| { | |
| "paper_id": "2018", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T07:27:39.857365Z" | |
| }, | |
| "title": "Leveraging Discriminative Training and Improved Neural Network Architecture and Optimization Method", | |
| "authors": [ | |
| { | |
| "first": "\u8d99\u5049\u6210", | |
| "middle": [ | |
| "\uf02a" | |
| ], | |
| "last": "\u3001\u5f35\u4fee\u745e", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "National Taiwan Normal University", | |
| "location": {} | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "\uf02a", | |
| "middle": [], | |
| "last": "\u3001\u7f85\u5929\u5b8f", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "National Taiwan Normal University", | |
| "location": {} | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "\uf02a", | |
| "middle": [], | |
| "last": "\u3001\u9673\u67cf\u7433 \uf02a", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "National Taiwan Normal University", | |
| "location": {} | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Wei-Cheng", | |
| "middle": [], | |
| "last": "Chao", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "National Taiwan Normal University", | |
| "location": {} | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Hsiu-Jui", | |
| "middle": [], | |
| "last": "Chang", | |
| "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": "[email protected]" | |
| }, | |
| { | |
| "first": "Berlin", | |
| "middle": [], | |
| "last": "Chen", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "National Taiwan Normal University", | |
| "location": {} | |
| }, | |
| "email": "[email protected]" | |
| }, | |
| { | |
| "first": "\u95dc\u9375\u8a5e\uff1a\u4e2d\u6587\u5927\u8a5e\u5f59\u9023\u7e8c\u8a9e\u97f3\u8fa8\u8b58\u3001\u8072\u5b78\u6a21\u578b\u3001\u9451\u5225\u5f0f\u8a13\u7df4\u3001\u77e9\u9663\u5206\u89e3\u3001\u4f86\u56de", | |
| "middle": [], | |
| "last": "\u91dd\u6cd5", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "National Taiwan Normal University", | |
| "location": {} | |
| }, | |
| "email": "" | |
| } | |
| ], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "\u6458\u8981 \u672c\u8ad6\u6587\u63a2\u8a0e\u8072\u5b78\u6a21\u578b\u4e0a\u7684\u6539\u9032\u5c0d\u65bc\u5927\u8a5e\u5f59\u9023\u7e8c\u4e2d\u6587\u8a9e\u97f3\u8fa8\u8b58\u7684\u5f71\u97ff\u3002\u5728\u57fa\u790e\u8072 \u5b78\u6a21\u578b\u7684\u8a13\u7df4\u4e0a\uff0c\u6709\u5225\u65bc\u4ee5\u5f80\u8a9e\u97f3\u8fa8\u8b58\u901a\u5e38\u4f7f\u7528\u4ea4\u4e92\u71b5(Cross Entropy)\u4f5c\u70ba\u6df1 \u5ea6\u985e\u795e\u7d93\u7db2\u8def\u76ee\u6a19\u51fd\u6578\uff0c\u6211\u5011\u4f7f\u7528 Lattice-free Maximum Mutual Information (LF-MMI)\u505a\u70ba\u5e8f\u5217\u5f0f\u9451\u5225\u8a13\u7df4\u7684\u76ee\u6a19\u51fd\u6578\u3002LF-MMI \u4f7f\u5f97\u80fd\u5920\u85c9\u7531\u5716\u5f62\u8655\u7406\u5668 (Graphical Processing Unit, GPU)\u4e0a\u5feb\u901f\u5730\u9032\u884c\u524d\u5411\u5f8c\u5411\u904b\u7b97\uff0c\u4e26\u4e14\u627e\u51fa\u6240\u6709\u53ef \u80fd\u8def\u5f91\u7684\u5f8c\u9a57\u6a5f\u7387\uff0c\u7701\u53bb\u50b3\u7d71\u9451\u5225\u5f0f\u8a13\u7df4\u524d\u9700\u8981\u63d0\u524d\u751f\u6210\u8a5e\u5716(Word Lattices) \u7684\u6b65\u9a5f\u3002\u91dd\u5c0d\u9019\u6a23\u7684\u8a13\u7df4\u65b9\u5f0f\uff0c\u985e\u795e\u7d93\u7db2\u8def\u7684\u90e8\u5206\u901a\u5e38\u4f7f\u7528\u6240\u8b02\u7684\u6642\u9593\u5ef6\u9072\u985e \u795e\u7d93\u7db2\u8def(Time-Delay Neural Network, TDNN)\u505a\u70ba\u8072\u5b78\u6a21\u578b\u53ef\u9054\u5230\u4e0d\u932f\u7684\u8fa8\u8b58 \u6548\u679c\u3002\u56e0\u6b64\uff0c\u672c\u7bc7\u8ad6\u6587\u5c07\u57fa\u65bc TDNN \u6a21\u578b\u52a0\u6df1\u985e\u795e\u7d93\u7db2\u8def\u5c64\u6578\uff0c\u4e26\u85c9\u7531\u534a\u6b63 \u4ea4\u4f4e\u79e9\u77e9\u9663\u5206\u89e3\u4f7f\u5f97\u6df1\u5c64\u985e\u795e\u7d93\u7db2\u8def\u8a13\u7df4\u904e\u7a0b\u66f4\u52a0\u7a69\u5b9a\u3002\u53e6\u4e00\u65b9\u9762\uff0c\u70ba\u4e86\u589e\u52a0 \u6a21\u578b\u7684\u4e00\u822c\u5316\u80fd\u529b(Generalization Ability)\uff0c\u6211\u5011\u4f7f\u7528\u4f86\u56de\u91dd\u6cd5(Backstitch)\u7684\u512a \u5316\u7b97\u6cd5\u3002\u5728\u4e2d\u6587\u5ee3\u64ad\u65b0\u805e\u7684\u8fa8\u8b58\u4efb\u52d9\u986f\u793a\uff0c\u4e0a\u8ff0\u5169\u7a2e\u6539\u9032\u65b9\u6cd5\u7684\u7d50\u5408\u80fd\u8b93 TDNN-LF-MMI \u7684\u6a21\u578b\u5728\u5b57\u932f\u8aa4\u7387(Character Error Rate, CER)\u6709\u76f8\u7576\u986f\u8457\u7684\u964d \u4f4e\u3002", | |
| "pdf_parse": { | |
| "paper_id": "2018", | |
| "_pdf_hash": "", | |
| "abstract": [ | |
| { | |
| "text": "\u6458\u8981 \u672c\u8ad6\u6587\u63a2\u8a0e\u8072\u5b78\u6a21\u578b\u4e0a\u7684\u6539\u9032\u5c0d\u65bc\u5927\u8a5e\u5f59\u9023\u7e8c\u4e2d\u6587\u8a9e\u97f3\u8fa8\u8b58\u7684\u5f71\u97ff\u3002\u5728\u57fa\u790e\u8072 \u5b78\u6a21\u578b\u7684\u8a13\u7df4\u4e0a\uff0c\u6709\u5225\u65bc\u4ee5\u5f80\u8a9e\u97f3\u8fa8\u8b58\u901a\u5e38\u4f7f\u7528\u4ea4\u4e92\u71b5(Cross Entropy)\u4f5c\u70ba\u6df1 \u5ea6\u985e\u795e\u7d93\u7db2\u8def\u76ee\u6a19\u51fd\u6578\uff0c\u6211\u5011\u4f7f\u7528 Lattice-free Maximum Mutual Information (LF-MMI)\u505a\u70ba\u5e8f\u5217\u5f0f\u9451\u5225\u8a13\u7df4\u7684\u76ee\u6a19\u51fd\u6578\u3002LF-MMI \u4f7f\u5f97\u80fd\u5920\u85c9\u7531\u5716\u5f62\u8655\u7406\u5668 (Graphical Processing Unit, GPU)\u4e0a\u5feb\u901f\u5730\u9032\u884c\u524d\u5411\u5f8c\u5411\u904b\u7b97\uff0c\u4e26\u4e14\u627e\u51fa\u6240\u6709\u53ef \u80fd\u8def\u5f91\u7684\u5f8c\u9a57\u6a5f\u7387\uff0c\u7701\u53bb\u50b3\u7d71\u9451\u5225\u5f0f\u8a13\u7df4\u524d\u9700\u8981\u63d0\u524d\u751f\u6210\u8a5e\u5716(Word Lattices) \u7684\u6b65\u9a5f\u3002\u91dd\u5c0d\u9019\u6a23\u7684\u8a13\u7df4\u65b9\u5f0f\uff0c\u985e\u795e\u7d93\u7db2\u8def\u7684\u90e8\u5206\u901a\u5e38\u4f7f\u7528\u6240\u8b02\u7684\u6642\u9593\u5ef6\u9072\u985e \u795e\u7d93\u7db2\u8def(Time-Delay Neural Network, TDNN)\u505a\u70ba\u8072\u5b78\u6a21\u578b\u53ef\u9054\u5230\u4e0d\u932f\u7684\u8fa8\u8b58 \u6548\u679c\u3002\u56e0\u6b64\uff0c\u672c\u7bc7\u8ad6\u6587\u5c07\u57fa\u65bc TDNN \u6a21\u578b\u52a0\u6df1\u985e\u795e\u7d93\u7db2\u8def\u5c64\u6578\uff0c\u4e26\u85c9\u7531\u534a\u6b63 \u4ea4\u4f4e\u79e9\u77e9\u9663\u5206\u89e3\u4f7f\u5f97\u6df1\u5c64\u985e\u795e\u7d93\u7db2\u8def\u8a13\u7df4\u904e\u7a0b\u66f4\u52a0\u7a69\u5b9a\u3002\u53e6\u4e00\u65b9\u9762\uff0c\u70ba\u4e86\u589e\u52a0 \u6a21\u578b\u7684\u4e00\u822c\u5316\u80fd\u529b(Generalization Ability)\uff0c\u6211\u5011\u4f7f\u7528\u4f86\u56de\u91dd\u6cd5(Backstitch)\u7684\u512a \u5316\u7b97\u6cd5\u3002\u5728\u4e2d\u6587\u5ee3\u64ad\u65b0\u805e\u7684\u8fa8\u8b58\u4efb\u52d9\u986f\u793a\uff0c\u4e0a\u8ff0\u5169\u7a2e\u6539\u9032\u65b9\u6cd5\u7684\u7d50\u5408\u80fd\u8b93 TDNN-LF-MMI \u7684\u6a21\u578b\u5728\u5b57\u932f\u8aa4\u7387(Character Error Rate, CER)\u6709\u76f8\u7576\u986f\u8457\u7684\u964d \u4f4e\u3002", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Abstract", | |
