u8a9e\u53e5 u \u7684\u7af6\u722d \u72c0\u614b\u5e8f\u5217\uff0c\u65e9\u671f\u7684\u8072\u5b78\u6a21\u578b\u900f\u904e CE \u7b2c\u4e00\u968e\u6bb5\u7684\u8a13\u7df4\u9650\u5236\u7522\u751f\u7af6\u722d\u5e8f\u5217\u7684\u5047\u8aaa\u7a7a\u9593\uff0c\u4f7f \u5f97\u7af6\u722d\u7684\u5e8f\u5217\u53ea\u80fd\u5f9e CE \u8a13\u7df4\u5f8c\u7684\u8a5e\u5716\u4e2d\u7522\u751f\u3002\u800c LF-MMI \u900f\u904e\u4e00\u4e9b\u5be6\u4f5c\u4e0a\u7684\u6a5f\u5236\u907f\u958b \u4e0a\u8ff0\u5197\u9918\u7684\u6b65\u9a5f\uff0c\u53ef\u4ee5\u5728\u8a13\u7df4\u6642\u76f4\u63a5\u8a08\u7b97\u6240\u6709\u7684\u7af6\u722d\u5e8f\u5217\u3002" |
| }, |
| "TABREF3": { |
| "type_str": "table", |
| "num": null, |
| "content": "<table><tr><td>\u7d50\u5408\u9451\u5225\u5f0f\u8a13\u7df4\u8207\u6a21\u578b\u5408\u4f75\u65bc\u534a\u76e3\u7763\u5f0f\u8a9e\u97f3\u8fa8\u8b58\u4e4b\u7814\u7a76 \u7f85\u5929\u5b8f\u8207\u9673\u67cf\u7433 27 \u7d50\u5408\u9451\u5225\u5f0f\u8a13\u7df4\u8207\u6a21\u578b\u5408\u4f75\u65bc\u534a\u76e3\u7763\u5f0f\u8a9e\u97f3\u8fa8\u8b58\u4e4b\u7814\u7a76 29 \u7f85\u5929\u5b8f\u8207\u9673\u67cf\u7433 \u7d50\u5408\u9451\u5225\u5f0f\u8a13\u7df4\u8207\u6a21\u578b\u5408\u4f75\u65bc\u534a\u76e3\u7763\u5f0f\u8a9e\u97f3\u8fa8\u8b58\u4e4b\u7814\u7a76 31</td></tr><tr><td>\u8868 2. AMI \u6703\u8b70\u4e4b\u8a13\u7df4\u3001\u767c\u5c55\u8207\u6e2c\u8a66\u96c6 5.2 \u5be6\u9a57\u7d50\u679c\u8207\u5206\u6790 (Results and Discussion) \u4ee5\u770b\u51fa\u9019\u5169\u7a2e\u65b9\u6cd5\u7684\u6cdb\u7528\u6027\uff0c\u4e0d\u6703\u53d7\u5230\u4e0d\u540c\u7684\u8a13\u7df4\u6e96\u5247\u5f71\u97ff\u9032\u6b65\u6210\u6548\u3002\u53e6\u4e00\u65b9\u9762\uff0c\u5f9e\u5408 \u74b0\u5883\u4e0b\uff0c\u6211\u5011\u7686\u53ef\u900f\u904e\u7c21\u55ae\u5730\u6539\u8b8a\u8d85\u53c3\u6578\uff0c\u518d\u4ee5\u97f3\u6846\u5c64\u7d1a\u8207\u5047\u8aaa\u5c64\u7d1a\u7684\u5408\u4f75\u9054\u5230\u66f4\u597d\u7684</td></tr><tr><td>\u4f75\u7684\u89c0\u9ede\u4f86\u770b\uff0c\u97f3\u6846\u5c64\u7d1a\u5408\u4f75\u5728\u5404\u500b\u968e\u6bb5\uff0c\u6bd4\u8d77\u55ae\u4e00\u7cfb\u7d71\u7684\u6e96\u5ea6\u7686\u80fd\u63d0\u5347 0.5 \u81f3 1.5 \u7684 \u8fa8\u8b58\u7d50\u679c\u3002</td></tr><tr><td>\u8a9e\u6599\u55ae\u4f4d 5.2.1 \u52a0\u5165NCE\u8207\u8a5e\u5716\u7684\u5f71\u97ff (NCE and Lattice for supervision) \u8a13\u7df4\u96c6 \u767c\u5c55\u96c6 \u6e2c\u8a66\u96c6 1 \u6e2c\u8a66\u96c6 2 WER\uff0c\u8b49\u660e\u4e86\u5408\u4f75\u6a21\u578b\u7684\u6280\u8853\u61c9\u7528\u65bc\u534a\u76e3\u7763\u5f0f\u74b0\u5883\u7684\u6709\u6548\u6027\u3002\u53e6\u4e00\u65b9\u9762\uff0c\u5047\u8aaa\u5c64\u7d1a\u7684\u5408 \u7e3d\u8a08 \u8868 6. \u4e0d\u540c\u534a\u76e3\u7763\u6e96\u5247\u7684\u6a21\u578b\u5408\u4f75</td></tr><tr><td>\u5c0f\u6642\u6578 \u8868 3 \u4e2d\u5448\u73fe\u7684\u662f\u662f\u5426\u52a0\u5165 NCE \u6b0a\u91cd\u548c\u8a5e\u5716\u65bc\u8a13\u7df4\u4e2d\u7684\u7d50\u679c\u3002\u7b2c\u4e8c\u6b04\u4e2d\u7684 lm-scale \u70ba\u7b2c\u4e8c 70.09 7.81 8.71 8.97 95.79 \u4f75\u6548\u679c\u66f4\u52dd\u65bc\u97f3\u6846\u5c64\u7d1a\u7684\u5408\u4f75\uff0c\u9019\u6a23\u7684\u7d50\u679c\u6b78\u529f\u65bc\u5047\u8aaa\u5408\u4f75\u662f\u505a\u5728\u5404\u500b ASR \u7684\u8a5e\u5716\u4e0a\uff0c [Table 6. Results on model combination in conjunction with different semi-supervised (Huang & Hasegawa-Johnson, 2010)\u3002\u6211\u5011\u53ef\u4ee5 \u7a31\u5f0f(7)\u70ba\u7d66\u4e88\u6a21\u578b\u53c3\u6578 \u548c\u8072\u5b78\u7279\u5fb5 \u689d\u4ef6\u4e0b\uff0c\u53c3\u8003\u8f49\u5beb\u5e8f\u5217 \u7684\u689d\u4ef6\u71b5 | , \u3002\u5f0f \u8a9e\u53e5\u6578 97,222 10,882 13,059 12,612 \u968e\u6bb5\u7684\u8a9e\u97f3\u6a21\u578b\u91cd\u65b0\u8a55\u5206\u7684\u7e2e\u653e\u5e38\u6578\u3001Beam \u70ba\u641c\u5c0b\u8a5e\u5716\u4fdd\u7559\u7684\u7a2e\u5b50\u500b\u6578\u3001Tol \u70ba\u5728\u8a13\u7df4 \u800c\u97f3\u6846\u5408\u4f75\u5247\u662f\u985e\u795e\u7d93\u7db2\u8def\u7684\u8f38\u51fa\u3002\u6bd4\u8d77\u97f3\u6846\u5408\u4f75\uff0c\u5047\u8aaa\u5c64\u7d1a\u7684\u5408\u4f75\u66f4\u63a5\u8fd1\u65bc\u8fa8\u8b58\u76ee\u6a19 criteria] 133,775 \u6642\u7528\u5230\u7684\u8a5e\u5716\u7684\u5141\u8a31\u97f3\u6846\u4f4d\u79fb(Frame shift)\uff0c1 \u4ee3\u8868 30ms\uff0c\u9019\u88e1\u57fa\u65bc\u7d93\u9a57\u4e0a\u7684\u8a2d\u7f6e\uff0c\u4e26\u7121 \u7684\u8a5e\uff0c\u56e0\u6b64\u6548\u679c\u8f03\u597d\u3002\u4f46\u53e6\u4e00\u65b9\u9762\uff0c\u97f3\u6846\u5408\u4f75\u96d6\u7136\u5728\u6548\u679c\u4e0a\u7565\u8f38\u5047\u8aaa\u5408\u4f75\uff0c\u4f46\u7531\u65bc\u53ef\u76f4 F-COMB H-COMB (7)\u7684\u6539\u8b8a\u53ef\u5229\u7528\u8cc7\u8a0a\u91cf\u5c0d\u8f49\u5beb\u7684\u300c\u54c1\u8cea\u300d\u5efa\u6a21\uff0c\u4e26\u4e14\u81ea\u7136\u5730\u52a0\u5165 LF-MMI \u76ee\u6a19\u51fd\u6578\uff0c\u5728 \u4e0d\u7528\u4fe1\u5fc3\u904e\u6ffe\u5668\u7684\u60c5\u6cc1\u4e0b\u4e5f\u80fd\u63d0\u5347\u8a13\u7df4\u7d50\u679c\u3002 4. \u6a21 \u578b \u5408 \u4f75 \u6280 \u8853 \u61c9 \u7528 \u65bc \u8072 \u5b78 \u6a21 \u578b (MODEL COMBINATION OF ACOUSITIC MODELING) \u6a21\u578b\u5408\u4f75\u7684\u6210\u679c\u53ef\u900f\u904e\u4fee\u6b63\u5404\u5225\u6a21\u578b\u7684\u932f\u8aa4\u3001\u6e1b\u5c11\u8f03\u5dee\u9078\u64c7\u7684\u53ef\u80fd\u6027\u3001\u589e\u52a0\u6a21\u578b\u641c\u5c0b\u6642 \u53ef\u7528\u65bc\u6458\u8981\u3001\u60c5\u7dd2\u8207\u5c0d\u8a71\uff1b\u6700\u5f8c\u662f\u8a9e\u97f3\u7684\u90e8\u5206\uff0c\u53ef\u5206\u70ba\u8033\u639b\u5f0f\u8fd1\u8ddd\u96e2\u9ea5\u514b\u98a8\u3001\u56fa\u5b9a\u5f0f\u9060 \u8ddd\u96e2\u9ea5\u514b\u98a8\u3002\u672c\u5be6\u9a57\u53ea\u7528\u5230\u4e86\u8a9e\u97f3\u8a9e\u6599\u3002\u8868 2 \u70ba AMI \u7684\u57fa\u672c\u7d71\u8a08\u6578\u64da\uff0c\u7531\u65bc\u539f\u5148 AMI \u7684\u8a13\u7df4\u4e2d\u4e26\u6c92\u6709\u7528\u5230\u767c\u5c55\u96c6\uff0c\u56e0\u6b64\u5be6\u969b\u8a13\u7df4\u96c6\u70ba\u8a13\u7df4\u96c6\u52a0\u767c\u5c55\u96c6\u3002\u6211\u5011\u7528\u8a5e\u932f\u8aa4\u7387(Word error rate, WER)\u548c\u8a5e\u4fee\u5fa9\u7387(WER recovery