| { | |
| "paper_id": "O16-1028", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T08:05:03.303110Z" | |
| }, | |
| "title": "Deep Neural Networks for Audio Event Detection", | |
| "authors": [ | |
| { | |
| "first": "Jhih-Wei", | |
| "middle": [], | |
| "last": "\u9673\u667a\u5049", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "National Taipei University of Technology", | |
| "location": {} | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Chen", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "National Taipei University of Technology", | |
| "location": {} | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Chia-Hsin", | |
| "middle": [], | |
| "last": "\u5289\u4f73\u946b", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Taipei University of Technology", | |
| "location": {} | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Liu", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Taipei University of Technology", | |
| "location": {} | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Yuan-Fu", | |
| "middle": [], | |
| "last": "\u5ed6\u5143\u752b", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Taipei University of Technology", | |
| "location": {} | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Liao", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Taipei University of Technology", | |
| "location": {} | |
| }, | |
| "email": "[email protected]" | |
| } | |
| ], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "", | |
| "pdf_parse": { | |
| "paper_id": "O16-1028", | |
| "_pdf_hash": "", | |
| "abstract": [], | |
| "body_text": [], | |
| "back_matter": [], | |
| "bib_entries": { | |
| "BIBREF0": { | |
| "ref_id": "b0", | |
| "title": "SCREAM AND GUNSHOT DETECTION IN NOISY ENVIRONMENTS", | |
| "authors": [ | |
| { | |
| "first": "L", | |
| "middle": [], | |
| "last": "Gerosa", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Valenzise", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Tagliasacchi", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "F", | |
| "middle": [], | |
| "last": "Antonacci", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Sarti", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2007, | |
| "venue": "EURASIP European Signal Processing Conference", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "L. Gerosa, G. Valenzise, M. Tagliasacchi, F. Antonacci, A. Sarti, \"SCREAM AND GUNSHOT DETECTION IN NOISY ENVIRONMENTS,\" in EURASIP European Signal Processing Conference (EUSIPCO 2007), Poland, Sept, 2007", | |
| "links": null | |
| }, | |
| "BIBREF1": { | |
| "ref_id": "b1", | |
| "title": "Polyphonic Sound Event Detection Using Multi Label Deep Neural Networks", | |
| "authors": [ | |
| { | |
| "first": "Emre", | |
| "middle": [], | |
| "last": "Cakir", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Toni", | |
| "middle": [], | |
| "last": "Heittola", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Heikki", | |
| "middle": [], | |
| "last": "Huttunen", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Tuomas", | |
| "middle": [], | |
| "last": "Virtanen", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2015, | |
| "venue": "IEEE International Joint Conference on Neural Networks (IJCNN)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Emre Cakir, Toni Heittola, Heikki Huttunen and Tuomas Virtanen, \"Polyphonic Sound Event Detection Using Multi Label Deep Neural Networks,\" in IEEE International Joint Conference on Neural Networks (IJCNN), 2015.", | |
| "links": null | |
| }, | |
| "BIBREF2": { | |
| "ref_id": "b2", | |
| "title": "MULTI-LABEL VS. COMBINED SINGLE-LABEL SOUND EVENT DETECTION WITH DEEP NEURAL NETWORKS", | |
| "authors": [ | |
| { | |
| "first": "Emre", | |
| "middle": [], | |
| "last": "Cakir", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Toni", | |
| "middle": [], | |
| "last": "Heittola", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Heikki", | |
| "middle": [], | |
| "last": "Huttunen", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Tuomas", | |
| "middle": [], | |
| "last": "Virtanen", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2015, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Emre Cakir, Toni Heittola, Heikki Huttunen and Tuomas Virtanen, \"MULTI-LABEL VS. COMBINED SINGLE-LABEL SOUND EVENT DETECTION WITH DEEP NEURAL NETWORKS ,\" 2015.", | |
| "links": null | |
| }, | |
| "BIBREF3": { | |
| "ref_id": "b3", | |
| "title": "Contextdependent sound event detection", | |
| "authors": [ | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Heittola", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Mesaros", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Eronen", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Virtanen", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2013, | |