| "sec_num": null | |
| } | |
| ], | |
| "body_text": [ | |
| { | |
| "text": "\u8fd1\u5e7e\u5e74\u4f86\uff0c\u8a9e\u97f3\u8fa8\u8b58\u6280\u8853\u5df2\u6709\u4e86\u9577\u8db3\u7684\u9032\u6b65\u3002\u5176\u4e2d\uff0c\u96a8\u8457\u6df1\u5ea6\u5b78\u7fd2\u6280\u8853\u4ee5\u53ca\u96fb\u8166\u904b\u7b97\u80fd \u529b\u7684\u7a81\u7834\u6027\u767c\u5c55\uff0c\u8072\u5b78\u6a21\u578b\u5316\u6280\u8853\u5df2\u5f9e\u50b3\u7d71\u7684\u9ad8\u65af\u6df7\u5408\u6a21\u578b\u7d50\u5408\u96b1\u85cf\u5f0f\u99ac\u53ef\u592b\u6a21\u578b (Gaussian Mixture Model-Hidden Markov Model, GMM-HMM) (Rabiner, 1989) (Gales & Yang, 2008) (Bahl, Brown, de Souza & Mercer, 1986) (Vesel\u00fd, Ghoshal, Burget & Poveyet, 2013) \u3002\u8fd1\u5e74\u4f86\u70ba\u4e86\u6e1b\u5c11\u6642\u9593\u53ca\u7a7a\u9593\u8907\u96dc\u5ea6\uff0c\u6709\u5b78\u8005\u5c0d\u65bc Maximum Mutual Information (MMI)\u8a13\u7df4\uff0c\u63d0\u51fa\u4e86\u6240\u8b02\u7684 Lattice-free \u7684\u65b9\u5f0f\uff0c\u4f7f\u7522\u751f\u8a5e\u5716\u7684\u6b65\u9a5f\u80fd \u5920\u5728 GPU \u4e0a\u5b8c\u6210 (Povey et al., 2016) \uff0c\u56e0\u800c\u8b93\u9451\u5225\u5f0f\u8a13\u7df4\u5f97\u4ee5\u505a\u5230\u7aef\u5c0d\u7aef\u7684\u8a13\u7df4\u65b9\u5f0f (Hadian, Sameti, Povey & Khudanpur, 2018 )\uff0c\u56e0\u800c\u5927\u5e45\u7e2e\u6e1b\u4e86\u8072\u5b78\u6a21\u578b\u8a13\u7df4\u6240\u9700\u6642\u9593\u3002 \u50b3\u7d71 DNN-HMM \u6a21\u578b\u7528\u65bc\u8a9e\u97f3\u8fa8\u8b58\u7684\u7f3a\u9ede\u5728\u65bc\u7121\u6cd5\u5145\u5206\u5229\u7528\u8a9e\u97f3\u4fe1\u865f\u4e4b\u6642\u9593\u4f9d\u8cf4 \u6027\uff1b\u800c\u5982\u540c\u5728 (Graves, Mohamed & Hinton, 2013) ", | |
| "cite_spans": [ | |
| { | |
| "start": 130, | |
| "end": 145, | |
| "text": "(Rabiner, 1989)", | |
| "ref_id": "BIBREF12" | |
| }, | |
| { | |
| "start": 146, | |
| "end": 166, | |
| "text": "(Gales & Yang, 2008)", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 167, | |
| "end": 205, | |
| "text": "(Bahl, Brown, de Souza & Mercer, 1986)", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 206, | |
| "end": 247, | |
| "text": "(Vesel\u00fd, Ghoshal, Burget & Poveyet, 2013)", | |
| "ref_id": "BIBREF18" | |
| }, | |
| { | |
| "start": 351, | |
| "end": 371, | |
| "text": "(Povey et al., 2016)", | |
| "ref_id": "BIBREF10" | |
| }, | |
| { | |
| "start": 394, | |
| "end": 434, | |
| "text": "(Hadian, Sameti, Povey & Khudanpur, 2018", | |
| "ref_id": "BIBREF2" | |
| }, | |
| { | |
| "start": 503, | |
| "end": 535, | |
| "text": "(Graves, Mohamed & Hinton, 2013)", | |
| "ref_id": "BIBREF1" | |
| } | |
| ], | |
| "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": "\u2190 ,", | |
| "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": "\u03b8\u662f\u66f4\u65b0\u7684\u53c3\u6578\uff0cx_n \u662f\u7b2c n \u500b\u8fed\u4ee3\u7684\u6a23\u672c g(x_n,\u03b8_n)\u662f f(x,\u03b8)\u95dc\u65bc\u03b8\u7684\u5c0e\u51fd\u6578 \u2190 ,", | |
| "eq_num": "(3)" | |
| } | |
| ], | |
| "section": "\u7dd2\u8ad6 (INTRODUCTION)", | |
| "sec_num": "1." | |
| }, | |
| { | |
| "text": "\u2190 , 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 Gales, M., & Yang, S. (2008 . The application of hidden markov models in speech recognition. Foundations and Trends\u00ae in Signal Processing, 1(3), 195-304. doi: 10.1561/2000000004 ", | |
| "cite_spans": [ | |
| { | |
| "start": 25, | |
| "end": 58, | |
| "text": "de Souza, P., & Mercer, R. (1986)", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 175, | |
| "end": 186, | |
| "text": "ICASSP 1986", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 187, | |
| "end": 201, | |
| "text": ". doi: 10.1109", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 202, | |
| "end": 214, | |
| "text": "/ICASSP.1986", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 215, | |
| "end": 251, | |
| "text": ".1169179 Gales, M., & Yang, S. (2008", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 317, | |
| "end": 401, | |
| "text": "Foundations and Trends\u00ae in Signal Processing, 1(3), 195-304. doi: 10.1561/2000000004", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "\u7dd2\u8ad6 (INTRODUCTION)", | |
| "sec_num": "1." | |
| } | |
| ], | |
| "back_matter": [], | |
| "bib_entries": { | |
| "BIBREF0": { | |
| "ref_id": "b0", | |
| "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 '06", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "369--376", | |
| "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 '06, 369-376. doi: 10.1145/1143844.1143891", | |
| "links": null | |
| }, | |
| "BIBREF1": { | |
| "ref_id": "b1", | |
| "title": "Speech recognition with deep recurrent neural networks", | |
| "authors": [ | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Graves", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Mohamed", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [ | |
| "E" | |
| ], | |
| "last": "Hinton", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2013, | |
| "venue": "Proceedings of ICASSP 2013", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": { | |
| "DOI": [ | |
| "10.1109/ICASSP.2013.6638947" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Graves, A., Mohamed, A.-r., & Hinton, G. E. (2013). Speech recognition with deep recurrent neural networks. In Proceedings of ICASSP 2013. doi: 10.1109/ICASSP.2013.6638947", | |
| "links": null | |
| }, | |
| "BIBREF2": { | |
| "ref_id": "b2", | |
| "title": "End-to-end speech recognition using lattice-free MMI", | |
| "authors": [ | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Hadian", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Sameti", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Povey", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Khudanpur", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2018, | |
| "venue": "Proceedings of Interspeech", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": { | |
| "DOI": [ | |
| "10.21437/Interspeech.2018-1423" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Hadian, H., Sameti, H., Povey, D., & Khudanpur, S. (2018). End-to-end speech recognition using lattice-free MMI. In Proceedings of Interspeech 2018. doi: 10.21437/Interspeech.2018-1423", | |