rate, WRR)\u4f5c\u70ba\u8a55\u4f30\u3002WRR \u5982\u4e0b\uff1a WRR \u7279\u5730\u8abf\u52d5\u3002\u7b2c\u4e00\u6b04\u7531\u4e0a\u5230\u4e0b\uff0cBaseline \u662f\u53ea\u7528 16 \u5c0f\u6642\u8a13\u7df4\u7684\u8072\u5b78\u6a21\u578b\uff1bNo weight (NW) \u63a5\u5728\u97f3\u6846\u968e\u6bb5\u5c31\u5408\u4f75\uff0c\u4e0d\u9700\u8981\u5404\u5225 ASR \u7522\u751f\u8a5e\u5716\uff0c\u56e0\u6b64\u6709\u66f4\u597d\u7684\u5373\u6642\u6027\u3002 Dev Eval WRR Dev Eval WRR \u5247\u662f\u76f4\u63a5\u52a0\u5165 62 \u5c0f\u6642\u672a\u8f49\u5beb\u7684\u8a9e\u6599\uff1bBest Phone Path(BPP)\u57fa\u65bc NW \u4e4b\u4e0a\uff0c\u52a0\u5165 NCE \u7684 \u6b0a\u91cd\u7684\u6700\u4f73\u8fa8\u8b58\u7d50\u679c\uff1bLattice for supervision (LS)\u5247\u662f\u57fa\u65bc BPP \u4e26\u52a0\u5165\u8a5e\u5716\u3002\u70ba\u4e86\u8a08\u7b97\u65b9 \u8868 4. \u4e0d\u540c\u7db2\u8def\u7684\u8a2d\u5b9a\u5dee\u7570 +NW 25.9 26.1 35% 25.7 26.0 39% [Table 4. The table shows that different training criteria in the experiment. We combine four TDNN model generated by different random seed at both frame level TDNN-0 +BPP 25.4 25.5 48% 25.3 25.5 50% \u4fbf\uff0cLS \u5c07\u539f\u5148\u7684\u8a5e\u5716\u5207\u6210 1.5 \u79d2\u7684\u584a(Chunks)\uff0c\u518d\u7528\u5411\u524d\u5411\u5f8c\u7b97\u6cd5\u8a08\u7b97\uff1bOracle \u5247\u662f\u5c07 78 \u5c0f\u6642\u7684\u8a9e\u6599\u76f4\u63a5\u52a0\u5165\u8a13\u7df4\u3002Dev \u548c Eval \u5206\u5225\u662f\u6e2c\u8a66\u96c6 1 \u548c\u6e2c\u8a66\u96c6 2\u3002\u5f9e\u5be6\u9a57\u4e2d\u7684\u7d50\u679c and hypothesis level.] +LS 25.1 25.1 57% 24.9 25.0 60% (10) \u53ef\u4ee5\u770b\u51fa\uff0c\u76f4\u63a5\u9032\u884c\u50b3\u7d71\u4e8c\u968e\u6bb5\u8a13\u7df4\u6642\uff0c\u4fbf\u53ef\u7a0d\u5fae\u63d0\u5347 WRR \u70ba 24%\uff0c\u9019\u8981\u6b78\u529f\u65bc\u826f\u597d TDNN0 TDNN1 TDNN2 TDNN3 \u7684\u5047\u8aaa\u7a7a\u9593\u4f86\u9054\u5230\u66f4\u597d\u7684\u6a21\u578b\u6548\u80fd\u3002\u5728\u8072\u5b78\u6a21\u578b\u7684\u5408\u4f75\u53ef\u5206\u70ba\u97f3\u6846\u5c64\u7d1a\u548c\u5047\u8aaa\u5c64\u7d1a\u3002\u5169 \u8005\u7684\u6bd4\u8f03\u7d00\u9304\u65bc\u8868 1\uff0c\u524d\u8005\u56e0\u70ba\u662f\u5728\u8072\u5b78\u6a21\u578b\u7684\u8f38\u51fa\u76f4\u63a5\u5408\u4f75\uff0c\u56e0\u6b64\u5177\u6709\u8f03\u5feb\u7684\u5373\u6642\u6027\u3002 \u5f8c\u8005\u5247\u662f\u5728\u6a21\u578b\u7522\u751f\u8a5e\u5716\u5f8c\u5408\u4f75\uff0c\u8207\u89e3\u78bc\u6a19\u6e96\u66f4\u76f8\u95dc\uff0c\u6709\u8f03\u597d\u7684\u8fa8\u8b58\u7d50\u679c\u3002 \u8868 1. \u5408\u4f75\u65b9\u5f0f\u6bd4\u8f03 \u9762\u5411 \u97f3\u6846\u5c64\u7d1a\u5408\u4f75 \u5047\u8aaa\u5c64\u7d1a\u5408\u4f75 \u89e3\u78bc\u8a5e\u5716 \uf0b7 \u5f37\u5236\u5171\u4eab\u8a5e\u5716\u4e2d\u7684\u6642\u9593\u540c\u6b65\u7684 \u72c0\u614b \uf0b7 \u50c5\u9700\u8655\u7406\u6574\u9ad4\u7684\u55ae\u500b\u8a5e\u5716 \uf0b7 \u4e0d\u9700\u8981\u6642\u9593\u540c\u6b65\u7684\u72c0\u614b \uf0b7 \u9700\u5148\u5404\u5225\u8655\u7406\u6574\u9ad4\u6a21\u578b\u6578\u7684\u8a5e \u5716\u518d\u5408\u4f75 \u4e8b\u5f8c\u6a5f\u7387 \uf0b7 \u65e8\u5728\u7522\u751f\u66f4\u597d\u7684\u97f3\u6846\u4e8b\u5f8c\u6a5f\u7387 \u6216\u89c0\u5bdf\u53ef\u80fd\u6027\uff0c\u5f9e\u800c\u7522\u751f\u66f4\u597d \u7684\u8a5e\u5716 \uf0b7 \u65e8\u5728\u7522\u751f\u66f4\u597d\u7684\u5047\u8aaa\u4e8b\u5f8c\u6a5f \u7387\uff0c\u5176\u8207\u89e3\u78bc\u6a19\u6e96\u66f4\u5bc6\u5207\u76f8\u95dc \u5716 1. \u5169\u7a2e\u5c64\u7d1a\u7684\u8072\u5b78\u5206\u6578\u5408\u4f75\u3002\u5de6\u65b9\u70ba\u97f3\u6846\u5c64\u7d1a\uff0c\u53f3\u65b9\u70ba\u5047\u8aaa\u5c64\u7d1a\u3002 \u5047\u8aaa\u5c64\u7d1a\u5408\u4f75\u5247\u662f\u5229\u7528 ASR \u7cfb\u7d71\u7d93\u904e\u4e00\u822c\u7684\u89e3\u78bc\u6a5f\u5236\u7522\u751f\u7684\u8a5e\u5716\uff0c\u7d66\u4e88\u4e0d\u540c\u6b0a\u91cd\u548c\u640d\u5931 \u51fd\u6578\u9032\u884c\u5408\u4f75\u3002\u76f8\u8f03\u65bc\u97f3\u6846\u5c64\u7d1a\u5408\u4f75\uff0c\u5047\u8aaa\u5c64\u7d1a\u5408\u4f75\u53ef\u5141\u8a31\u975e\u540c\u6b65\u6642\u9593\u8f38\u51fa\uff0c\u4f46\u56e0\u70ba\u9700 \u8981\u5408\u4f75\u5404\u5225 ASR \u7684\u8f38\u51fa\u7d50\u679c\uff0c\u56e0\u6b64\u8f03\u70ba\u8cbb\u6642\u3002 * argmin \u2211 , \u2211 | , (9) \u5176\u4e2dh \u70ba\u5404\u5225 ASR \u7cfb\u7d71\u89e3\u78bc\u6642\u7522\u751f\u7684\u8a5e\u5e8f\u5217\u3002 \u70ba\u5408\u4f75\u7684\u6a21\u578b\u7e3d\u6578\u3002 \u70ba\u5404\u5225\u6a21\u578b\u6df7\u548c \u6b0a\u91cd\uff0c\u4e14 \u22650 ,\u2211 1 \u3002 \u70ba\u8a5e\u5c64\u7d1a\u7684\u640d\u5931\u51fd\u6578\uff0c\u9019\u88e1\u4f7f\u7528\u7de8\u8f2f\u8ddd\u96e2(Edit distance)\u3002 \u5f0f(9)\u53ef\u7406\u89e3\u70ba\u5404\u5225 ASR \u7522\u751f\u8a5e\u5716\u7684\u806f\u96c6\uff0c\u4e26\u900f\u904e\u6700\u5c0f\u5316\u8c9d\u5f0f\u6c7a\u7b56\u98a8\u96aa\u5c0d\u5408\u4f75\u7684\u8a5e\u5716\u89e3 \u78bc\u3002\u97f3\u6846\u5c64\u7d1a\u548c\u5047\u8aaa\u5c64\u7d1a\u7684\u793a\u610f\u5716\u53ef\u53c3\u8003\u5716 1\u3002 \u7684 LF-MMI \u7a2e\u5b50\u6a21\u578b\uff0c\u53ea\u7528 16 \u5c0f\u6642\u7684\u8a13\u7df4\u8a9e\u6599\u4fbf\u9054\u5230\u5c1a\u53ef\u7684\u8fa8\u8b58\u7387\uff0c\u4f46 WRR \u63d0\u5347\u4e0d\u5920 \u53ef\u80fd\u662f\u672a\u8f49\u5beb\u8a9e\u6599\u76f8\u5c0d\u4f7f\u7528\u592a\u5c11\u5c0e\u81f4\u4fee\u5fa9\u7387\u4e0d\u4f73\u3002\u4f7f\u7528 NCE \u53ef\u9032\u4e00\u6b65\u5730\u63d0\u5347 WRR \u81f3 \u8a2d\u5b9a\u5dee\u7570 (\u8207 TDNN0) \u57fa\u65bc\u8868 2 \u7684\u8a2d \u5b9a +Proportional shrink \u8207 TDNN1 \u521d\u59cb 6. \u7d50\u8ad6 (CONCLUSION AND FUTURE WORK) +L2-regularization \u5316\u4e0d\u540c 5.1.1 \u534a\u76e3\u7763\u5f0f\u5be6\u9a57\u6d41\u7a0b\u8207\u8a2d\u5b9a (Semi-supervised Setup) \u672c\u5be6\u9a57\u5c07 AMI \u539f\u5148\u7684\u8a13\u7df4\u96c6\u5207\u5272\u6210 16 \u5c0f\u6642\u7684\u76e3\u7763(\u8f49\u5beb)\u8a9e\u6599\u548c 62 \u5c0f\u6642\u7684\u975e\u76e3\u7763(\u672a\u8f49\u5beb) 33%\uff0c\u9019\u4e5f\u8b49\u660e\u4e86\u52a0\u5165\u97f3\u6846\u5c64\u7d1a\u7684\u689d\u4ef6\u71b5\u80fd\u6709\u6548\u8f14\u52a9\u534a\u76e3\u7763\u5f0f\u8a13\u7df4\uff0c\u56e0\u6b64\u5f8c\u7e8c\u5be6\u9a57\u63a2\u8a0e \u8868 5. \u97f3\u6846\u548c\u5047\u8aaa\u5c64\u7d1a\u7684\u8072\u5b78\u5206\u6578\u7d50\u5408 \u7686\u6703\u4ee5 NCE \u70ba\u4e3b\uff1b\u6700\u5f8c\u5247\u662f\u52a0\u5165\u6574\u500b\u8a5e\u5716\u5f8c\u53ef\u5c07 WRR \u63d0\u5347\u81f3 45%\uff0c\u53ef\u770b\u51fa\u591a\u4fdd\u7559\u5e7e\u500b [Table 5. Results on model combinations including frame-level and hypothesislevel \u8a9e\u6599\uff0c\u767c\u5c55\u96c6\u548c\u6e2c\u8a66\u96c6\u3002\u6574\u9ad4\u5be6\u9a57\u7684\u8a13\u7df4\u70ba\u5169\u968e\u6bb5\uff0c\u7b2c\u4e00\u968e\u6bb5\u70ba\u5229\u7528 16 \u5c0f\u6642\u7684\u76e3\u7763\u8a9e\u6599 \u641c\u5c0b\u7684\u53ef\u80fd\u6027\u5f8c\uff0c\u589e\u52a0\u7684\u8a08\u7b97\u7a7a\u9593\u80fd\u8f14\u52a9\u6a21\u578b\u7684\u8a13\u7df4\uff0c\u9032\u4e00\u6b65\u63d0\u5347\u8fa8\u8b58\u7d50\u679c\u3002 combination.] \u8a13\u7df4\u7a2e\u5b50\u6a21\u578b\uff0c\u4ee5\u53ca\u518d\u4f7f\u7528 62 \u5c0f\u6642\u7684\u975e\u76e3\u7763\u8a9e\u6599\u63d0\u5347\u6a21\u578b\u6548\u80fd\u3002\u6574\u9ad4\u5be6\u9a57\u7684\u8a73\u7d30\u67b6\u69cb\u53ef \u8868 3. \u52a0\u5165 NCE \u8207\u8a5e\u5716\u7684\u5f71\u97ff TDNN1 TDNN1 TDNN2 TDNN3 FCOMB HCOMB \u7684\u7814\u7a76\u65b9\u5411\u6703\u91dd\u5c0d\u6709\u6548\u6027\u8207\u5373\u6642\u6027\u5169\u500b\u65b9\u5411\u7e7c\u7e8c\u7814\u7a76\u3002\u6839\u64da\u9019\u6b21\u7684\u5be6\u9a57\u7d50\u679c\uff0c\u6211\u5011\u5f97\u77e5 \u53c3\u8003\u5716 2\u3002LF-MMI \u7684\u8a2d\u5b9a\u8207(Povey et al., 2016)\u4e00\u6a23\uff0c\u7279\u5fb5\u662f 40 \u7dad MFCC \u548c 100 \u7dad\u7684 i-vector\uff0c\u985e\u795e\u7d93\u7db2\u8def\u662f\u4f7f\u7528\u6642\u9593\u5ef6\u9072\u7db2\u8def(Time-delay neural network, TDNN) (Peddinti, Povey & Khudanpur, 2015)\u3002\u5be6\u9a57\u5206\u70ba\u8a13\u7df4\u6e96\u5247\u7684\u6709\u6548\u6027\uff0c\u4ee5\u53ca\u5f8c\u8655\u7406\u7684\u6a21\u578b\u5408\u4f75\u3002\u9019\u88e1 [Table 3. Negative conditional entropy and lattice for supervision] Supervision lm-scale Beam Tol Dev Eval WRR Baseline ---27.2 27.8 -Test Test Test Test Test \u53ef\u900f\u904e\u6a21\u64ec\u4e0d\u78ba\u5b9a\u6027\u6216\u66f4\u6539\u53c3\u6578\u7684\u6a21\u578b\u5408\u4f75\u63d0\u5347\u6e96\u5ea6\uff0c\u672a\u4f86\u6703\u7e7c\u7e8c\u671d\u5982\u4f55\u5229\u7528\u672a\u8f49\u5beb\u8a9e Test Baseline 27.5 26.7 26.5 26.5 \u6599\u8207\u4e92\u88dc\u591a\u6a23\u6027\u7684\u5408\u4f75\u7e7c\u7e8c\u7814\u7a76\uff0c\u5982 1) \u5229\u7528\u4e0d\u540c\u6a21\u578b\u7a2e\u985e\uff0c\u4ee5\u7522\u751f\u66f4\u597d\u7684\u4e92\u88dc\u6027\u3002\u53e6 25.6 25.5 \u4e00\u65b9\u9762\uff0c2) \u8f49\u5beb\u8a9e\u6599\u8207\u672a\u8f49\u5beb\u8a9e\u6599\u7684\u6bd4\u4f8b\uff0c\u8981\u5230\u591a\u5c11\u624d\u80fd\u9054\u6210\u6700\u597d\u7684 WRR\uff1b\u518d\u8005\uff0c\u5118 \u9700\u8981\u6ce8\u610f\u7684\u662f\u5408\u4f75\u6642\u7684\u6b0a\u91cd\u7686\u70ba\u6a21\u578b\u6578\u91cf\u7684\u5012\u6578(e.g. M \u500b\u6a21\u578b\uff0c\u6b0a\u91cd\u70ba 1/M) \u3002 NW 0 0 1 26.2 26.8 BPP 26.1 25.2 25.5 25.1 24.5 24.4 \u7ba1\u9019\u6b21\u900f\u904e\u6a21\u578b\u5408\u4f75\u5f97\u5230\u4e86\u4e0d\u932f\u7684\u7d50\u679c\uff0c\u4f46\u540c\u6642\u4e5f\u4ed8\u51fa\u76f8\u8f03\u65bc\u55ae\u4e00 ASR \u7cfb\u7d71\u66f4\u9ad8\u6602\u7684\u904b 24% BPP 0 0 1 26.0 26.2 LS 25.6 25.1 25.5 24.9 24.4 24.2 \u7b97\u8cc7\u6e90\uff0c\u5373\u4fbf\u662f\u76f8\u8f03\u65bc\u5047\u8aaa\u5408\u4f75\u8f03\u70ba\u8f15\u91cf\u7684\u97f3\u6846\u5408\u4f75\u4e5f\u662f\u5982\u6b64\u3002\u56e0\u6b64 3) \u672a\u4f86\u6703\u52a0\u5165\u6a21 33% LS 0.5 4 1 25.5 25.7 45% Oracle 23.3 22.5 22.8 22.5 21.5 \u578b\u58d3\u7e2e(Model combination)\u7684\u6280\u8853\uff0c\u671f\u8a31\u6709\u4e00\u5929\u80fd\u5920\u4ee5\u5c11\u91cf\u7684\u8f49\u5beb\u8a9e\u6599\u4fbf\u9054\u5230\u6709\u6548\u4e14\u5373\u6642 21.3 \u7684\u8fa8\u8b58\u7d50\u679c\u3002</td></tr><tr><td>4.1 \u97f3\u6846\u5c64\u7d1a\u5408\u4f75 (Frame-level Combination) \u97f3\u6846\u5c64\u7d1a\u5408\u4f75\u662f\u6839\u64da\u67d0\u500b\u6642\u9593\u9ede\u4e2d\u97f3\u6846\u8f38\u51fa\u7684\u5c0d\u6578\u53ef\u80fd\u6027(Log likelihood)\uff0c\u7d66\u4e88\u4e0d\u540c\u7684 \u6b0a\u91cd\u5f8c\u5408\u4f75\u3002\u56e0\u70ba\u5408\u4f75\u7684\u662f\u97f3\u6846\uff0c\u6240\u4ee5\u5fc5\u9808\u4fdd\u6301\u8f38\u51fa\u6642\u9593\u7684\u540c\u6b65\u3002\u8072\u5b78\u6a21\u578b\u7684\u5c0d\u6578\u53ef\u80fd \u6027\u662f\u985e\u795e\u7d93\u7db2\u8def\u7684\u8f38\u51fa\u3002\u5f0f\u5b50\u5982\u4e0b\uff1a | , \u2211 | , (8) \u5f0f(8)\u4e2d \u70ba\u985e\u795e\u7d93\u7db2\u8def\u7684\u8f38\u51fa\uff0c\u4ee3\u8868\u8a9e\u53e5 u \u5728\u6642\u9593\u9ede t \u7684\u72c0\u614b S \u7684\u6a5f\u7387\u3002M \u70ba\u5408\u4f75\u7684\u6a21 \u578b\u7e3d\u6578\uff0c \u70ba\u5404\u5225\u6a21\u578b\u6df7\u548c\u6b0a\u91cd\uff0c\u4e14 0 , \u2211 1 \u3002\u03b1\u76f8\u4f3c\u65bc\u5728(Dietterich, 2000) \u4e2d\u7684\u5229\u7528\u5c0d\u89d2\u7dda\u77e9\u9663(Diagonal matrices)\u5c0d\u5404\u5225\u6a21\u578b\u7684\u7dda\u6027\u5408\u4f75(Linear ensemble)\u3002\u5408\u4f75\u5f8c \u7684\u97f3\u6846\u4e8b\u5f8c\u6a5f\u7387(Frame posterior)\u6703\u7576\u6210\u96b1\u85cf\u5f0f\u99ac\u53ef\u592b\u6a21\u578b(Hidden Markov model, HMM) \u7684\u8072\u5b78\u7279\u5fb5 O \u7684\u5c0d\u6578\u53ef\u80fd\u6027\uff0c\u4e26\u9032\u884c\u6a19\u6e96\u7684\u89e3\u78bc\u7a0b\u5e8f\u3002 5. \u5be6\u9a57 (EXPERIMENTS) 5.1 \u5be6\u9a57\u8a2d\u5b9a (Experimental Setup) \u5be6 \u9a57 \u4f7f \u7528 Kaldi (Povey et al., 2011) \u8a9e \u97f3 \u8b58 \u5225 \u5de5 \u5177 \u5305 \u3002 \u8a9e \u6599 \u5eab \u70ba AMI (Augmented Multi-party Interaction) (McCowan et al., 2005)\u3002AMI \u8a9e\u6599\u5eab\u662f\u4f86\u81ea\u6b50\u76df\u767c\u8d77\u7684\u6703\u8b70\u700f\u89bd (Meeting browser)\u8a08\u756b\uff0c\u5176\u4e2d\u5305\u542b\u60c5\u5883\u6703\u8b70 (Scenario meetings)\u548c\u975e\u60c5\u5883\u7684\u6703\u8b70\uff0c\u60c5\u5883\u6703 \u8b70\u662f\u6307\u660e\u78ba\u7684\u6703\u8b70\u76ee\u6a19\u3001\u6703\u8b70\u9593\u5f7c\u6b64\u6709\u95dc\u9023\uff0c\u5982\u5176\u4e2d\u4e00\u500b\u6703\u8b70\u7684\u4e3b\u984c\u70ba\u8a0e\u8ad6\u96fb\u8996\u9059\u63a7\u5668 \u7684\u8a2d\u8a08\uff1b\u53e6\u4e00\u65b9\u9762\u7684\u975e\u60c5\u5883\u6703\u8b70 (Non-scenario meetings)\u5247\u53cd\u4e4b\uff0c\u8f03\u6c92\u6709\u660e\u78ba\u7684\u4e3b\u984c\uff0c\u4e3b \u8981\u70ba\u82f1\u570b\u611b\u4e01\u5821\u5927\u5b78\u3001\u745e\u58eb Idiap \u7814\u7a76\u4e2d\u5fc3\u3001\u8377\u862d TNO \u4eba\u70ba\u56e0\u7d20\u7814\u7a76\u6240\u7684\u5b78\u751f\u6216\u7814\u7a76\u8005 \u7d44\u6210\u8a0e\u8ad6\u7684\u5c0f\u578b\u6703\u8b70\uff0c\u5982\u7dda\u6027\u4ee3\u6578\u3001\u5fae\u7a4d\u5206\u7b49\u3002AMI \u7684\u8a9e\u6599\u5eab\u4e5f\u5305\u542b\u4e86\u5f71\u50cf\u3001\u6587\u5b57\u3001\u8a9e \u97f3\uff0c\u5f71\u50cf\u7d00\u9304\u7684\u662f\u6703\u8b70\u8996\u89d2\u3001\u6295\u5f71\u6a5f\u756b\u9762\u548c\u767d\u677f\u66f8\u5beb\u8a18\u9304\uff1b\u6587\u5b57\u6709\u8a9e\u97f3\u8f49\u5beb\u3001\u5c0d\u8a71\u7279\u6027\uff0c \u5716 2. \u6574\u9ad4\u5be6\u9a57\u67b6\u69cb [Figure 2. A Flow chart of the experimental design.] Oracle ---23.5 23.1 -Proportional shrink \u548c L2-regularization \u4f86\u5f97\u5c0f\u3002\u5118\u7ba1\u5982\u6b64\uff0c\u4e0d\u8ad6\u662f\u90a3\u7a2e\u65b9\u5f0f\uff0c\u5728\u534a\u76e3\u7763\u5f0f \u5408\u7684\u554f\u984c\u3002\u53e6\u4e00\u65b9\u9762\uff0c\u96d6\u7136\u5728 WER \u4e0a\u770b\u5230\u6210\u6548\uff0c\u4e0d\u904e\u5728 WRR \u4e0a\u5247\u6c92\u6709\u986f\u8457\u5dee\u7570\u3002\u9019\u53ef (\u8868 5)\u4f86\u5f97\u66f4\u597d\u3002\u56e0\u6b64\u6211\u5011\u53ef\u5f97\u77e5\uff0c\u534a\u76e3\u7763\u6e96\u5247\u5728\u5404\u5225\u6e96\u78ba\u8207\u591a\u6a23\u6027\u4e0a\uff0c\u76f8\u8f03\u4f7f\u7528 \u932f\u8aa4\u7387 1%\uff0c\u800c\u8abf\u6574\u521d\u59cb\u5316\u4e5f\u6703\u4e9b\u5fae\u5730\u5f71\u97ff WER\uff0c\u4e5f\u518d\u4e00\u6b21\u8b49\u5be6\u4e86 LF-MMI \u5bb9\u6613\u904e\u5ea6\u64ec \u4e2d\u89c0\u5bdf\u5230 WER \u9032\u6b65\u7684\u5e45\u5ea6\u7d04\u70ba 0.5\uff0c\u6c92\u6709\u6bd4\u5404\u5225\u8a13\u7df4\u591a\u500b\u6a21\u578b\u4e26\u5728\u540c\u500b\u534a\u76e3\u7763\u6e96\u5247\u5408\u4f75 \u4e0a\u8ff0\u5be6\u9a57\u4e2d\u53ef\u770b\u51fa\u52a0\u5165 Proportional shrink \u548c L2-regularization \u53ef\u6bd4\u539f\u5148\u7684\u8a13\u7df4\u518d\u964d\u4f4e\u8a5e WRR \u70ba 60.8%\u3002\u9019\u4e9b\u4e0d\u540c\u6e96\u5247\u7684\u5408\u4f75\u96d6\u6709\u52a9\u65bc WER \u8207 WRR \u7684\u63d0\u5347\uff0c\u4f46\u53ef\u5f9e\u5be6\u9a57\u7d50\u679c \u5247\u5206\u5225\u662f\u97f3\u6846\u5c64\u7d1a\u548c\u5047\u8aaa\u5c64\u7d1a\u7684\u5408\u4f75\uff1bTest \u5247\u662f Dev \u548c Eval \u4e4b WER \u7684\u76f8\u52a0\u53d6\u5e73\u5747\u3002\u5f9e \u5c11\u90e8\u5206\u662f\u5169\u8005\u6301\u5e73\u3002\u9019\u88e1\u6211\u5011\u53ef\u5206\u6790 NW\u3001BPP \u548c LS \u9019\u4e09\u7a2e\u65b9\u6cd5\u5f7c\u6b64\u7684\u4e92\u88dc\uff0c\u6700\u597d\u7684 \u81f3\u53f3\u5206\u5225\u70ba\u4e0d\u540c\u65b9\u5f0f\u8a13\u7df4\u4e0b\u7684 LF-MMI \u8072\u5b78\u6a21\u578b\uff0c\u6bd4\u8f03\u7d00\u9304\u65bc\u8868 4\uff1bFCOMB \u548c HCOMB \u5c64\u7d1a\u7684\u5408\u4f75\u5341\u5206\u6709\u6548\uff0c\u4e14\u5047\u8aaa\u5c64\u7d1a\u7684\u5408\u4f75\u5728\u5927\u90e8\u5206\u7684\u60c5\u6cc1\u4e0b\uff0c\u4ecd\u52dd\u904e\u97f3\u6846\u5c64\u7d1a\u7684\u5408\u4f75\uff0c \u4f75\u7684\u6210\u6548\uff0c\u56e0\u6b64\u63a1\u7528\u7c21\u55ae\u5730\u8abf\u6574\u8a13\u7df4\u65b9\u5f0f\u9054\u6210\u591a\u6a23\u6027\u3002\u5be6\u9a57\u8a18\u9304\u65bc\u8868 5\uff0c\u5f9e\u7b2c\u4e00\u5217\u7531\u5de6 NW \u548c BPP\uff1b\u7a2e\u5b50\u6a21\u578b\u3001NW\u3001BPP \u548c LS\u3002\u5f9e\u5be6\u9a57\u7684\u7d50\u679c\u4e2d\u53ef\u770b\u51fa\u57fa\u65bc\u97f3\u6846\u5c64\u7d1a\u8207\u5047\u8aaa \u4f75\u3002\u96d6\u7136\u5408\u4f75\u7684\u589e\u76ca\u4e3b\u8981\u4f86\u81ea\u6a21\u578b\u7684\u5404\u5225\u6e96\u78ba\u8207\u591a\u6a23\u6027\uff0c\u4f46\u9019\u88e1\u4e3b\u8981\u662f\u63a2\u8a0e\u6574\u9ad4\u5b78\u7fd2\u5408 LS\u3002\u800c\u9019\u88e1\u5408\u4f75\u7684\u6a21\u578b\u662f\u8868\u4e8c\u7684\u8a13\u7df4\u7d50\u679c\u3002\u5408\u4f75\u7684\u65b9\u5f0f\u70ba\u7a2e\u5b50\u6a21\u578b\u8207 NW\uff1b\u7a2e\u5b50\u6a21\u578b\u3001 \u9019\u88e1\u63a2\u8a0e\u4e0d\u540c\u7684\u8a13\u7df4\u6a5f\u5236\u4e0b\uff0c\u6a21\u578b\u5408\u4f75\u7684\u6210\u6548\u3002\u53ef\u5206\u70ba\u97f3\u6846\u5c64\u7d1a\u7684\u5408\u4f75\u53ca\u5047\u8aaa\u5c64\u7d1a\u7684\u5408 \u8868 6 \u4e2d\u7684\u7b2c\u4e00\u5217\u5206\u5225\u70ba\u97f3\u6846\u5c64\u7d1a\u5408\u4f75\u8207\u5047\u8aaa\u5c64\u7d1a\u5408\u4f75\u3002\u7b2c\u4e8c\u6b04\u7531\u4e0a\u81f3\u4e0b\u70ba NW\u3001BPP \u548c 5.2.2 \u6a21\u578b\u5408\u4f75\u65bc\u534a\u76e3\u7763\u5f0f\u8a13\u7df4 (Model combination in semi-supervised training) 5.2.3 \u4e0d\u540c\u534a\u76e3\u7763\u5f0f\u6e96\u5247\u7684\u6a21\u578b\u5408\u4f75 (Model Combination and Semi-supervised Training) \u53c3\u8003\u6587\u737b (REFERENCES)</td></tr></table>", |
| "html": null, |
| "text": "\u672c\u8ad6\u6587\u63a2\u8a0e\u5169\u7a2e\u601d\u8def\u65bc\u534a\u76e3\u7763\u5f0f LF-MMI\u3002\u5176\u4e00\uff0c\u5229\u7528 NCE \u6b0a\u91cd\u8207\u8a5e\u5716\u6a21\u64ec\u672a\u8f49\u5beb\u8a9e\u6599 \u7684\u4e0d\u78ba\u5b9a\u6027\uff1b\u5176\u4e8c\uff0c\u63a2\u8a0e\u4e0d\u540c\u5c64\u7d1a\u7684\u5408\u4f75\uff0c\u8f03\u5feb\u7684\u97f3\u6846\u5c64\u7d1a\u5408\u4f75\u548c\u8f03\u6e96\u7684\u5047\u8aaa\u5c64\u7d1a\u5408\u4f75\u3002 \u5be6\u9a57\u7d50\u679c\u5f97\u77e5\uff0c\u5728\u7121\u9700\u4fe1\u5fc3\u904e\u6ffe\u5668\u7684\u8a9e\u6599\u6311\u9078\u4e0b\uff0c\u9019\u5169\u7a2e\u601d\u8def\u53ef\u76f4\u63a5\u61c9\u7528\u65bc\u534a\u76e3\u7763\u5f0f LF-MMI\uff0c\u4e26\u80fd\u6709\u6548\u5730\u964d\u4f4e WER \u8207\u63d0\u5347 WRR \u4e14\u76f8\u8f14\u76f8\u6210\uff0c\u6700\u7d42 WRR \u70ba 60.8%\u3002\u672a\u4f86" |
| } |
| } |
| } |
| } |