| "venue": "EURASIP Journal on Audio, Speech, and Music Processing", | |
| "volume": "2013", | |
| "issue": "1", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "T. Heittola, A. Mesaros, A. Eronen, and T. Virtanen, \"Contextdependent sound event detection,\" in EURASIP Journal on Audio, Speech, and Music Processing, vol. 2013, no. 1, 2013, p. 1.", | |
| "links": null | |
| }, | |
| "BIBREF4": { | |
| "ref_id": "b4", | |
| "title": "Convolutional, long short-term memory, fully connected deep neural networks", | |
| "authors": [ | |
| { | |
| "first": "Tara", | |
| "middle": [ | |
| "N" | |
| ], | |
| "last": "Sainath", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Oriol", | |
| "middle": [], | |
| "last": "Vinyals", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Andrew", | |
| "middle": [], | |
| "last": "Senior", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Has\u00b8im", | |
| "middle": [], | |
| "last": "Sak", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2015, | |
| "venue": "IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Tara N. Sainath, Oriol Vinyals, Andrew Senior, Has\u00b8im Sak, \"Convolutional, long short-term memory, fully connected deep neural networks,\" in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2015.", | |
| "links": null | |
| }, | |
| "BIBREF5": { | |
| "ref_id": "b5", | |
| "title": "TUT Database for Acoustic Scene Classification and Sound Event Detection", | |
| "authors": [ | |
| { | |
| "first": "Annamaria", | |
| "middle": [], | |
| "last": "Mesaros", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Toni", | |
| "middle": [], | |
| "last": "Heittola", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Tuomas", | |
| "middle": [], | |
| "last": "Virtanen", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2016, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Annamaria Mesaros, Toni Heittola, Tuomas Virtanen, \"TUT Database for Acoustic Scene Classification and Sound Event Detection,\" in In 24rd European Signal Processing Conference 2016 (EUSIPCO 2016), 2016.", | |
| "links": null | |
| }, | |
| "BIBREF6": { | |
| "ref_id": "b6", | |
| "title": "Environmental sound classification with convolutional neural networks", | |
| "authors": [ | |
| { | |
| "first": "Karol", | |
| "middle": [ | |
| "J" | |
| ], | |
| "last": "Piczak", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2015, | |
| "venue": "Proc. of MLSP)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "1--6", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Karol J. Piczak, \"Environmental sound classification with convolutional neural networks,\" in Proc. of MLSP). IEEE, 2015, pp. 1-6.", | |
| "links": null | |
| }, | |
| "BIBREF7": { | |
| "ref_id": "b7", | |
| "title": "Evaluation of Pooling Operations in Convolutional Architectures for Object Recognition", | |
| "authors": [ | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Scherer", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Muller", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Behnke", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": null, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "D. Scherer, A. Muller, and S. Behnke. \" Evaluation of Pooling Operations in Convolutional Architectures for Object Recognition \". In ICANN. 2010", | |
| "links": null | |
| }, | |
| "BIBREF8": { | |
| "ref_id": "b8", | |
| "title": "Regularization of NNs using DropConnect", | |
| "authors": [ | |
| { | |
| "first": "Li", | |
| "middle": [], | |
| "last": "Wan", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Matthew", | |
| "middle": [], | |
| "last": "Zeiler", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Sixin", | |
| "middle": [], | |
| "last": "Zhang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Yann", | |
| "middle": [], | |
| "last": "Lecun", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Rob", | |
| "middle": [], | |
| "last": "Fergus", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2013, | |
| "venue": "ICML", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Li Wan, Matthew Zeiler, Sixin Zhang, Yann LeCun, and Rob Fergus. \"Regularization of NNs using DropConnect,\" In ICML, 2013.", | |
| "links": null | |
| }, | |
| "BIBREF9": { | |
| "ref_id": "b9", | |
| "title": "Multi-Label Classification: An Overview", | |
| "authors": [ | |
| { | |
| "first": "Grigorios", | |
| "middle": [], | |
| "last": "Tsoumakas", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Ioannis", | |