| "links": null | |
| }, | |
| "BIBREF3": { | |
| "ref_id": "b3", | |
| "title": "Deep residual learning for image recognition", | |
| "authors": [ | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "He", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "X", | |
| "middle": [], | |
| "last": "Zhang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Ren", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Sun", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2016, | |
| "venue": "Proceedings of the IEEE conference on computer vision and pattern recognition", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": { | |
| "DOI": [ | |
| "10.1109/CVPR.2016.90" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image recognition. (2016). In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR 2016). doi: 10.1109/CVPR.2016.90", | |
| "links": null | |
| }, | |
| "BIBREF4": { | |
| "ref_id": "b4", | |
| "title": "Deep neural networks for acoustic modeling in speech recognition: the shared views of four research groups", | |
| "authors": [ | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Hinton", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "L", | |
| "middle": [], | |
| "last": "Deng", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Yu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [ | |
| "E" | |
| ], | |
| "last": "Dahl", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Mohamed", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "N", | |
| "middle": [], | |
| "last": "Jaitly", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "B", | |
| "middle": [], | |
| "last": "\u2026kingsbury", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2012, | |
| "venue": "IEEE Signal processing magazine", | |
| "volume": "29", | |
| "issue": "6", | |
| "pages": "82--97", | |
| "other_ids": { | |
| "DOI": [ | |
| "10.1109/MSP.2012.2205597" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Hinton, G., Deng, L., Yu, D., Dahl, G. E., Mohamed, A.-r., Jaitly, N., \u2026Kingsbury, B. (2012). Deep neural networks for acoustic modeling in speech recognition: the shared views of four research groups. IEEE Signal processing magazine, 29(6), 82-97. doi: 10.1109/MSP.2012.2205597", | |
| "links": null | |
| }, | |
| "BIBREF5": { | |
| "ref_id": "b5", | |
| "title": "On the difficulty of training recurrent neural networks", | |
| "authors": [ | |
| { | |
| "first": "R", | |
| "middle": [], | |
| "last": "Pascanu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Mikolov", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Bengio", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2013, | |
| "venue": "Proceedings of ICML 2013", | |
| "volume": "28", | |
| "issue": "", | |
| "pages": "1310--1318", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Pascanu, R., Mikolov, T., & Bengio, Y. (2013). On the difficulty of training recurrent neural networks. In Proceedings of ICML 2013, 28, 1310-1318.", | |
| "links": null | |
| }, | |
| "BIBREF7": { | |
| "ref_id": "b7", | |
| "title": "Semi-orthogonal low-rank matrix factorization for deep neural networks", | |
| "authors": [], | |
| "year": 2018, | |
| "venue": "Proceedings of Interspeech", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": { | |
| "DOI": [ | |
| "10.21437/Interspeech.2018-1417" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Semi-orthogonal low-rank matrix factorization for deep neural networks. In Proceedings of Interspeech 2018. doi: 10.21437/Interspeech.2018-1417", | |
| "links": null | |
| }, | |
| "BIBREF8": { | |
| "ref_id": "b8", | |
| "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 | |
| }, | |
| "BIBREF9": { | |
| "ref_id": "b9", | |
| "title": "A time-restricted self-attention layer for ASR", | |
| "authors": [ | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Povey", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Hadian", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "P", | |
| "middle": [], | |
| "last": "Ghahremani", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Li", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Khudanpur", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2018, | |
| "venue": "Proceedings of ICASSP 2018", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": { | |
| "DOI": [ | |
| "10.1109/ICASSP.2018.8462497" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Povey, D., Hadian, H., Ghahremani, P., Li, K., & Khudanpur, S. (2018) A time-restricted self-attention layer for ASR. In Proceedings of ICASSP 2018. doi: 10.1109/ICASSP.2018.8462497", | |
| "links": null | |
| }, | |
| "BIBREF10": { | |
| "ref_id": "b10", | |
| "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 | |
| }, | |
| "BIBREF11": { | |
| "ref_id": "b11", | |
| "title": "Parallel training of DNNs with natural gradient and parameter averaging", | |
| "authors": [ | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Povey", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "X", | |
| "middle": [], | |
| "last": "Zhang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Khudanpur", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2014, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Povey, D., Zhang, X., & Khudanpur, S. (2014). Parallel training of DNNs with natural gradient and parameter averaging. Retrieved from https://arxiv.org/abs/1410.7455", | |
| "links": null | |
| }, | |
| "BIBREF12": { | |
| "ref_id": "b12", | |
| "title": "A tutorial on hidden markov models and selected applications in speech recognition", | |
| "authors": [ | |
| { | |
| "first": "L", | |
| "middle": [ | |
| "R" | |
| ], | |
| "last": "Rabiner", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1989, | |
| "venue": "Proceedings of the IEEE", | |
| "volume": "77", | |
| "issue": "", | |
| "pages": "257--286", | |
| "other_ids": { | |
| "DOI": [ | |
| "10.1109/5.18626" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Rabiner, L. R. (1989). A tutorial on hidden markov models and selected applications in speech recognition. Proceedings of the IEEE, 77(2), 257-286. doi: 10.1109/5.18626", | |
| "links": null | |
| }, | |
| "BIBREF13": { | |
| "ref_id": "b13", | |
| "title": "An overview of gradient descent optimization algorithms", | |
| "authors": [ | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Ruder", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2016, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Ruder, S. (2016). An overview of gradient descent optimization algorithms. Retrieved from https://arxiv.org/abs/1609.04747", | |