| "middle": [], | |
| "last": "Katakis", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2007, | |
| "venue": "Int J Data Warehousing and Mining", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "1--13", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Grigorios Tsoumakas, Ioannis Katakis, \"Multi-Label Classification: An Overview,\" in Int J Data Warehousing and Mining, 2007, pp. 1-13", | |
| "links": null | |
| } | |
| }, | |
| "ref_entries": { | |
| "TABREF0": { | |
| "text": "Mixture Model, GMM) \u70ba \u57fa \u790e \u7684 \u8a9e \u8005 \u8fa8 \uf9fc \u6280 \u8853 \u3001 \u4ee5 \u652f \u6301 \u5411 \uf97e \u6a5f (Support VectorMachine,SVM) \u70ba \u57fa \u790e \u7684 \u8a9e \u8005 \u8fa8 \uf9fc \u6280 \u8853 \uff0c \u7d50 \u5408 \u9ad8 \u65af \u6df7 \u5408 \u6a21 \u578b \u8207 \u652f \u6301 \u5411 \uf97e \u6a5f (Hybrid GMM-SVM)\u4e4b\u96d9\u6a21\u578b\u7684\u8a9e\u8005\u8fa8\uf9fc\u6280\u8853\u3002 \u6bd4\u8cfd\u63d0\u4f9b\u7684 TUT \u8cc7\u6599\u5eab\uff0c\u6e2c\u8a66\u5927\u6703\u7d66\u4e88\u7684 GMM baseline \u8207\u6211\u5011 \u63d0\u51fa\u7684 DNN \u65b9\u6cd5\uff0cDCASE2016 \u6bd4\u8cfd\u985e\u578b\u5206\u70ba 4 \u9805\uff0c\u6211 \u5011\u9078\u64c7\u5176\u4e2d\u7684\u7b2c\u4e09\u500b\u4efb\u52d9\uff0cSound 86\uff0c\u50b3\u7d71\u6a21\u578b GMM \u5247\u70ba 0.91\uff0cDNN \u7684 F1 \u70ba 26.80%\uff0cGMM \u7684 F1\u5247\u70ba 23.40%\uff0c 03\uff0c\u800c DNN \u5247\u70ba 0.96\uff0c\u4ee5 F1 \u5206\u6578\u4f86\u770b\uff0cGMM \u70ba 17.60%\uff0c\u800c DNN \u5247\u70ba 12.80%\uff0c\u96d6 \u7136 DNN \u7684 F1 \u5206\u6578\u6bd4 GMM \u5dee\uff0c\u4f46\u662f\u56e0\u9019\u6b21\u6bd4\u8cfd\u4e3b\u8981\u662f\u6bd4\u932f\u8aa4\u7387\uff0c\u6240\u4ee5\u4ee5\u6574\u9ad4\u4f86\u770b\uff0c 60% 0.97 6.20% 0.96 12.80% \u4e94\u3001\u7d50\u8ad6 \u672c\u7814\u7a76\u4f7f\u7528 DNN\uff0c\u5efa\u7acb\u97f3\u8a0a\u4e8b\u4ef6\u8072\u5b78\u5075\u6e2c\u7cfb\u7d71\u3002\u4e26\u5229\u7528 Dropout \u9054\u5230\u6700\u4f73\u5316 DNN \u4e8b\u4ef6 \u6a21\u578b\u3002\u964d\u4f4e\u901a\u9053\u96dc\u8a0a\u5e72\u64fe\u8207\u80cc\u666f\u74b0\u5883\u7684\u5f71\u97ff\u3002\u5be6\u9a57\u7d50\u679c\u986f\u793a DNN \u4f7f\u7528\u4e8c\u5c64\u96b1\u85cf\u5c64\uff0c\u795e \u7d93\u5143\u6578\u70ba 64 \u6642\uff0c\u53ef\u5728 DCASE2016 \u6bd4\u8cfd\u6e2c\u8a66\u8cc7\u6599\u4e2d\u5f97\u5230\u6700\u4f73\u7d50\u679c\u3002\u82e5\u8207\u50b3\u7d71 GMM \u6bd4 \u8f03\uff0c\u5176\u5834\u666f\u5075\u6e2c\u932f\u8aa4\u7387\u53ef\u5f9e 0.91 \u964d\u81f3 0.86\u3001F1 \u5206\u6578\u4e26\u5f9e 23.4%\u63d0\u5347\u5230 26.8%\uff0c\u6b64\u5916\u91dd \u5c0d\u5ba4\u5167\u74b0\u5883\u7684\u97f3\u8a0a\u4e8b\u4ef6\u5075\u6e2c\u5be6\u9a57\uff0c\u932f\u8aa4\u7387\u53ef\u5f9e 1.06 \u964d\u81f3 0.86\uff0cF1 \u5206\u6578\u4e26\u5f9e 8.9%\u63d0\u5347\u5230 27.7%\uff0c\u800c\u5c0d\u6236\u5916\u74b0\u5883\u7684\u97f3\u8a0a\u4e8b\u4ef6\u5075\u6e2c\u5be6\u9a57\uff0c\u932f\u8aa4\u7387\u53ef\u5f9e 1.03 \u964d\u81f3 0.96\uff0cF1 \u5206\u6578\u4e26\u5f9e 17.6%\u5230 12.8%\uff0c\u56e0\u70ba\u6bd4\u8cfd\u4e3b\u8981\u662f\u770b\u932f\u8aa4\u7387\uff0c\u6240\u4ee5\u5f9e\u7e3d\u7d50\u679c\u4f86\u770b\uff0cDNN \u65b9\u6cd5\u6bd4 GMM \u65b9", | |
| "type_str": "table", | |
| "content": "<table><tr><td>\u4e00\u3001\u7c21\u4ecb \u4e8c\u3001\u76f8\u95dc\u7814\u7a76 \u8207\u975e\u4e8b\u4ef6\u672c\u8eab\u7684\u97f3\u6846\u7528\u4ee5\u8a13\u7df4\u6240\u6709\u53ef\u80fd\u7684 DNNs \u6a21\u578b[7]\u3002 (\u4e00) DNN \u539f\u7406 \u6709\u9ad8\u6297\u96dc\u8a0a\u80fd\u529b[9]\u3002 \u8868\u4e00\u3001DCASE2016\uff0cTask3 \u8a13\u7df4\u8cc7\u6599\u5eab \u7684\u7b2c 0 \u7dad\uff0c\u6240\u4ee5\u5171 19 \u7dad MFCCs\uff0c\u518d\u52a0\u4e0a\u4e00\u968e\u8207\u4e8c\u968e\u5c0e\u6578\u7d44\u6210\u5171 59 \u7dad\u7279\u5fb5\u7684\u5411\u91cf\uff0c\u5176 fn:\u932f\u8aa4\u5224\u5b9a\u70ba\u932f\u8aa4\u3002 \u8868\u56db\u3001\u5ba4\u5167\u74b0\u5883\u97f3\u8a0a\u4e8b\u4ef6\u5075\u6e2c\u932f\u8aa4\u7387\u8207 F1 \u5206\u6578</td></tr><tr><td>\u8072\u97f3\u662f\u4eba\u985e\u611f\u77e5\u74b0\u5883\u7684\u91cd\u8981\u8cc7\u8a0a\uff0c\u4e5f\u662f\u53cd\u6620\u4eba\u985e\u884c\u70ba\u7684\u91cd\u8981\u7279\u5fb5\u3002\u5c24\u5176\u662f\u5728\u67d0\u4e9b\u74b0\u5883 \u4e2d\uff0c\u4e00\u4e9b\u7279\u6b8a\u7684\u8072\u97f3\u4ee3\u8868\u4e86\u67d0\u7a2e\u72c0\u6cc1\u6b63\u5728\u767c\u751f\uff0c\u4f8b\u5982\uff1a\u5728\u8fa6\u516c\u5ba4\u88e1\uff0c\u6709\u9375\u76e4\u8072\u3001\u958b\u95dc\u9580 \u8072\u3001\u7b11\u8072\u3001\u73bb\u7483\u7834\u788e\u8072\u2026\u7b49\uff0c\u5728\u5c45\u5bb6\u74b0\u5883\u4e2d\uff0c\u6709\u71d2\u958b\u6c34\u8072\u3001\u5b30\u5152\u54ed\u8072\u3001\u8dcc\u5012\u8072\u3001\u958b\u9580\u8072\u2026 \u7b49\uff0c\u6216\u662f\u5728\u8857\u982d\u74b0\u5883\u4e0b\uff0c\u6709\u5587\u53ed\u8072\u3001\u78b0\u649e\u8072\u3001\u69cd\u64ca\u8072\u2026\u7b49\u3002 \u8072\u97f3\u4e8b\u4ef6\u5075\u6e2c\u7684\u5be6\u969b\u61c9\u7528\u5f88\u5ee3\u6cdb\uff0c\u4f8b\u5982\uff1a\u7f8e\u570b\u897f\u96c5\u5716\u653f\u5e9c\u65e5\u524d\u516c\u958b\u5c55\u793a\u4e00\u5957\u69cd\u8072\u5075\u6e2c\u7cfb \u7d71\uff1aShotSpotter\uff0c\u7528\u4ee5\u66f4\u6709\u6548\u5730\u904f\u6b62\u3001\u6253\u64ca\u57ce\u5e02\u72af\u7f6a[1]\u3002\u6216\u662f\u5e74\u8001\u7684\u9577\u8f29\u5e7e\u4e4e\u90fd\u7368\u81ea\u5728 \u5bb6\u88e1\uff0c\u5728\u5bb6\u4e2d\u6709\u53ef\u80fd\u6703\u767c\u751f\u4e8b\u60c5\uff0c\u4f8b\u5982\uff1a\u5fd8\u8a18\u81ea\u5df1\u6b63\u5728\u71d2\u958b\u6c34\uff0c\u5c0e\u81f4\u5f15\u767c\u706b\u707d\u3001\u5728\u6d74\u5ba4 \u8dcc\u5012\uff0c\u7121\u6cd5\u53ca\u6642\u6c42\u6551\u6cbb\u7642\uff0c\u9802\u6a13\u7a97\u6236\u88ab\u5c0f\u5077\u6253\u7834\uff0c\u5c0d\u5bb6\u88e1\u8ca1\u7269\u641c\u522e\u2026\u7b49\u3002\u6b64\u6642\u82e5\u6709\u97f3\u8a0a \u4e8b\u4ef6\u8072\u97f3\u5075\u6e2c\u7cfb\u7d71\uff0c\u5c31\u53ef\u4ee5\u5373\u6642\u63d0\u4f9b\u63f4\u52a9\u3002 \u50b3\u7d71\u7684\u8072\u97f3\u4e8b\u4ef6\u6280\u8853\u4e3b\u8981\u53ef\u5206\u70ba\u4e09\u500b\u4e3b\uf9ca\u7684\u6280\u8853\uf9d0\u5225\uff0c\u5206\u5225\u662f\u4ee5\u9ad8\u65af\u6df7\u5408\u6a21\u578b(Gaussian \u7136\u800c\uff0c\u5728\u5be6\u969b\u61c9\u7528\u74b0\u5883\u4e2d\uff0c\u82e5\u9047\u5230\u5e72\u64fe\u5075\u6e2c\u56e0\u7d20\uff0c\uf9b5\u5982\uff1a\u592a\u591a\u80cc\u666f\u96dc\u8a0a\u8072\u97f3\u7684\u5e72\u64fe\u6216\u9304 \u97f3\u54c1\u8cea\u592a\u5dee\u7b49\uff0c\u50b3\u7d71\u4ee5\u9ad8\u65af\u6df7\u5408\u6a21\u578b\u70ba\u57fa\u790e\u7684\u8a9e\u8005\u8fa8\uf9fc\u6280\u8853\u53ca\u4ee5\u652f\u6301\u5411\uf97e\u6a5f\u70ba\u57fa\u790e\u7684\u97f3 \u8a0a\u8fa8\uf9fc\u6280\u8853\uff0c\u56e0\uf967\u5177\u5099\u74b0\u5883\u9069\u61c9\u7684\u80fd\uf98a\u53ca\u5c0d\u65bc\u932f\u8aa4\u5bb9\u5fcd\u7684\u7a0b\ufa01\u592a\u4f4e\uff0c\u5e38\u6703\u5c0e\u81f4\u8fa8\uf9fc\u7cfb\u7d71 \u7684\u8fa8\uf9fc\u6027\u80fd\u7121\u6cd5\u7dad\u6301\u3002\u800c\u5c0d\u65bc\u7d50\u5408\u9ad8\u65af\u6df7\u5408\u6a21\u578b\u8207\u652f\u6301\u5411\uf97e\u6a5f\u7b49\uf978\u8005\u7684\u6a21\u578b\u800c\u8a00\uff0c\u96d6\u7136 \u8a72\u6280\u8853\u64f7\u53d6\uf978\uf9d0\u6a21\u578b\u8fa8\uf9fc\u6280\u8853\u7684\u512a\u9ede\uff0c\u4f46\u662f\u5176\u4ea6\uf967\u5177\u5099\u74b0\u5883\u9069\u61c9\u8207\u7cfb\u7d71\u5bb9\u932f\u7684\u80fd\uf98a\u3002\u9019 \u4e3b\u8981\u662f\u56e0\u70ba SVM \u5c6c\u65bc\u6dfa\u5c64\u5206\u6790\u6280\u8853\uff0c\u56e0\u6b64\u8a13\u7df4\u51fa\u4f86\u7684\u6a21\u578b\uff0c\u4ecd\u6613\u53d7\u8a0a\u865f\u7684\u8868\u9762\u8b8a\u6613\u5e72 \u64fe[2][4]\u3002 \u6700\u8fd1\u5e7e\u5e74\uff0c\u6df1\u5c64\u985e\u795e\u7d93\u7db2\u8def\u88ab\u5927\u91cf\u61c9\u7528\uff0c\u56e0\u5176\u53ef\u5c0d\u8a0a\u865f\u505a\u6df1\u5c64\u5206\u6790\uff0c\u5b78\u7fd2\u8a0a\u865f\u7684\u96b1\u6027\u7d50 \u69cb\uff0c\u56e0\u6b64\u8a13\u7df4\u51fa\u4f86\u7684\u6a21\u578b\u8f03\u4e0d\u6613\u53d7\u74b0\u5883\u96dc\u8a0a\uff0c\u4e0d\u5339\u914d\u7684\u9304\u97f3\u8a2d\u5b9a\u2026\u7b49\u7b49\u5f71\u97ff\uff0c\u5177\u6709\u5f37\u9375 \u6027\uff0c\u53ef\u80fd\u8f03\u9069\u5408\u88ab\u61c9\u7528\u5230\u8072\u97f3\u4e8b\u4ef6\u5075\u6e2c\u7cfb\u7d71[3]\u3002\u6240\u4ee5\u5728\u8ad6\u6587\u4e2d\uff0c\u6211\u5011\u5c07\u63a1\u7528\u6df1\u5c64\u985e\u795e \u7d93\u7db2\u8def\uff0c\u5be6\u505a\u5c45\u5bb6\u8207\u6236\u5916\u5169\u5834\u666f\u7684\u97f3\u8a0a\u4e8b\u4ef6\u5075\u6e2c\u7cfb\u7d71\uff0c\u4e26\u64da\u6b64\u53c3\u52a0 DCASE2016 \u8a55\u6bd4\uff0c \u5229\u7528\u5176\u5177\u516c\u4fe1\u529b\u7684\u8a9e\u6599\uff0c\u5c0b\u627e\u6700\u4f73\u7684 DNN \u8a2d\u5b9a\u3002 \u76ee\u524d\u70ba\u6b62\u6548\u80fd\u8f03\u9ad8\u7684\u8072\u97f3\u4e8b\u4ef6\u5075\u6e2c\u6a21\u578b\u5927\u81f4\u5206\u70ba\u4e0b\u5217\u5e7e\u7a2e: (1) \u50b3\u7d71\u9ad8\u65af\u6df7\u5408\u6a21\u578b (Gaussian Mixture Model, GMM)\u3002 (2)\u652f\u6301\u5411\u91cf\u6a5f\u5668(Support Vector Machine, SVM)\u3002 (3) hybrid GMM/SVM\u3002 \u9996\u5148\u7528\u9ad8\u65af\u6df7\u5408\u6a21\u578b(GMM)\u4f86\u4ee3\u8868\u8072\u97f3\u6a21\u578b\u7684\u4e3b\u8981\u7406\u7531\u6709\u5169\u500b\uff0c\u7b2c\u4e00\u500b\u7406\u7531\u662f\u9ad8\u65af\u6df7 \u5408\u6a21\u578b\u7684\u6bcf\u500b\u57fa\u672c\u5bc6\u5ea6\u51fd\u6578\u53ef\u4ee5\u6a21\u64ec\u51fa\u4e00\u4e9b\u8072\u97f3\u4e8b\u4ef6\u7684\u7279\u5fb5\u3002\u56e0\u6b64\u6211\u5011\u53ef\u4ee5\u7528\u9ad8\u65af\u6df7\u5408 \u6a21\u578b\u4e2d\u7b2c i \u500b\u5e73\u5747\u503c\u4f86\u4ee3\u8868\u7b2c i \u500b\u8072\u97f3\u7279\u5fb5\u7684\u983b\u8b5c\u5f62\u72c0\uff0c\u800c\u7528\u5171\u8b8a\u7570\u77e9\u9663\u4f86\u4ee3\u8868\u983b\u8b5c\u5f62 \u5c07\u6240\u6709\u4e8b\u4ef6\u6a21\u578b\u8a13\u7df4\u5b8c\u7562\u4e4b\u5f8c\uff0c\u5c07\u6e2c\u8a66\u97f3\u6a94\u7684\u97f3\u6846\u500b\u5225\u9001\u5165\u5404\u500b\u4e8b\u4ef6\u6a21\u578b\uff0c\u5728\u6e2c\u8a66\u6642\u5373 \u53ef\u4ee5\u5f97\u5230\u5404\u97f3\u6a94\u7684\u97f3\u6846\u5728\u4e0d\u540c\u4e8b\u4ef6\u6642\uff0c\u70ba\u4e8b\u4ef6\u672c\u8eab\u6216\u4e0d\u662f\u4e8b\u4ef6\u672c\u8eab\u7684\u5206\u6578\u503c\uff0c\u6b64\u5916\u70ba\u6c42 \u7a69\u5b9a\u5224\u65b7\uff0c\u6211\u5011\u518d\u4ee5 moving average \u6c42\u53d6\u97f3\u6846\u7684\u5e73\u5747\u5206\u6578\u7576\u4f5c\u6700\u5f8c\u7684\u5224\u65b7\u4f9d\u64da\uff0c\u56e0\u6b64\uff0c \u6700\u5f8c\u5206\u6578\u7684\u8a08\u7b97\u65b9\u5f0f\u5982\u5716\u4e09\u6240\u793a\u7684\u5206\u6578\u8a08\u7b97\u793a\u610f\u5716\u3002 \u5834\u666f \u8072\u97f3\u4e8b\u4ef6 \u97f3\u6a94 \u8072\u97f3\u4e8b\u4ef6 \u97f3\u6a94 \u5834\u666f \u8072\u97f3\u4e8b\u4ef6 \u97f3\u6a94 \u4e2d\u7684\u4e00\u968e\u8207\u4e8c\u968e\u5c0e\u6578\u63a5\u8003\u616e\u524d\u5f8c\u5404\u56db\u500b\u97f3\u6846\u3002 GMM DNN DNN \u67b6\u69cb\u5716\u4e3b\u8981\u5206\u70ba\u8f38\u5165\u5c64\u3001\u96b1\u85cf\u5c64\u548c\u8f38\u51fa\u5c64\u5176\u67b6\u69cb\u5982\u5716\u4e94\u6240\u793a\u3002\u8f38\u5165\u5c64\u5728\u7db2\u8def\u67b6\u69cb \u4e2d\u70ba\u8f38\u5165\u8a0a\u606f\u4e4b\u4e00\u65b9\uff0c\u5176\u795e\u7d93\u5143\u6578\u76ee\u8996\u8f38\u5165\u7279\u5fb5\u53c3\u6578\u6578\u91cf\u800c\u5b9a\u3002\u800c\u96b1\u85cf\u5c64\u4ecb\u65bc\u8f38\u5165\u8207\u8f38 \u51fa\u5c64\u9593\uff0c\u53ef\u4ee5\u70ba\u8907\u6578\u5c64\u6578\uff0c\u5176\u4f7f\u7528\u975e\u7dda\u6027\u6fc0\u6d3b\u51fd\u6578\u4f86\u8403\u53d6\u8cc7\u8a0a\uff0c\u96b1\u85cf\u5c64\u4e2d\u7684\u795e\u7d93\u5143\u6578\u91cf \u9700\u8981\u7d93\u7531\u5be6\u969b\u6e2c\u8a66\u8abf\u6574\u800c\u5b9a\uff0c\u96b1\u85cf\u5c64\u6578\u91cf\u4e5f\u8ddf\u795e\u7d93\u5143\u4e00\u6a23\u90fd\u9700\u7d93\u5be6\u9a57\u7372\u5f97\u7406\u60f3\u5c64\u6578\u3002\u6700 \u5f8c\uff0c\u8f38\u51fa\u5c64\u5728\u7db2\u8def\u67b6\u69cb\u4e2d\u70ba\u63d0\u4f9b\u8cc7\u6599\u8f38\u51fa\u4e4b\u4e00\u65b9\uff0c\u901a\u5e38\u4ee5\u4e00\u5c64\u8868\u793a\uff0c\u5176\u795e\u7d93\u5143\u6578\u76ee\u8996\u8f38 \u51fa\u7684\u5167\u5bb9\u800c\u5b9a\u3002\u6700\u5f8c\uff0c\u6df1\u5c64\u795e\u7d93\u7db2\u8def\u901a\u5e38\u5177\u5099\u81f3\u5c11\u4e8c\u500b\u4ee5\u4e0a\u7684\u96b1\u85cf\u5c64\uff0c\u591a\u51fa\u7684\u96b1\u85cf\u5c64\u662f \u70ba\u4e86\u63d0\u4f9b\u66f4\u9ad8\u7684\u62bd\u8c61\u5c64\u6b21\uff0c\u63d0\u9ad8\u6a21\u578b\u7684\u80fd\u529b\u3002 \u5716\u516d\u3001\u795e\u7d93\u5143\u904b\u7b97\u65b9\u5f0f\u793a\u610f\u5716 (\u4e8c) DNN \u8a13\u7df4 \u7531\u65bc\u5728\u6df1\u5c64\u985e\u795e\u7d93\u7db2\u8def\u4e2d\uff0c\u6240\u9700\u8abf\u6574\u7684\u7cfb\u7d71\u53c3\u6578\u592a\u591a\uff0c\u56e0\u6b64 DNN \u8a13\u7df4\u901a\u5e38\u4f7f\u7528 Gradient \u5c45\u5bb6 (object) Rustling 41 Drawer 23 \u6236\u5916 (object) Banging #. of layers 1 2 15 (object) Snapping 42 Glass jingling 26 Bird singing 162 Cupboard 27 Object impact 155 Car passing by 74 Cutlery 56 People walking 24 \uf06c GMM \u53c3\u6578\u8a2d\u5b9a\uff1a Event ER F1 ER F1 ER F1 (\u56db)\u5be6\u9a57\u7d50\u679c cupboard 1.00 0.00% 0.94 15.6% 0.93 22.00% \u4f9d\u64da\u4e3b\u8fa6\u55ae\u4f4d\u7d66\u4e88\u7684 GMM baseline \u8a2d\u5b9a\u6a19\u6e96\uff0c\u5728\u5be6\u9a57\u7576\u4e2d\uff0c\u6211\u5011\u5c07\u6bcf\u4e00\u500b\u4e8b\u4ef6\u8a13\u7df4\u70ba \u9996\u5148\u6e2c\u8a66\u97f3\u6a94\u5206\u70ba\u5c45\u5bb6\u8207\u6236\u5916\uff0c\u6211\u5011\u5206\u5225\u4f7f\u7528 DCASE2016 \u5927\u6703\u7d66\u7684 GMM \u548c\u6211\u5011\u63d0\u51fa cutlery 1.02 0.00% 1.00 0.00% 0.56 62.80% Children shouting 23 Dishes 94 Washing dishes 60 People speaking 41 Water tap running 37 People walking \u5169\u500b\u6a21\u578b\uff0c\u5206\u5225\u70ba\u4e8b\u4ef6\u672c\u8eab\u3001\u975e\u4e8b\u4ef6\u7684\u5176\u4ed6\u8072\u97f3\u3002\u6bcf\u500b GMM \u6a21\u578b\u4f7f\u7528\u6df7\u5408\u6578\u70ba 8\uff0c\u6240 \u4ee5\u4e00\u500b\u4e8b\u4ef6\u7684 GMM \u6a21\u578b\u7e3d\u6df7\u5408\u6578\u70ba 16(\u4e8b\u4ef6\u672c\u8eab+\u975e\u4e8b\u4ef6\u672c\u8eab)\u3002 dishes 1.16 2.50% 0.98 3.70% 0.87 41.90% \u7684 DNN \u6a21\u578b\u505a\u6e2c\u8a66\uff0c\u5728 DNN \u7cfb\u7d71\u4e2d\u6211\u5011\u6e2c\u8a66\u4e86\u4f7f\u7528\u4e00\u5c64\u548c\u4e8c\u5c64\u96b1\u85cf\u5c64\u7684\u60c5\u6cc1\u3002\u97f3\u8a0a drawer 1.19 0.00% 1.00 8.80% 0.92 26.00% 32 Wind blowing \u4e8b\u4ef6\u5075\u6e2c\u5206\u70ba\u5169\u5927\u985e\u5be6\u9a57\uff0c\u5171 3 \u500b\u5b50\u5be6\u9a57\uff0c\u5169\u5927\u985e\u5305\u62ec(1)\u5834\u666f\u5075\u6e2c\u548c(2)\u5c45\u5bb6\u8207\u6236\u5916\u97f3 glass_jingling 