| "links": null | |
| }, | |
| "BIBREF14": { | |
| "ref_id": "b14", | |
| "title": "Long Short-Term Memory Based Recurrent Neural Network Architectures for Large Vocabulary Speech Recognition", | |
| "authors": [ | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Sak", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Senior", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "F", | |
| "middle": [], | |
| "last": "Beaufays", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2014, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": { | |
| "arXiv": [ | |
| "arXiv:1402.1128" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Sak, H., Senior, A., & Beaufays, F. (2014). Long Short-Term Memory Based Recurrent Neural Network Architectures for Large Vocabulary Speech Recognition. Retrieved from arXiv:1402.1128", | |
| "links": null | |
| }, | |
| "BIBREF15": { | |
| "ref_id": "b15", | |
| "title": "Very deep convolutional networks for large-scale image recognition", | |
| "authors": [ | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Simonyan", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Zisserman", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2014, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Simonyan, K. & Zisserman, A. (2014). Very deep convolutional networks for large-scale image recognition. Retrieved from https://arxiv.org/abs/1409.1556", | |
| "links": null | |
| }, | |
| "BIBREF17": { | |
| "ref_id": "b17", | |
| "title": "SRILM-an extensible language modeling toolkit", | |
| "authors": [ | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Stolcke", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2002, | |
| "venue": "Proceedings of ICSLP 2002", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "901--904", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Stolcke, A. (2002). SRILM-an extensible language modeling toolkit. In Proceedings of ICSLP 2002, 901-904.", | |
| "links": null | |
| }, | |
| "BIBREF18": { | |
| "ref_id": "b18", | |
| "title": "Sequence-discriminative training of deep neural networks", | |
| "authors": [ | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Vesel\u00fd", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Ghoshal", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "L", | |
| "middle": [], | |
| "last": "Burget", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Poveyet", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2013, | |
| "venue": "Proceedings of Interspeech 2013", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "2345--2349", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Vesel\u00fd, K., Ghoshal, A., Burget, L., & Poveyet, D. (2013). Sequence-discriminative training of deep neural networks. In Proceedings of Interspeech 2013, 2345-2349.", | |
| "links": null | |
| }, | |
| "BIBREF19": { | |
| "ref_id": "b19", | |
| "title": "Phoneme recognition using time-delay neural networks", | |
| "authors": [ | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Waibel", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Hanazawa", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Hinton", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Shikano", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "K", | |
| "middle": [ | |
| "J" | |
| ], | |
| "last": "Lang", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1989, | |
| "venue": "IEEE Transactions on Acoustics, Speech, and Signal Processing", | |
| "volume": "37", | |
| "issue": "3", | |
| "pages": "328--339", | |
| "other_ids": { | |
| "DOI": [ | |
| "10.1109/29.21701" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Waibel, A., Hanazawa, T., Hinton, G., Shikano, K., & Lang, K. J. (1989). Phoneme recognition using time-delay neural networks. IEEE Transactions on Acoustics, Speech, and Signal Processing, 37(3), 328-339. doi: 10.1109/29.21701", | |
| "links": null | |
| }, | |
| "BIBREF20": { | |
| "ref_id": "b20", | |
| "title": "MATBN: A Mandarin Chinese broadcast news corpus. International journal of computational linguistics & Chinese language processing", | |
| "authors": [ | |
| { | |
| "first": "H.-M", | |
| "middle": [], | |
| "last": "Wang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "B", | |
| "middle": [], | |
| "last": "Chen", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J.-W", | |
| "middle": [], | |
| "last": "Kuo", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S.-S", | |
| "middle": [], | |
| "last": "Cheng", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2005, | |
| "venue": "Special Issue on Annotated Speech Corpora", | |
| "volume": "10", | |
| "issue": "2", | |
| "pages": "219--236", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Wang, H.-M., Chen, B., Kuo, J.-W., & Cheng, S.-S. (2005). MATBN: A Mandarin Chinese broadcast news corpus. International journal of computational linguistics & Chinese language processing, Special Issue on Annotated Speech Corpora, 10(2), 219-236.", | |
| "links": null | |
| }, | |
| "BIBREF21": { | |
| "ref_id": "b21", | |
| "title": "Backstitch: counteracting finite-sample bias via negative steps", | |
| "authors": [ | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Wang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "V", | |
| "middle": [], | |
| "last": "Peddinti", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Xu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "X", | |
| "middle": [], | |
| "last": "Zhang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Povey", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Khudanpur", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2017, | |
| "venue": "Proceedings of Interspeech", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": { | |
| "DOI": [ | |
| "10.21437/Interspeech.2017-1323" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Wang, Y., Peddinti, V., Xu, H., Zhang, X., Povey, D., & Khudanpur, S. (2017). Backstitch: counteracting finite-sample bias via negative steps. In Proceedings of Interspeech 2017. doi: 10.21437/Interspeech.2017-1323", | |
| "links": null | |
| } | |
| }, | |
| "ref_entries": { | |
| "FIGREF0": { | |
| "num": null, | |
| "type_str": "figure", | |
| "uris": null, | |
| "text": "., & Rich, C. (2014). Do deep nets really need to be deep? In Proceedings ofNIPS 2014NIPS , 2, 2654NIPS -2662" | |
| }, | |
| "TABREF0": { | |
| "type_str": "table", | |
| "num": null, | |
| "text": "\uff0c\u8f49\u8b8a\u6210\u4ee5\u4f7f\u7528\u4ea4\u4e92\u71b5(Cross Entropy)\u4f5c\u70ba\u640d\u5931\u51fd\u6578\u7684\u6df1\u5ea6\u985e\u795e\u7d93\u7db2\u8def\u7d50\u5408\u96b1 \u85cf\u5f0f\u99ac\u53ef\u592b\u6a21\u578b(Deep Neural Network-Hidden Markov Model, DNN-HMM) (Hinton et al., 2012)\u3002DNN-HMM \u5c07\u4ee5\u5f80\u7528 GMM \u8a08\u7b97\u7684\u751f\u6210\u6a5f\u7387\u900f\u904e DNN \u7684\u8f38\u51fa\u5c64\u6240\u4ee3\u8868\u7684\u4e8b\u5f8c\u6a5f \u7387\u4f86\u8fd1\u4f3c\uff0c\u8f38\u5165\u7279\u5fb5\u4f7f\u7528\u7576\u524d\u5e40\u9084\u6709\u76f8\u9130\u7684\u5e40\uff0c\u8f38\u51fa\u5247\u548c GMM-HMM \u5e38\u7528\u7684 Triphone \u5171\u4eab\u72c0\u614b\u76f8\u540c\uff0c\u4ee5\u5f97\u5230\u66f4\u4f4e\u7684\u8a5e\u932f\u8aa4\u7387(Word Error Rate, WER)\u6216\u5b57\u932f\u8aa4\u7387(Character Error Rate, CER)\u3002\u53e6\u4e00\u65b9\u9762\uff0c\u9032\u4e00\u6b65\u900f\u904e\u9451\u5225\u5f0f\u8a13\u7df4\u4f30\u6e2c\u7684\u8072\u5b78\u6a21\u578b\u5728\u8a9e\u97f3\u8fa8\u8b58\u7684\u8868\u73fe \u4e0a\u5f80\u5f80\u6bd4\u50c5\u4ee5\u4ea4\u4e92\u71b5\u505a\u70ba\u6df1\u5ea6\u985e\u795e\u7d93\u7db2\u8def\u640d\u5931\u51fd\u6578\u7684\u8a13\u7df4\u65b9\u5f0f\u4f86\u7684\u597d\u3002\u4f46\u7531\u65bc\u50b3\u7d71\u4e0a\u9032 \u884c\u9451\u5225\u5f0f\u8a13\u7df4\u9700\u8981\u4f7f\u7528\u5148\u9032\u884c\u4ea4\u4e92\u71b5\u8a13\u7df4\u7684\u8072\u5b78\u6a21\u578b\u4f86\u7522\u751f\u8a5e\u5716(Word Lattices)\uff0c\u624d\u80fd\u518d \u9032\u884c\u4e0b\u4e00\u6b65\u8072\u5b78\u6a21\u578b\u9451\u5225\u5f0f\u8a13\u7df4", | |
| "html": null, | |
| "content": "<table/>" | |
| }, | |
| "TABREF1": { | |
| "type_str": "table", | |
| "num": null, | |
| "text": "Backstitch \u662f\u4fee\u6539 SGD \u589e\u9032\u5c0d\u6c92\u6709\u770b\u904e\u8cc7\u6599\u7684\u6548\u80fd(Wang et al., 2017)\uff1b\u9019\u500b\u65b9\u6cd5\u5206\u6210\u5169 \u500b\u6b65\u9a5f\uff0c\u5148\u7528\u4e00\u500b\u8f03\u5c0f\u7684\u8ca0\u5b78\u7fd2\u7387\uff0c\u518d\u8ddf\u8457\u4e00\u500b\u8f03\u5927\u7684\u5b78\u7fd2\u7387\uff0c\u548c\u5c0d\u6297\u8a13\u7df4\u7684\u6982\u5ff5\u5f88\u50cf\uff0c \u53ef\u4ee5\u6d88\u9664\u6709\u9650\u6578\u64da\u96c6\u7684\u504f\u898b\uff0c\u7528\u540c\u6a23\u7684 Minibatch \u4f46\u91cd\u65b0\u8a08\u7b97\u68af\u5ea6\uff0c\u50b3\u7d71\u7684 SGD \u55ae\u4e00\u7684 \u8fed\u4ee3\u5982\u4e0b\u5f0f:", | |
| "html": null, | |
| "content": "<table><tr><td>\u7d50\u5408\u9451\u5225\u5f0f\u8a13\u7df4\u8072\u5b78\u6a21\u578b\u4e4b\u985e\u795e\u7d93\u7db2\u8def\u67b6\u69cb\u53ca\u512a\u5316\u65b9\u6cd5\u7684\u6539\u9032</td><td>\u8d99\u5049\u6210 \u7b49 39</td></tr><tr><td>2. \u8072\u5b78\u6a21\u578b (ACOUSTIC-MOLEL)</td><td/></tr><tr><td colspan=\"2\">2.1 \u57fa \u672c \u8072 \u5b78 \u6a21 \u578b -\u6642 \u9593 \u5ef6 \u985e \u795e \u7d93 \u7db2 \u8def (Time-Delay Neural Network,</td></tr><tr><td>TDNN)</td><td/></tr><tr><td colspan=\"2\">TDNN \u5728 1989 \u5e74\u88ab\u63d0\u51fa(Waibel et al., 1989)\uff0c\u6700\u521d\u7528\u65bc\u97f3\u7d20\u8fa8\u8b58\uff1b\u57fa\u65bc TDNN \u6240\u7522\u751f\u7684 \u6240\u63d0\u5230\uff0c\u57fa\u65bc\u905e\u8ff4\u985e\u795e\u7d93(Recurrent Neural \u6a21\u578b\u67b6\u69cb\uff0c\u80fd\u9069\u7528\u65bc\u8655\u7406\u8a9e\u97f3\u6240\u64c1\u6709\u7279\u5fb5\u5411\u91cf\u5e8f\u5217\u4e4b\u6642\u9593\u9577\u5ea6\u4e0d\u4e00\u81f4\u7684\u7279\u6027\u3002TDNN \u5c0d Network, RNN)\u80fd\u5c0d\u65bc\u5e8f\u5217\u6027\u8cc7\u6599\u80fd\u5920\u6709\u597d\u7684\u5efa\u6a21\u6548\u679c\u7684\u60f3\u6cd5\u6240\u767c\u5c55\u7684 RNN-HMM \u8072\u5b78 \u6bcf\u4e00\u500b\u96b1\u85cf\u5c64\u7684\u8f38\u51fa\u5728\u6642\u9593\u4e0a\u9032\u884c\u64f4\u5c55\uff0c\u5373\u6bcf\u500b\u96b1\u85cf\u5c64\u6536\u5230\u7684\u8f38\u5165\u6703\u6709\u524d\u4e00\u5c64\u5728\u4e0d\u540c\u6642 \u6a21\u578b\u5176\u8fa8\u8b58\u6548\u679c\u537b\u662f\u4e0d\u5982 DNN-HMM \u6a21\u578b\u4f86\u7684\u597d\uff0c\u56e0\u6b64\u4ee5\u9577\u77ed\u671f\u8a18\u61b6(Long Short-Term \u523b\u7684\u8f38\u51fa\u3002\u8a9e\u97f3\u5728\u8003\u616e\u4e0a\u4e0b\u6587\u9577\u6642\u9593\u76f8\u95dc\u6027\u5f88\u91cd\u8981\uff0cTDNN \u7684\u512a\u9ede\u5728\u53ef\u4ee5\u6bd4\u50b3\u7d71 DNN Memory, LSTM)\u53d6\u4ee3\u7c21\u55ae RNN \u6240\u5f62\u6210\u7684\u8072\u5b78\u6a21\u578b(LSTM-HMM) (Sak, Senior & Beaufays, 2014)\uff0c\u89e3\u6c7a\u4e86 RNN-HMM \u68af\u5ea6\u6d88\u5931\u7684\u554f\u984c\uff0c\u5728\u8a9e\u97f3\u8fa8\u8b58\u4e0a\u80fd\u5920\u9054\u5230\u6bd4 DNN-HMM \u597d\u7684 \u770b\u66f4\u9577\u7684\u6642\u9593\uff0c\u800c\u4e14\u901f\u5ea6\u4e0d\u6703\u6bd4 DNN \u5728\u8a13\u7df4\u548c\u8fa8\u8b58(\u89e3\u78bc)\u6642\u4f86\u7684\u6162\u3002</td></tr><tr><td colspan=\"2\">\u6548\u679c\u3002\u4f46\u5728\u5be6\u52d9\u4e0a\uff0c\u9019\u6a23\u7684\u8072\u5b78\u6a21\u578b\u5f88\u96e3\u50cf DNN-HMM \u4e00\u6a23\u5e73\u884c\u5316\u8a13\u7df4(Pascanu, Mikolov</td></tr><tr><td colspan=\"2\">& Bengio, 2013)\uff0c\u4ee5\u81f4\u65bc\u6a21\u578b\u8a13\u7df4\u6642\u9593\u7684\u589e\u52a0\u3002\u53e6\u4e00\u65b9\u9762\uff0c\u4e5f\u7531\u65bc\u5176\u6a21\u578b\u67b6\u69cb\u8f03\u70ba\u8907\u96dc</td></tr><tr><td colspan=\"2\">\u4f7f\u5f97\u904b\u7b97\u91cf\u8f03\u5927\uff0c\u8f03\u4e0d\u9069\u5408\u9700\u5373\u6642\u53cd\u61c9\u7684\u8a9e\u97f3\u8fa8\u8b58\u4efb\u52d9\uff1b\u76f8\u5c0d\u4f86\u8aaa\u6642\u9593\u5ef6\u9072\u985e\u795e\u7d93\u7db2\u8def</td></tr><tr><td colspan=\"2\">(Time-Delay Neural Network, TDNN) (Waibel, Hanazawa, Hinton, Shikano & Lang, 1989)\u53ef</td></tr><tr><td colspan=\"2\">\u4ee5\u5305\u542b\u6b77\u53f2\u548c\u672a\u4f86\u8f38\u51fa\u3001\u5c0d\u9577\u6642\u9593\u4f9d\u8cf4\u6027\u7684\u8a9e\u97f3\u8a0a\u865f\u5efa\u6a21\uff0c\u4f7f TDNN-HMM \u8207\u50b3\u7d71</td></tr><tr><td colspan=\"2\">DNN-HMM \u8a13\u7df4\u6548\u7387\u4e5f\u76f8\u4eff\uff0c\u56e0\u6b64\u5728\u4f7f\u7528 LF-MMI \u9032\u884c\u9451\u5225\u5f0f\u8a13\u7df4\u6642\uff0c\u8072\u5b78\u6a21\u578b\u7684\u985e\u795e \u5716 1</td></tr><tr><td>\u7d93\u7db2\u8def\u90e8\u5206\u901a\u5e38\u662f\u4f7f\u7528 TDNN\u3002</td><td/></tr><tr><td colspan=\"2\">\u5f9e\u7d93\u9a57\u4e0a\u4f86\u770b\uff0c\u985e\u795e\u7d93\u7db2\u8def\u7684\u6df1\u5ea6\u5c0d\u6a21\u578b\u7684\u6027\u80fd\u975e\u5e38\u91cd\u8981(Ba & Rich, 2014)\uff0c\u589e\u52a0\u5c64 \u503c\u66f4\u65b0\u901f\u7387\u7684\u5feb\u6162\uff0c\u6108\u5927\u7684\u5b78\u7fd2\u7387\u6703\u66f4\u5feb\u9054\u5230\u534a\u6b63\u4ea4\u7684\u7d50\u679c\uff0c\u4f46\u662f\u8a2d\u7f6e\u592a\u5927\u6703\u8b8a\u5f97\u5f88\u4e0d \u6578\u4e4b\u5f8c\u80fd\u6709\u66f4\u52a0\u8907\u96dc\u7684\u7279\u5fb5\u64f7\u53d6\u80fd\u529b\u3002\u5c0d\u65bc TDNN \u800c\u8a00\uff0c\u589e\u52a0\u5c64\u6578\u53ef\u4ee5\u8aaa\u662f\u63d0\u53d6\u66f4\u9577\u6642 \u7a69\u5b9a\uff0c\u5728\u63a5\u8fd1\u534a\u6b63\u4ea4\u77e9\u9663\u7684\u6642\u5019 0.125 \u7684\u8a2d\u7f6e\u662f\u6700\u597d\u7684\uff0c\u6578\u5b78\u4e0a\u53ef\u4ee5\u9054\u5230\u5e73\u65b9\u6536\u6582\u3002\u4ee4X\u662f \u9593\u7684\u7279\u5fb5\uff1b\u6211\u5011\u5e0c\u671b\u52a0\u6df1 TDNN \u7684\u7db2\u8def\u5c64\u6578\u4f86\u9054\u5230\u66f4\u597d\u7684\u7d50\u679c\uff0c\u4f46\u4ee5\u5f80\u7684\u5be6\u9a57\u767c\u73fe\u6df1\u5ea6 \u6bcf\u6b21M\u66f4\u65b0\u7684\u503c\uff0c\u6240\u4ee5\u6211\u5011\u505a\u4e00\u6b21\u66f4\u65b0M \u2190 M X\uff0c\u6211\u5011\u5e0c\u671btr MX 0\u4ee5\u9054\u5230\u6b63\u4ea4\u6548 \u7684\u7db2\u8def\u5e38\u6709\u9000\u5316\u554f\u984c\uff0c\u985e\u795e\u7d93\u7db2\u8def\u7684\u6df1\u5ea6\u4e4b\u589e\u52a0\u6e96\u78ba\u7387\u53cd\u800c\u6703\u4e0b\u964d\u3002\u56e0\u6b64\u672c\u7bc7\u8ad6\u6587\u5c07\u6bd4 \u8f03\u4e26\u7d50\u5408\u7576\u524d\u5148\u9032\u7684\u8072\u5b78\u6a21\u578b\u8a13\u7df4\u65b9\u6cd5\uff0c\u4f8b\u5982(Povey et al., 2018)\u5c0d\u7db2\u8def\u7684\u77e9\u9663\u5206\u89e3\u8a13\u7df4 \u679c\uff0c\u4e0b\u5f0f\u70ba\u66f4\u65b0\u516c\u5f0f:</td></tr><tr><td colspan=\"2\">\u53ef\u4ee5\u4f7f\u7db2\u8def\u8a13\u7df4\u66f4\u7a69\u5b9a\uff0c\u4ee5\u671f\u9054\u5230\u66f4\u4f73\u7684\u8a9e\u97f3\u8fa8\u8b58\u8868\u73fe\u3002\u53e6\u4e00\u65b9\u9762\uff0c\u68af\u5ea6\u4e0b\u964d\u662f\u57f7\u884c\u512a \u5316\u7684\u6700\u6d41\u884c\u7684\u7b97\u6cd5\u4e4b\u4e00\uff0c\u4e5f\u662f\u8fc4\u4eca\u70ba\u6b62\u512a\u5316\u985e\u795e\u7d93\u7db2\u7d61\u7684\u6700\u5e38\u7528\u65b9\u6cd5\u3002\u800c\u5e38\u898b\u7684\u512a\u5316\u7b97 \u2190 (1)</td></tr><tr><td colspan=\"2\">\u6cd5\u6709\u96a8\u6a5f\u68af\u5ea6\u4e0b\u964d\u6cd5(Stochastic Gradient Descent, SGD)\u3001RMSprop\u3001Adam\u3001Adagrad\u3001 \u03b1\u662f\u4e00\u500b\u7e2e\u653e\u7684\u53c3\u6578\uff0cI \u662f\u55ae\u4f4d\u77e9\u9663\u4e0d\u8003\u616e\u5e38\u6578\u9805\uff0c\u6211\u5011\u8981\u4f7ftr MM P I 0\uff0c Adadelta (Ruder, 2016)\u7b49\u6f14\u7b97\u6cd5\uff1b\u5176\u4e2d\uff0cSGD \u7b97\u6cd5\u5728\u8a9e\u97f3\u8fa8\u8b58\u4efb\u52d9\u4e0a\u6700\u88ab\u5ee3\u70ba\u4f7f\u7528\u3002\u800c \u56e0\u70baMM P\uff0c\u6240\u4ee5tr P P 0\u79fb\u9805\u4e4b\u5f8c\u03b1 \uff0c\u56e0\u70ba P \u662f\u5c0d\u7a31\u77e9\u9663\u6240\u4ee5 \u672c\u8ad6\u6587\u5247\u63a1\u7528\u4f86\u56de\u91dd\u6cd5(Backstitch) (Wang et al., 2017)\u505a\u70ba\u6a21\u578b\u512a\u5316\u7684\u6f14\u7b97\u6cd5\uff1b\u5b83\u662f\u4e00\u7a2e P PP \uff0c\u70ba\u4e86\u8a08\u7b97\u4e0a\u8f03\u5feb\u6703\u4f7f\u7528\u03b1 \u3002\u5716 1 \u662f TDNN+NF(Networks Factorized) \u57fa\u65bc SGD \u4e0a\u7684\u6539\u9032\uff0c\u5e0c\u671b\u80fd\u5920\u85c9\u7531\u5169\u6b65\u9a5f\u7684\u66f4\u65b0 Minibatch\uff0c\u4ee5\u9054\u5230\u66f4\u597d\u7684\u6548\u679c\u3002 \u5167\u90e8\u67b6\u69cb\uff0c1536 \u7dad\u7684\u96b1\u85cf\u5c64\u7d93\u77e9\u9663\u5206\u89e3\u5f8c\u8b8a\u6210 1536*160*1536\uff0cSMAT \u662f\u8981\u505a\u6b63\u4ea4\u9650\u5236</td></tr><tr><td colspan=\"2\">\u7e3d\u5408\u4ee5\u4e0a\u6240\u8ff0\uff0c\u6211\u5011\u8a8d\u70ba\u52a0\u5165\u5c0d\u7db2\u8def\u7684\u77e9\u9663\u5206\u89e3\u4f86\u53ef\u9806\u5229\u8a13\u7df4\u66f4\u6df1\u5c64\u7684\u985e\u795e\u7d93\u7db2\u8def \u7684\u77e9\u9663\uff0c\u5f8c\u9762\u518d\u63a5\u4e0a\u7dda\u6027\u6574\u6d41\u51fd\u6578(ReLU)\u548c\u6279\u6b21\u6a19\u6e96\u5316(Batch Normalization)\u3002</td></tr><tr><td colspan=\"2\">\u6a21\u578b\uff1b\u540c\u6642\uff0c\u4f7f\u7528 Backstitch \u4ea6\u53ef\u63d0\u5347\u6a21\u578b\u6cdb\u5316\u6027\uff0c\u6700\u7d42\u80fd\u4f7f\u8fa8\u8b58\u7d50\u679c\u66f4\u52a0\u9032\u6b65\u3002\u56e0\u6b64\uff0c</td></tr><tr><td colspan=\"2\">\u672c\u8ad6\u6587\u5c07\u5206\u5225\u6bd4\u8f03\u4f7f\u7528 TDNN-LF-MMI\uff0cTDNN-LF-MMI \u52a0\u5165\u534a\u6b63\u4ea4\u4f4e\u79e9\u77e9\u9663\u5206\u89e3\uff0c</td></tr><tr><td colspan=\"2\">TDNN-LF-MMI \u52a0\u5165\u534a\u6b63\u4ea4\u4f4e\u79e9\u77e9\u9663\u5206\u89e3\u53ca\u4f86\u56de\u91dd\u6cd5\u512a\u5316\u7b97\u6cd5\u7684\u8fa8\u8b58\u6548\u679c\uff0c\u6700\u7d42\u5728</td></tr><tr><td colspan=\"2\">TDNN-LF-MMI \u52a0\u5165\u534a\u6b63\u4ea4\u4f4e\u79e9\u77e9\u9663\u5206\u89e3\u53ca\u4f86\u56de\u91dd\u6cd5\u512a\u5316\u7b97\u6cd5\u9054\u5230\u8f03\u4f73\u7684\u4e2d\u6587\u5ee3\u64ad\u65b0\u805e</td></tr><tr><td>\u8a9e\u97f3\u8fa8\u8b58\u7684 CER \u8868\u73fe\u3002</td><td/></tr></table>" | |
| }, | |
| "TABREF2": { | |
| "type_str": "table", | |
| "num": null, | |
| "text": "Normalization \u5408\u8d77\u4f86\u7a31\u70ba TDNNF \u5c64\uff0c\u5c64\u6578\u53ef\u6bd4\u4ee5\u524d\u7684 TDNN \u7db2\u8def(9 \u5c64\u4ee5\u5167)\u90fd\u9084\u6df1\u3002\u5982 \u5716 2 \u6240\u793a\uff0c\u6700\u4e0b\u5c64\u70ba\u96a8\u8457\u6642\u9593\u8f38\u5165\u4e4b\u7279\u5fb5\uff0c\u6700\u4e0a\u5c64\u70ba\u591a\u4efb\u52d9\u7684\u8f38\u51fa\uff0cLF-MMI \u7684\u76ee\u6a19\u51fd \u6578(Povey et al., 2016)\u5c0d\u61c9\u7531\u6c7a\u7b56\u6a39\u5b9a\u7fa9\u7684 Senone \u51fd\u6578\uff0cCross Entropy Regularization \u70ba \u8f14\u52a9\u6b63\u898f\u5316\u8a13\u7df4\u8f38\u51fa\uff0c\u4e2d\u9593\u7684 TDNNF \u5c64\u6709\u6377\u5f91\u9023\u7d50(Skip Connections) \u3002", | |
| "html": null, | |
| "content": "<table><tr><td>44</td><td colspan=\"3\">\u7d50\u5408\u9451\u5225\u5f0f\u8a13\u7df4\u8072\u5b78\u6a21\u578b\u4e4b\u985e\u795e\u7d93\u7db2\u8def\u67b6\u69cb\u53ca\u512a\u5316\u65b9\u6cd5\u7684\u6539\u9032 \u7d50\u5408\u9451\u5225\u5f0f\u8a13\u7df4\u8072\u5b78\u6a21\u578b\u4e4b\u985e\u795e\u7d93\u7db2\u8def\u67b6\u69cb\u53ca\u512a\u5316\u65b9\u6cd5\u7684\u6539\u9032</td><td>\u8d99\u5049\u6210 \u7b49 41 \u8d99\u5049\u6210 \u7b49 43 \u8d99\u5049\u6210 \u7b49</td></tr><tr><td colspan=\"5\">Batch Normalization \u4e5f\u9054\u5230\u5f88\u597d\u7684\u6548\u679c\uff0c\u5728\u8fa8\u8b58\u7387\u548c\u89e3\u78bc\u901f\u5ea6\u90fd\u6709\u5f88\u5927\u63d0\u5347\u3002 2001 \u5e74\u81f3 2002 \u5e74\u4e2d\u592e\u901a\u8a0a\u793e(Central News Agency, CNA)\u7684\u6587\u5b57\u65b0\u805e\u8a9e\u6599\uff0c\u5167\u542b\u4e00\u5104\u4e94 \u5343\u842c\u500b\u4e2d\u6587\u5b57\uff0c\u7d93\u65b7\u8a5e\u5f8c\u7d04\u6709\u516b\u5343\u842c\u500b\u8a5e(\u672c\u8ad6\u6587\u4f7f\u8a5e\u5178\u7d04\u4e03\u842c\u4e8c\u5343\u8a5e)\u3002\u672c\u8ad6\u6587\u662f\u4f7f\u7528 \u8868 3. \u6539\u9032\u6a21\u578b\u5728\u6e2c\u8a66\u96c6\u4e00\u7684\u5be6\u9a57\u7d50\u679c \u5f9e\u8a13\u7df4\u904e\u7a0b\u7684\u6e96\u78ba\u7387\u4f86\u770b\uff0c\u6709\u6c92\u6709\u4f7f\u7528 Backstitch \u770b\u4e0d\u51fa\u592a\u5927\u5dee\u7570\uff0c\u4f46\u662f\u53cd\u61c9\u5728\u5b57 [Table 3. Experiment results for test sets] \u7684\u932f\u8aa4\u7387\u4e0a\u6709\u9032\u6b65\u3002\u6709\u505a\u77e9\u9663\u5206\u89e3\u7684\u6a21\u578b\u56e0\u70ba\u6709\u6b63\u4ea4\u9650\u5236\u7684\u66f4\u65b0\uff0c\u5728\u8fed\u4ee3 160 \u6b21\u5f8c\u6e96\u78ba</td></tr><tr><td colspan=\"5\">3.3 \u7d50\u5408\u534a\u6b63\u4ea4\u9650\u5236\u548c\u4f86\u56de\u91dd\u6cd5\u8a13\u7df4 (Combine Training) SRI Language Modeling Toolkit(SRILM) (Stolcke, 2002) \u4f86\u8a13\u7df4\u8a9e\u8a00\u6a21\u578b\u3002\u8a9e\u8a00\u6a21\u578b\u7684\u8a13 WER CER Parameters RTF \u7387\u6703\u8d85\u8d8a\u57fa\u672c\u7684 TDNN\u3002\u95dc\u65bc Backstitch \u65b9\u9762\uff0c\u5716 4 \u4e2d\u865b\u7dda\u70ba\u8a13\u7df4\u96c6\uff0c\u5be6\u7dda\u70ba\u9a57\u8b49\u96c6\uff0c</td></tr><tr><td colspan=\"5\">ReLU \u6211\u5011\u5617\u8a66\u5169\u7a2e\u8a13\u7df4\u53c3\u6578\u7684\u65b9\u6cd5\u7528\u5728\u8072\u5b78\u6a21\u578b\u7684\u6548\u679c\uff0c\u5728 Natural Gradient (NG) (Povey, \uf996\u96c6\u7531 2001 \u53ca 2002 \uf98e\u7684\u65b0\u805e\u8a9e\uf9be\u6240\u7be9\u9078\u51fa\uf92d\u7684\u3002\u6e2c\u8a66\u96c6\u4e00\u5305\u542b\u4e94\u5834\u9304\u97f3\u5728 2003/01/28\uff0c TDNN+Backstitch(9 \u5c64) 25.14 17.45 15M 0.42 \u85cd\u8272\u03b1=1 \u76f8\u8f03\u7d05\u8272\u03b1=0.3 \u9700\u8981\u591a\u5169\u500d\u7684\u8fed\u4ee3\u624d\u6703\u6536\u6582\u5230\u76f8\u540c\u6e96\u78ba\u7387\uff0c\u53ef\u80fd\u662f\u7b2c\u4e8c\u6b65\u9a5f\u6c92</td></tr><tr><td colspan=\"5\">1536 Zhang & Khudanpur, 2014)\u4e0a\u9762\u505a\u6539\u8b8a\uff0c\u505a SGD \u66f4\u65b0\u6642\u6bcf\u82e5\u5e72\u6b21\u5728 backstitch \u7b2c\u4e00\u6b65\u9a5f\u5f8c LF-MMI criteria Cross entropy regularization \u505a\u534a\u6b63\u4ea4\u9650\u5236\u7684\u66f4\u65b0\uff0c\u7b2c\u4e8c\u6b65\u9a5f\u7dad\u6301\u539f\u672c\u7684\u505a\u6cd5\uff0c\u5728\u9000\u56de\u548c\u524d\u9032\u6b65\u9a5f\u9593\u505a\u6b63\u4ea4\u9650\u5236\u3002 2003/01/29\uff0c2003/02/11\uff0c2003/03/07 \u548c 2003/04/03\uff0c\u6e2c\u8a66\u96c6\u4e8c\u70ba\u53ea\u9078\u64c7\uf9ba\u63a1\u8a2a\u8a18\u8005\u8a9e\uf9be \u4e26\uf984\u6389\uf9ba\u542b\u6709\u8a9e\u52a9\u8a5e\u4e4b\u8a9e\uf906\u3002\u53e6\u4e00\u65b9\u9762\uff0c\u8072\u5b78\u6a21\u578b\u90fd\u662f\u5728\u958b\u6e90\u7684\u8a9e\u97f3\u8fa8\u8b58\u5de5\u5177 Kaldi 23.98 16.27 18M 0.47 TDNN+NF(15 \u5c64) \u6709\u505a\u6b63\u4ea4\u5316\u5ef6\u9072\u4e86\u6536\u6582\u7684\u901f\u5ea6\u3002</td></tr><tr><td colspan=\"5\">1536 160 SMAT 160*1536*3 MAT 1536*160 TDNN+NF TDNN+NF (Povey et al., 2011)\u4e0a\u8a13\u7df4\uff0c\u9996\u5148\u5728\u8a9e\u97f3\u8a9e\u6599\u5eab\u4e0a\u8a13\u7df4\u5177\u8a9e\u8005\u8abf\u9069\u6027(Speaker-adaptive)\u9ad8 TDNN+NF+Backstitch(10 \u5c64) 23.30 15.64 13M 0.37 TDNN+NF TDNN+NF 3.4 Dropout \u65af\u6df7\u548c\u96b1\u99ac\u53ef\u592b\u6a21\u578b(GMM-HMM)\uff0c\u4e26\u5229\u7528\u8a72\u6a21\u578b\u4f86\u7372\u5f97\u6240\u6709\u8a13\u7df4\u8a9e\u53e5\u7684\u8a5e\u5716\u4f86\u6e96\u5099\u5f8c TDNN+NF+Backstitch(15 \u5c64) 22.56 15.15 18M 0.47 \u7e8c\u8072\u5b78\u6a21\u578b\u4e4b\u985e\u795e\u7d93\u7db2\u8def\u8a13\u7df4\u3002\u7136\u5f8c\u4f7f\u7528 TDNN-LF-MMI \u6e96\u5247\u4f86\u8a13\u7df4\u51fa\u4e00\u985e\u795e\u7d93\u7db2\u8def\uff1b \u5728\u6a5f\u5668\u5b78\u7fd2\u4e2d\uff0c\u6a21\u578b\u7684\u53c3\u6578\u592a\u591a\u6703\u767c\u751f\u904e\u5ea6\u64ec\u5408\u73fe\u8c61\uff0c\u5728\u6bcf\u500b\u6279\u6b21\u8a13\u7df4\u4e2d\u5ffd\u7565\u4e00\u4e9b\u7279\u5fb5 TDNN+NF+Backstitch(20 \u5c64) 22.75 15.26 23M 0.58 \u5176\u4e2d\uff0c\u6700\u4f73\u5316\u7684\u90e8\u5206\u5247\u4f7f\u7528 NG \u548c Backstich\uff0c\u9075\u5faa(Povey et al., 2016)\u4e2d\u63cf\u8ff0\u7684\u65b9\u6cd5\u4f86\u5275 \u6aa2\u6e2c\u5668\uff0c\u5728 TDNN \u4e2d\u662f\u6a6b\u8de8\u6642\u9593\u7684\uff0c(Povey et al., 2018)\u4e0d\u662f\u7528\u4e8c\u503c\u7684\u96f6\u4e00\u4e1f\u68c4\u906e\u7f69\uff0c\u800c \u5efa\u8072\u5b78\u6a21\u578b\uff0c\u5373\u5c0d\u65bc 5,600 \u500b\u4f9d\u8cf4\u65bc\u4e0a\u4e0b\u6587\u7684\u8a9e\u97f3\u4e2d\u7684\u6bcf\u4e00\u500b\u5177\u6709\u4e00\u500b\u72c0\u614b\u7684 HMM \u62d3 \u8868 4. \u6539\u9032\u6a21\u578b\u5728\u6e2c\u8a66\u96c6\u4e8c\u8207\u767c\u5c55\u96c6\u7684\u5be6\u9a57\u7d50\u679c \u662f\u7528\u4e00\u500b\u9023\u7e8c\u578b\u5747\u52fb\u5206\u5e03[1-2\u03b1, 1+2\u03b1]\u3002\u6211\u5011\u4f7f\u7528\u4e00\u500b\u4e1f\u68c4\u6392\u7a0b\u8868\uff0c\u5728\u8a13\u7df4\u4e00\u958b\u59cb\u8a2d\u5b9a \u64b2\uff0c\u4ee5\u539f\u59cb\u5e40\u901f\u7387\u7684\u4e09\u5206\u4e4b\u4e00\u64cd\u4f5c\u3002\u6709\u7528\u8b8a\u901f\u64fe\u52d5\u7684\u8cc7\u6599\u64f4\u5145\uff0c\u7279\u5fb5\u4f7f\u7528 40 \u7dad\u7684 MFCC [Table 4. Experiment results for other test sets] \u03b1=0.2\uff0c\u8a13\u7df4\u5230\u4e00\u534a\u63d0\u5347\u5230 0.5\uff0c\u6700\u5f8c\u53c8\u4e0b\u964d\u5230 0\uff0c\u9019\u500b\u8a2d\u5b9a\u5728\u666e\u901a\u672a\u5206\u89e3\u7684 TDNN \u770b\u8d77 \u4f86\u6c92\u6709\u6548\u679c\uff0c\u4f46\u5728\u5206\u89e3\u5f8c\u7684\u7db2\u8def\u67b6\u69cb\u4e2d\u6709\u660e\u986f\u6539\u5584\u3002 \u548c 3 \u7dad\u7684\u8072\u8abf\u7279\u5fb5\u52a0\u4e0a 100 \u7dad\u7684 i-vectors \u505a\u8abf\u9069\u3002 \u6e2c\u8a66\u96c6\u4e8c (CER) \u767c\u5c55\u96c6 (CER)</td></tr><tr><td colspan=\"5\">TDNN+NF 4.2 \u5be6\u9a57\u7d50\u679c (Experiment Results) TDNN 5.05 3.5 \u6377\u5f91\u9023\u7d50 (Skip Connections) \u6839\u64da\u5f71\u50cf\u8fa8\u8b58\u88e1 VGG (Simonyan & Zisserman, 2014)\u7684\u767c\u5c55\uff0c\u7d93\u9a57\u4e0a\u589e\u52a0\u5c64\u6578\u53ef\u4ee5\u589e\u52a0\u6e96 33.39 TDNN+Backstitch 4.88 33.15</td></tr><tr><td colspan=\"5\">TDNN+NF \u78ba\u5ea6\uff0c\u4f46\u662f\u6703\u589e\u52a0\u8a13\u7df4\u4e0a\u7684\u96e3\u5ea6\u3002ResNet (He, Zhang, Ren & Sun, 2016)\u5728\u5176\u4e0a\u9032\u884c\u4fee\u6539\u5728 TDNN+NF 3.69 23.56</td></tr><tr><td colspan=\"5\">\u7db2\u8def\u4e0a\u589e\u52a0\u6377\u5f91\uff0c\u53ef\u4ee5\u9632\u6b62\u985e\u795e\u7d93\u7db2\u8def\u592a\u6df1\u800c\u7121\u6cd5\u8a13\u7df4\uff0c\u6211\u5011\u5728 TDNNF \u88e1\u4e5f\u505a\u4e0a\u985e\u4f3c TDNN+NF+Backstitch 3.67 22.73 \u7dad \u7684\u6a5f\u5236\uff0c\u6bcf\u5c64\u52a0\u4e0a\u8f38\u5165\u66f4\u524d\u9762\u4e00\u5c64\u7684\u4e09\u5206\u4e4b\u4e8c\u548c\u524d\u4e00\u5c64\u76f8\u52a0\u7576\u6210\u65b0\u7684\u8f38\u5165\u3002 \u7684\u96b1\u85cf\u5c64\uff0c\u77e9\u9663\u5206\u89e3\u74f6\u9838 160 \u7dad\uff0c\u524d\u5f8c\u6587\u97f3\u7a97\u5404 33 \u5e40\u3002TDNN+NF \u5728\u96b1\u85cf\u5c64\u7dad\u5ea6\u8b8a\u5927\u6642 RTF(Real Time Factor)\u662f\u4e00\u500b\u5e38\u7528\u65bc\u5ea6\u91cf\u81ea\u52d5\u8a9e\u97f3\u8fa8\u8b58\u7cfb\u7d71\u89e3\u78bc\u901f\u5ea6\u7684\u503c\uff0c\u5982\u679c\u8655</td></tr><tr><td colspan=\"5\">t \u6548\u679c\u8f03\u597d(Waibel et al., 1989)\uff0c\u800c\u57fa\u672c\u7684 TDNN \u6c92\u6709\u9019\u7a2e\u8b8a\u5316\u3002 4. \u5be6\u9a57 (EXPERIMENTS) \u7406\u4e00\u6bb5\u9577\u5ea6\u70ba a \u7684\u97f3\u8a0a\u4fe1\u865f\u9700\u8981\u82b1\u8cbb\u6642\u9593 b\uff0c\u5247 RTF \u70ba b/a\uff0c\u5716 3 \u662f\u4e0d\u540c\u6a21\u578b\u89e3\u78bc\u901f\u7387\u6bd4 t-1 t+1 \u8868 2. \u57fa\u790e\u8a9e\u97f3\u8fa8\u8b58\u5be6\u9a57 \u8f03\uff0c\u53ef\u4ee5\u770b\u51fa\u4f7f\u7528 LSTM \u6703\u63d0\u5347\u5927\u91cf\u6642\u9593\uff0c\u800c\u5176\u4ed6\u6a21\u578b\u96a8\u8457\u53c3\u6578\u63d0\u6607\u6703\u7a0d\u5fae\u589e\u52a0\u6642\u9593\u3002</td></tr><tr><td colspan=\"2\">4.1 \u5be6\u9a57\u8a2d\u5b9a (Experiment Setups) [Table 2. Baseline experiment results]</td><td/><td/></tr><tr><td/><td>\u8868 1. \u4e2d\u6587\u5ee3\u64ad\u65b0\u805e\u7684\u5be6\u9a57\u8a9e\u6599 WER RTF</td><td>CER</td><td>Parameters</td><td>RTF</td></tr><tr><td/><td>[Table 1. MATBN]</td><td/><td/></tr><tr><td colspan=\"5\">\u5176\u4e2d\uff0c\u5f0f(3)\u70ba Backstitch \u7b2c\u4e00\u6b65\u9a5f\u9000\u56de\u66f4\u65b0\uff0c\u800c\u5f0f(4)\u70ba Backstitch \u7b2c\u4e8c\u6b65\u9a5f\u524d\u9032\u66f4\u65b0\uff1b \u03b1\u9019\u500b\u5e38\u6578\u6c7a\u5b9a\u8981\u505a\u591a\u5927\u6b65\u4f10\u7684\u66f4\u65b0\uff0c\u6211\u5011\u53ef\u4ee5\u8abf\u6574\u8981\u5e7e\u500b Minibatch \u505a\u4e00\u6b21\u9019\u7a2e\u66f4\u65b0\uff0c \u4f4e\u5e40\u7387\u548c\u5c07\u50b3\u7d71 3-state \u7684 HMM \u62d3\u58a3\u6539\u70ba 2-state \u7684 HMM\uff0c\u7aef\u5c0d\u7aef\u7684\u9451\u5225\u5f0f\u8a13\u7df4\u4e5f\u501f\u9452 \u7528\u65bc\u5be6\u9a57\u7684\u8a9e\u6599\u9577\u5ea6\u548c\u53e5\u6578\uff0c\u5305\u542b\u5167\u5834\u65b0\u805e\u8207\u5916\u5834\u65b0\u805e\uff0c\u5176\u4e2d\u5167\u5834\u65b0\u805e\u70ba\u65b0\u805e\u4e3b\u64ad\u8a9e\u6599\uff0c \u5716 3. \u4e0d\u540c\u6a21\u578b\u89e3\u78bc\u901f\u7387\u6bd4\u8f03 3.1 \u8072\u5b78\u6a21\u578b\u4e4b\u985e\u795e\u7d93\u7db2\u8def\u67b6\u69cb(Structure) \u6240\u4ee5\u4e0d\u7528\u4e8b\u5148\u751f\u6210\u8a13\u7df4\u8a9e\u53e5\u4e4b\u8a5e\u5716\u3002\u7531\u65bc TDNN \u76f8\u9130\u7bc0\u9ede\u7684\u8b8a\u5316\u901a\u5e38\u4e0d\u5927\uff0c\u800c\u4e14\u8a0a\u606f\u91cd \u8907\u6a5f\u7387\u5f88\u9ad8\uff0c\u6240\u4ee5\u53ef\u4ee5\u8df3\u904e\u4e00\u4e9b\u5e40\u7684\u8a08\u7b97\u3002LFMMI \u7684\u5be6\u9a57\u8a2d\u5b9a\u5728(Povey et al., 2016)\u662f\u964d \u53e3\u8a9e\u5c0f\u7d44\u8207\u516c\u5171\u96fb\u8996\u53f0\u5408\u4f5c\u9304\u88fd\uff0c\u5171\u8a08 197 \u500b\u5c0f\u6642\uff0c\u53d6\u5176\u4e2d\u90e8\u5206\u7528\u65bc\u5be6\u9a57\uff0c\u8868 1 \u70ba\u5be6\u969b 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 Taiwan-Broadcast News, MATBN)\u3002\u516c\u8996\u65b0\u805e\u8a9e\u6599\u662f 2001 \u5e74\u81f3 2003 \u5e74\u9593\u7531\u4e2d\u7814\u9662\u8cc7\u8a0a\u6240 3. \u6a21\u578b\u67b6\u69cb\u548c\u8a13\u7df4\u65b9\u6cd5 (METHODS) \u7684\u68af\u5ea6\u4f86\u8a13\u7df4\uff0c\u56e0\u70ba LF-MMI \u5728\u8a13\u7df4\u4e2d\u8a08\u7b97\u6240\u6709\u8def\u5f91\u7684\u5f8c\u9a57\u6a5f\u7387(Posterior Probability)\uff0c \u672c\u8ad6\u6587\u5be6\u9a57\u8a9e\u6599\u4f86\u81ea\u516c\u8996\u65b0\u805e(Wang, Chen, Kuo & Cheng, 2005) (Mandarin Across TDNN9\u5c64+LSTM3\u5c64 \u7d42\u6a21\u578b\u7684\u5b57\u932f\u8aa4\u7387\u90fd\u6709\u986f\u8457\u6539\u5584\u3002 \u6839\u64da\u539f\u59cb\u8ad6\u6587\u6bd4\u8f03\u6709\u6548\u7387\u7684\u8a2d\u5b9a\u662f\u03b1=1 \u548c n=4\u3002 \u751f\u6210\u4e00\u500b\u8a5e\u5716\uff0c\u8a5e\u5716\u8981\u6709\u6b63\u78ba\u7d50\u679c\u7684\u8def\u5f91\u548c\u8db3\u5920\u9760\u8fd1\u7684\u5176\u4ed6\u8def\u5f91\uff0c\u9451\u5225\u5f0f\u6a21\u578b\u7528\u76ee\u6a19\u662f \u8981\u63d0\u9ad8\u8d70\u6b63\u78ba\u8def\u5f91\u4e4b\u6a5f\u7387\uff0c\u964d\u4f4e\u8d70\u76f8\u4f3c\u8def\u5f91\u4e4b\u6a5f\u7387\u3002 MMI(Maximum Mutual Information)\u6e96\u5247\u662f\u8981\u52a0\u5927\u6b63\u78ba\u8def\u5f91\u548c\u5176\u4ed6\u8def\u5f91\u7684\u6a5f\u7387\u5dee\u3002 \u53e6\u4e00\u65b9\u9762\uff0cLF-MMI \u900f\u904e\u5728\u985e\u795e\u7d93\u7db2\u8def\u8f38\u51fa\u8a08\u7b97\u6240\u6709\u53ef\u80fd\u7684\u5e8f\u5217\uff0c\u6839\u64da\u9019\u4e9b\u5e8f\u5217\u8a08\u7b97 MMI 3.7 \u6e2c\u8a66\u96c6\u4e00 3.6 \u6e2c\u8a66\u96c6\u4e8c 1.4 307 \u540c\u6df1\u5ea6\u7684 TDNN+NF \u6bd4\u8f03\u5be6\u9a57\u4e2d 15 \u5c64\u8868\u73fe\u6700\u597d\u3002\u8868 4 \u53ef\u898b\u5728\u4e0d\u540c\u96e3\u5ea6\u6e2c\u8a66\u96c6\u89e3\u78bc\u5f8c\u6700 TDNN9\u5c64 \u5be6\u9a57\uff0c\u5206\u89e3\u5f8c\u7684\u7db2\u8def\u5728\u53c3\u6578\u4e0a\u53ea\u591a\u4e86\u4e00\u4e9b\uff0c\u89e3\u78bc\u901f\u5ea6\u76f8\u7576\u4f46\u662f\u5b57\u932f\u8aa4\u7387\u964d\u4f4e\u5f88\u591a\uff0c\u5728\u4e0d 1,957 \u6539\u9032\u6a21\u578b\u5be6\u9a57\u7d50\u679c\u5982\u8868 3\uff0c\u5206\u5225\u5be6\u505a\u4e86 TDNN \u548c TDNN+NF \u53ca\u5404\u81ea\u52a0\u4e0a ackstitch \u7684 TDNN+NF10\u5c64 2,001 \u767c\u5c55\u96c6 TDNN(CE) 27.84 19.17 15M 0.47 \u80fd\u3002\u8a13\u7df4\u958b\u59cb\u4e4b\u524d\u5148\u8981\u6709\u6240\u6709\u55ae\u8a5e\u5e8f\u5217\u7d44\u5408\uff0c\u900f\u904e\u4ea4\u53c9\u71b5\u8a13\u7df4\u4e00\u500b\u6a21\u578b\uff0c\u914d\u5408\u8a9e\u8a00\u6a21\u578b 114.7 38,556 TDNN+NF15\u5c64 \u8a13\u7df4\u96c6 TDNN(LF-MMI) 26.22 18.34 15M 0.42 \u8fd1\u5e74\u4f86\uff0c\u900f\u904e\u53ef\u8996\u70ba\u9451\u5225\u5f0f\u6a21\u578b(Discriminant Model)\u7684\u985e\u795e\u7d93\u7db2\u8def\u80fd\u5920\u6709\u6548\u63d0\u5347\u7cfb\u7d71\u6548 \u9577\u5ea6(\u5c0f\u6642) \u53e5\u6578 Attention(LF-MMI) 26.76 18.96 50M 1.66 TDNN+NF20\u5c64</td></tr><tr><td colspan=\"5\">CTC \u4e0a\u7279\u6b8a\u7684\u7a7a\u767d\u6a19\u7c64(Graves, Fern\u00e1ndez, Gomez & Schmidhuber, 2006)\uff0c\u5728\u5c11\u91cf\u8a9e\u6599\u4e0b \u5916\u5834\u5305\u542b\u63a1\u8a2a\u8a18\u8005\u8a9e\u53d7\u8a2a\u8005\u8a9e\u6599\u3002\u80cc\u666f\u8a9e\u8a00\u6a21\u578b\u4f7f\u7528 5-gram \u8a9e\u8a00\u6a21\u578b\uff0c\u8a13\u7df4\u8a9e\u6599\u4f86\u81ea [Figure 3. Decoding speed comparison]</td></tr></table>" | |
| } | |
| } | |
| } | |
| } |