1.10 0.00% 0.95 8.70% 0.70 54.00% 22 \uf06c DNN \u53c3\u6578\u8a2d\u5b9a\uff1a \u8a0a\u4e8b\u4ef6\u5075\u6e2c\u3002\u5176\u4e2d\u5b50\u5be6\u9a57\u4e00\u70ba\u5834\u666f\u5075\u6e2c\uff0c\u76ee\u7684\u662f\u8981\u5340\u5206\u5834\u666f\u662f\u5728\u5c45\u5bb6\u6216\u6236\u5916\u74b0\u5883\u3002\u5b50\u5be6 object_impact 1.06 19.30% 1.00 0.00% 0.99 1.50% \u72c0\u7684\u8b8a\u5316\u3002\u7b2c\u4e8c\u500b\u7406\u7531\u662f\u9ad8\u65af\u6df7\u5408\u6a21\u578b\u80fd\u5f88\u5e73\u6ed1\u5730\u8fd1\u4f3c\u4efb\u610f\u5f62\u72c0\u7684\u5bc6\u5ea6\u3002\u55ae\u4e00\u578b\u614b\u9ad8\u65af \u6df7\u5408\u8a9e\u8005\u6a21\u578b\u662f\u5229\u7528\u4e00\u500b\u5e73\u5747\u503c\u5411\u91cf\u548c\u5171\u8b8a\u7570\u77e9\u9663\u4f86\u4ee3\u8868\u8072\u97f3\u4e8b\u4ef6\u7279\u5fb5\u53c3\u6578\u7684\u5206\u4f48\u60c5 \u5f62\u3002\u800c\u5411\u91cf\u91cf\u5316\u6a21\u578b\u5247\u662f\u5229\u7528\u4e00\u7d44\u96e2\u6563\u7684\u7279\u5fb5\u6a23\u677f\u4f86\u4ee3\u8868\u8a9e\u8005\u7684\u5206\u4f48\u3002\u9ad8\u65af\u6df7\u5408\u6a21\u578b\u53ef \u4ee5\u8aaa\u662f\u7d50\u5408\u4e86\u4e0a\u8ff0\u5169\u7a2e\u6a21\u578b\u7684\u512a\u9ede\uff0c\u5b83\u5229\u7528\u4e86\u4e00\u7d44\u96e2\u6563\u7684\u9ad8\u65af\u51fd\u6578\uff0c\u52a0\u4e0a\u9ad8\u65af\u51fd\u6578\u5177\u6709 \u7684\u5e73\u5747\u503c\u5411\u91cf\u548c\u5171\u8b8a\u7570\u77e9\u9663\u4f7f\u5f97\u5b83\u6709\u66f4\u597d\u7684\u6a21\u578b\u80fd\u529b\u3002 \u6b64\u5916\uff0c\u652f\u6301\u5411\u91cf\u6a5f\u5668(SVM)\u7684\u512a\u52e2\u5728\u65bc\u4f7f\u7528\u4e0a\u76f8\u7576\u5bb9\uf9e0\uff0cSVM \u4e3b\u8981\u8981\u627e\u51fa\u4e00\u500b\u8d85\u5e73\u9762 (hyperplane)\uff0c\u4f7f\u4e4b\u5c07\uf978\u500b\uf967\u540c\u7684\u96c6\u5408\u5206\u958b\u3002\u4ee5\u4e8c\u7dad\u7684\uf9b5\u5b50\uf92d\uf96f\uff0c\u6211\u5011\u5e0c\u671b\u80fd\u627e\u51fa\u4e00\u689d \u5206\u754c\u7dda\u80fd\u5920\u5c07\u76ee\u6a19\u96c6\u5408\u7684\u6a23\u672c\u9ede\u548c\u975e\u76ee\u6a19\u96c6\u5408\u7684\u8cc7\u6599\u9ede\u5206\u958b\uff0c\u800c\u4e14\u6211\u5011\u9084\u5e0c\u671b\u9019\u689d\u5206\u9694 \u7dda\u8ddd\uf9ea\u9019\uf978\u500b\u96c6\u5408\u7684\u908a\u754c(margin)\u8d8a\u5927\u8d8a\u597d\uff0c\u9019\u6a23\u6211\u5011\u624d\u80fd\u5920\u5f88\u660e\u78ba\u7684\u5206\u8fa8\u9019\u500b\u6a23\u672c\u9ede \u662f\u5c6c\u65bc\u90a3\u500b\u96c6\u5408\u3002 \u6700\u5f8c\uff0cHybrid GMM/SVM \u662f\u9ad8\u65af\u6df7\u5408\u6a21\u578b(GMM)\u548c\u5224\u5225\u652f\u6301\u5411\u91cf\u6a5f(SVM)\u7684\u7d50\u5408\u3002 GMM \u5c0d SVM \u7684\u8f38\u51fa\u505a\u8abf\u6574\uff0c\u5be6\u73fe SVM \u7684\u6982\u7387\u8f38\u51fa\uff0c\u4ee5\u9054\u5230\u8fa8\u8b58\u7387\u63d0\u5347\u7684\u76ee\u7684\u3002 \u4e09\u3001\u6df1\u5c64\u985e\u795e\u7d93\u7db2\u8def\u97f3\u8a0a\u4e8b\u4ef6\u8fa8\u8a8d\u7cfb\u7d71 \u5728\u672c\u5be6\u9a57\u4e2d\u6211\u5011\u4f7f\u7528 DNN \u505a\u97f3\u8a0a\u4e8b\u4ef6\u8fa8\u8a8d\u7cfb\u7d71\uff0c\u4e26\u7528\u4ee5\u53c3\u52a0 DCASE2016 \u6bd4\u8cfd\u3002 DCASE2016 \u4e3b\u8fa6\u55ae\u4f4d\u63d0\u4f9b\u7684\u8072\u97f3\u4e8b\u4ef6\u8cc7\u6599\u5206\u6210\u74b0\u5883\u8207\u4e8b\u4ef6\u5169\u985e\uff0c\u5171\u6709\u5169\u7a2e\u74b0\u5883\u8207 18 \u7a2e\u97f3\u8a0a\u4e8b\u4ef6\uff0c\u74b0\u5883\u5305\u62ec\u5c45\u5bb6\u8207\u6236\u5916\uff0c\u5176\u4e2d\u5c45\u5bb6\u74b0\u5883\u4e2d\u6709 11 \u7a2e\u97f3\u8a0a\u4e8b\u4ef6\uff0c\u6236\u5916\u5247\u6709 7 \u7a2e \u97f3\u8a0a\u4e8b\u4ef6[6]\u3002\u5716\u4e00\u662f\u6211\u5011\u4f7f\u7528\u7684 DNN \u97f3\u8a0a\u4e8b\u4ef6\u8fa8\u8a8d\u7cfb\u7d71\u67b6\u69cb\u3002 \u5716\u4e00\u3001DNN \u97f3\u8a0a\u4e8b\u4ef6\u5075\u6e2c\u7cfb\u7d71 \u5728\u6b64\u67b6\u69cb\u4e2d\u6211\u5011\u4f7f\u7528\u591a\u7d44 DNN \u5efa\u7acb\u97f3\u8a0a\u4e8b\u4ef6\u5075\u6e2c\u7cfb\u7d71\u6a21\u578b\uff0c\u56e0\u70ba\u5728\u4e0d\u540c\u74b0\u5883\u4e2d\uff0c\u5404\u7a2e \u4e8b\u4ef6\u90fd\u6709\u53ef\u80fd\u540c\u6642\u767c\u751f\uff0c\u6240\u4ee5\u6bcf\u4e00\u7a2e\u4e8b\u4ef6\u90fd\u9700\u8981\u5efa\u7acb\u4e00\u500b\u7368\u7acb\u7684 DNN \u6a21\u578b\uff0c\u5e73\u884c\u505a\u6e2c \u8a66\uff0c\u4e3b\u8981\u70ba\u4e86\u78ba\u4fdd\u7576\u4e8b\u4ef6\u540c\u6642\u767c\u751f\u6642\uff0c\u53ef\u4ee5\u540c\u6642\u88ab\u7cfb\u7d71\u5075\u6e2c\u5230[5]\uff0cDNN \u6a21\u578b\u8a13\u7df4\u7684\u793a \u610f\u5716\u5982\u5716\u4e8c\u3002 \u5716\u4e8c\u3001DNN \u8a13\u7df4\u6a21\u578b\u793a\u610f\u5716 \u5176\u4e2d\u5efa\u7acb\u5404\u500b\u97f3\u8a0a\u4e8b\u4ef6\u6a21\u578b\u6642\uff0c\u9996\u5148\u5c07\u8a13\u7df4\u7684\u97f3\u6a94\u97f3\u6846\u5316\u4e4b\u5f8c\uff0c\u518d\u64f7\u53d6\u97f3\u8a0a\u4e8b\u4ef6\u7684\u7279\u5fb5 \u53c3\u6578(MFCCs)\uff0c\u6211\u5011\u5148\u5c07\u6240\u6709\u8a13\u7df4\u7528\u97f3\u6a94\u53d6\u6885\u723e\u5012\u983b\u8b5c\u53c3\u6578\uff0c\u518d\u500b\u5225\u6536\u96c6\u5404\u7a2e\u4e8b\u4ef6\u672c\u8eab \u5716\u4e09\u3001\u97f3\u8a0a\u4e8b\u4ef6\u5206\u6578\u8a08\u7b97\u793a\u610f\u5716 \u9019\u6b21\u7684 DCASE2016 \u6311\u6230[6]\u63d0\u4f9b\u4e86\u5341\u516b\u7a2e\u4e8b\u4ef6\u8072\u97f3\uff0c\u5176\u4e8b\u4ef6\u8207\u4e8b\u4ef6\u4e4b\u9593\u64c1\u6709\u540c\u6642\u767c\u751f\u7684 \u6a5f\u6703\uff0c\u6545\u6700\u5f8c\u5075\u6e2c\u7684\u7d50\u679c\u6a19\u7c64\u9700\u70ba\u591a\u91cd\u6a19\u7c64[3-5]\uff0c\u7d50\u679c\u8f38\u51fa\u898f\u5b9a\u7684\u683c\u5f0f\u5982\u5716\u56db\u3002 \u5716\u56db\u3001\u97f3\u8a0a\u4e8b\u4ef6\u5075\u6e2c\u7d50\u679c\u8f38\u51fa\u65b9\u5f0f \u5716\u4e94\u3001\u591a\u5c64 DNN \u67b6\u69cb\u5716 DNN \u4e2d\u55ae\u4e00\u7684\u795e\u7d93\u5143\u7684\u904b\u7b97\u65b9\u5f0f\u5982\u5716\u516d\u6240\u793a\uff0c\u7531\u8f38\u5165\u7684\u53c3\u6578 X \u8207\u9023\u7d50\u6b0a\u503c W\uff0c\u9032\u884c\u9023 \u4e58\u52a0\u7684\u52d5\u4f5c\uff0c\u6b64\u4e00\u6b65\u9a5f\u53ef\u85c9\u7531\u96c6\u6210\u51fd\u6578(Summation Function)\u5b8c\u6210\uff0c\u96c6\u6210\u51fd\u6578\u7684\u76ee\u7684\u5728\u65bc \u5c07\u524d\u4e00\u5c64\u4e4b\u8f38\u51fa\u7d93\u7531\u7db2\u8def\u7684\u9023\u7d50\u6b0a\u91cd\u503c\u532f\u96c6\u81f3\u795e\u7d93\u5143\u4e2d\uff0c\u901a\u5e38\u662f\u4ee5\u51fd\u6578\u7684\u65b9\u5f0f\u52a0\u4ee5\u8868\u9054 \u5176\u516c\u5f0f\u5982\u516c\u5f0f(1)\uff0c\u5176\u4e2d W \u70ba\u9023\u7d50\u6b0a\u503c\uff0cX \u70ba\u8f38\u5165\u8b8a\u6578\uff0cb \u70ba\u8a72\u795e\u7d93\u5143\u7684\u504f\u6b0a\u503c\u3002\u7d93\u7531 \u6b64\u516c\u5f0f\u904b\u7b97\u5f8c\uff0c\u5176\u8f38\u51fa\u6578\u503c\u8d8a\u5927\uff0c\u5247\u4ee3\u8868\u795e\u7d93\u5143\u88ab\u6fc0\u767c\uff1b\u8f38\u51fa\u6578\u503c\u8d8a\u5c0f\uff0c\u5247\u53cd\u4e4b\u3002\u6700\u5f8c \u518d\u7d93\u7531\u4f5c\u7528\u51fd\u6578 (Activation Function)\u904b\u7b97\u8f38\u51fa\uff0c\u6210\u70ba\u4e0b\u4e00\u5c64\u795e\u7d93\u5143\u7684\u8f38\u5165\u503c\u3002 \u4e0d\u518d\u96a8\u6a5f\u9078\u64c7\u5ffd\u7565\u800c\u662f\u5168\u90e8\u795e\u7d93\u5143\u7684\u8f38\u51fa\u7684\u5e73\u5747\u503c\u5982\u5716\u4e03(b)\u6240\u793a\uff0c\u5982\u6b64\u53ef\u4ee5\u4f7f\u6a21\u578b\u64c1 \u6b64\u516c\u5f0f\u7684\u7b26\u865f\u5b9a\u7fa9\u5982\u4e0b\uff1atp:\u6b63\u78ba\u5224\u5b9a\u70ba\u6b63\u78ba\uff1bfp:\u6b63\u78ba\u5224\u5b9a\u70ba\u932f\u8aa4\uff1btn:\u932f\u8aa4\u5224\u5b9a\u70ba\u6b63\u78ba\uff1b \u6846\u9593\u7684\u8b8a\u5316\u592a\u5287\u70c8\uff0c\u6211\u5011\u5c07\u5169\u500b\u97f3\u6846\u4e4b\u9593\u53d6 20ms \u91cd\u758a\u3002\u5728\u5be6\u9a57\u8a2d\u5b9a\u4e2d\u4e0d\u63a1\u7528 MFCCs \u51fa\u73fe\u67d0\u4e9b\u7279\u5fb5\u50c5\u50c5\u5728\u5176\u4ed6\u7279\u5fb5\u4e0b\u624d\u6709\u6548\u679c\u7684\u60c5\u6cc1\uff1b\u4f46\u5728\u6e2c\u8a66\u8a9e\u6599\u8f38\u5165\u6642\uff0c\u96b1\u85cf\u5c64\u7279\u5fb5\u5c07 (1) Descent\uff0cGradient Descent \u7684\u516c\u5f0f\u5982(2)\u8868\u793a\u3002 (2) \u5176\u4e2d \u70ba\u6700\u4f73\u5316\u6642\u8981\u8abf\u6574\u7684\u53c3\u6578\uff0c \u662f\u521d\u59cb\u5b78\u7fd2\u7387\uff0c \u70ba\u7576\u524d\u7684\u68af\u5ea6\uff0c\u6b64\u5916\u5728\u4f7f\u7528 Gradient Descent \u6f14\u7b97\u6cd5\u524d\uff0c\u6211\u5011\u5fc5\u9808\u5148\u5b9a\u7fa9\u4e00\u500b Cost Function \u624d\u80fd\u8a08\u7b97\u68af\u5ea6\uff0c\u6700\u5e38\u7528\u7684 Cost Function \u70ba cross entropy\uff0c\u5176\u516c\u5f0f\u5982\u516c\u5f0f(3)\u8868\u793a\uff1a (3) \u5176\u4e2d\uff0ca \u70ba DNN \u8f38\u51fa\u6a5f\u7387\u503c\uff0cy \u70ba\u6b63\u78ba\u985e\u5225\u6307\u6a19\u7b54\u6848\u3002 (\u4e09) DNN \u6700\u4f73\u5316 \u6b64\u5916\u70ba\u5efa\u7acb\u5177\u5f37\u5065\u6027\u7684 DNN \u6a21\u578b\uff0c\u5e38\u5229\u7528 Dropout \u6f14\u7b97\u6cd5\u3002Dropout \u5728\u57f7\u884c\u6642\uff0c\u6211\u5011\u5c07 \u96a8\u6a5f\u9078\u64c7\u5ffd\u7565\u96b1\u85cf\u5c64\u8403\u53d6\u51fa\u7684\u8a9e\u6599\u7279\u5fb5\uff0c\u65b9\u5f0f\u5982\u5716\u4e03(a)\u6240\u793a\uff0c\u6bcf\u500b\u6279\u6b21\u7684 DNN \u6a21\u578b\u8a13 \u7df4\u904e\u7a0b\u4e2d\uff0c\u56e0\u70ba\u6bcf\u6b21\u96a8\u6a5f\u5ffd\u7565\u7684\u96b1\u85cf\u5c64\u8a9e\u6599\u7279\u5fb5\u90fd\u4e0d\u4e00\u6a23\uff0c\u6240\u4ee5\u4f7f\u6bcf\u6b21\u6a21\u578b\u4e2d\u8a13\u7df4\u5230\u7684 \u985e\u795e\u7d93\u7db2\u8def\u90fd\u5c07\u662f\u4e0d\u540c\u7684\u6a23\u5f0f\uff0c\u6bcf\u6b21\u8a13\u7df4\u90fd\u5982\u540c\u505a\u4e00\u500b\u65b0\u7684\u97f3\u8a0a\u4e8b\u4ef6\u6a21\u578b\uff1b\u9664\u6b64\u4e4b\u5916\uff0c \u96b1\u85cf\u5c64\u8a9e\u6599\u7279\u5fb5\u90fd\u662f\u4ee5\u4e00\u5b9a\u6a5f\u7387\u96a8\u6a5f\u51fa\u73fe\uff0c\u56e0\u6b64\u4e0d\u80fd\u4fdd\u8b49\u6bcf\u5169\u500b\u96b1\u85cf\u5c64\u8a9e\u6599\u7279\u5fb5\u6bcf\u6b21\u90fd \u540c\u6642\u51fa\u73fe\uff0c\u9019\u6a23\u6b0a\u91cd\u7684\u66f4\u66ff\u4e0d\u518d\u4f9d\u8cf4\u65bc\u6709\u56fa\u5b9a\u95dc\u4fc2\u96b1\u85cf\u5c64\u8a9e\u6599\u7279\u5fb5\u7684\u5171\u540c\u4f5c\u7528\uff0c\u5c31\u4e0d\u6703 C= -\u5716\u4e03\u3001DNN \u6a21\u578b Dropout \u8a13\u7df4\u793a\u610f\u5716 \u56db\u3001\u5834\u666f\u8207\u97f3\u8a0a\u4e8b\u4ef6\u5075\u6e2c\u5be6\u9a57\u7d50\u679c \u5be6\u9a57\u4f7f\u7528 DCASE2016 event detection in real life audio(Task3)\u3002Task3 \u4efb\u52d9\u662f\u5728\u8a55\u4f30\u6211\u5011\u65e5\u5e38\u751f\u6d3b\u4e2d\u7684\u97f3\u8a0a\u4e8b \u4ef6\uff0c\u5176\u4e2d\u7684\u8072\u6e90\u90fd\u662f\u5728\u5177\u6709\u80cc\u666f\u97f3\u5e72\u64fe\u7684\u60c5\u5f62\u4e0b\u7684\u591a\u91cd\u4e8b\u4ef6\u3002\u4efb\u52d9\u8981\u6c42\u662f\u7576\u4e8b\u4ef6\u767c\u751f \u6642\uff0c\u7cfb\u7d71\u80fd\u5426\u6b63\u78ba\u5075\u6e2c\u5230\u4e8b\u4ef6\u7684\u767c\u751f\uff0c\u9084\u6709\u7576\u591a\u500b\u4e8b\u4ef6\u540c\u6642\u767c\u751f\u6642\uff0c\u7cfb\u7d71\u662f\u5426\u80fd\u540c\u6642\u5224 \u65b7\u51fa\u591a\u500b\u4e8b\u4ef6[8][10]\u3002 (\u4e00)\u3001\u5be6\u9a57\u8cc7\u6599\u5eab DCASE2016 \u6bd4\u8cfd\u63d0\u4f9b\u7684 TUT \u8a9e\u6599\u4e2d\u5206\u70ba\u5c45\u5bb6\u8207\u6236\u5916 2 \u500b\u4e0d\u540c\u7684\u5834\u666f\uff0c\u5728\u4e0d\u540c\u5834\u666f\u5404\u6709 \u4e0d\u540c\u7684\u8072\u97f3\u4e8b\u4ef6\uff0c\u9019\u4e9b\u9304\u97f3\u5728\u591a\u500b\u4e0d\u540c\u7684\u4f4d\u7f6e\u9304\u88fd\uff0c\u5305\u62ec\u4e0d\u540c\u7684\u8857\u9053\uff0c\u4e0d\u540c\u7684\u5bb6\u5ead\u3002\u6bcf \u6b21\u9304\u97f3\u6642\u9304\u88fd\u4e00\u500b 3~5 \u5206\u9418\u9577\uff0c44.1 kHz \u53d6\u6a23\u7387\u7684\u97f3\u6a94\uff0c\u6bcf\u500b\u97f3\u6a94\u9577\u5ea6\u7686\u4e0d\u540c\uff0c\u4e8b\u4ef6\u9577 \u77ed\u4e5f\u4e0d\u4e00\u6a23\uff0c\u4e26\u4ee5\u4eba\u5de5\u4f9d\u4e8b\u4ef6\u767c\u751f\u6642\u9593\u4f4d\u7f6e\uff0c\u7d66\u4e88\u6a19\u7c64\uff0c\u7576\u4f5c\u6a19\u6e96\u7b54\u6848\u3002\u8cc7\u6599\u5eab\u5167\u5bb9\u5305 \u542b\u8868\u4e00\u7684\u8a13\u7df4\u8cc7\u6599\u8207\u8868\u4e8c\u7684\u6e2c\u8a66\u8cc7\u6599\uff1a \u8868\u4e8c\u3001DCASE2016\uff0cTask3 \u6e2c\u8a66\u8cc7\u6599\u5eab \u74b0\u5883 \u97f3\u6a94\u500b\u6578 \u97f3\u6a94\u7e3d\u9577 \u5c45\u5bb6 10 36min16s \u6236\u5916 12 42min R-09 waveform recorder \u505a\u9304\u97f3\u3002 (\u4e8c)\u5be6\u9a57\u8a2d\u5b9a \u6211\u5011\u5148\u5c07\u6240\u6709\u8a13\u7df4\u8cc7\u6599\u5eab\u4e2d\u97f3\u8a0a\u6a94\u6848\u53d6\u6885\u723e\u5012\u983b\u8b5c\u53c3\u6578(MFCCs)\uff0c\u5c07\u5176\u97f3\u6846\u5316\u4e4b\u5f8c\uff0c\u5c07 \u6bcf\u500b\u4e8b\u4ef6\u7684\u97f3\u6846\u9001\u9032\u9ad8\u65af\u6df7\u5408\u6a21\u578b\u8a13\u7df4\u3002\u6700\u5f8c\u4f7f\u7528\u6e2c\u8a66\u8cc7\u6599\u5eab\u4e2d\u7684\u97f3\u8a0a\u6a94\u6848\u9032\u884c\u6e2c\u8a66 \u6642\uff0c\u5148\u5f97\u5230\u7684\u6bcf\u500b\u97f3\u6846\u7684\u5206\u6578\uff0c\u518d\u7d93\u904e\u8207\u95a5\u503c\u6b64\u5c0d\u5224\u65b7\uff0c\u5982\u5206\u6578\u5230\u9054\u6a19\u6e96\uff0c\u518d\u5c07\u4e8b\u4ef6\u6a19 \u8a18\u5beb\u5165\u6587\u672c\uff0c\u6700\u5f8c\u518d\u62ff\u6587\u672c\u8207\u6b63\u78ba\u7b54\u6848\u76f8\u4e92\u6bd4\u5c0d\uff0c\u5f97\u5230\u932f\u8aa4\u7387\u8207 F1 \u5206\u6578\uff0c\u4ee5\u4e0b\u8a73\u7d30\u8aaa \u660e\u5404\u90e8\u5206\u8a2d\u5b9a\u7d30\u7bc0\uff1a \uf06c \u524d\u8655\u7406\uff1a \u6c42\u53d6\u6885\u723e\u5012\u983b\u8b5c\u53c3\u6578\u6642\uff0c\u5176\u4e2d\u6ffe\u6ce2\u5668\u6578\u91cf\u70ba 40\uff0c\u6885\u723e\u5012\u983b\u8b5c\u53c3\u6578\u70ba 20 \u7dad\u3001\u983b\u7387\u7bc4\u570d\u53d6 0 Hz ~22050 Hz\u3001\u5085\u7acb\u8449\u8f49\u63db\u70ba 2048\uff0c\u97f3\u6a94\u4f7f\u7528\u7684\u97f3\u6846\u5927\u5c0f\u70ba 40ms\u3002\u6b64\u5916\u70ba\u4e86\u907f\u514d\u97f3 object_rustling 1.09 7.00% 1.00 0.00% 1.00 0.00% \u9996\u5148\u5728\u4e00\u524d\u7f6e\u5be6\u9a57\u4e2d\uff0c\u6211\u5011\u5148\u6e2c\u8a66\u8f38\u5165\u97f3\u6846\u6578\u8207\u985e\u795e\u7d93\u5143\u6578\u76ee\uff0c\u7d93\u628a\u97f3\u6846\u6578\u8a2d\u5b9a\u70ba 1\u3001 5 \u6216 9 \u505a\u6e2c\u8a66\uff0c\u7d50\u679c\u5728\u97f3\u6846\u6578\u70ba 1 \u6642\uff0c\u64c1\u6709\u8f03\u4f73\u7d50\u679c\u3002\u795e\u7d93\u5143\u6578\u5247\u66fe\u7d93\u6e2c\u8a66 32\u300164 \u6216 128\u3002\u7d50\u679c\u5728\u795e\u7d93\u5143\u6578\u70ba 64\u3001Dropout \u8a66\u904e 0.9\u30010.7\u30010.5\uff0c\u7d50\u679c 0.7 \u6642\u6548\u679c\u8f03\u597d\u3002\u56e0\u6b64\uff0c \u5728\u4ee5\u4e0b\u5be6\u9a57\u4e2d\u7686\u63a1\u7528\u8f38\u5165\u97f3\u6846\u6578\u70ba 1\uff0c\u795e\u7d93\u5143\u6578 64 \u8207 Dropout \u70ba 0.7 \u7684\u8a2d\u5b9a\uff0c\u4e26\u9032\u4e00\u6b65 \u6e2c\u8a66 DNN \u7684\u5c64\u6578\u3002 (\u4e09)\u8a55\u4f30\u65b9\u5f0f \u7cfb\u7d71\u7684\u8a55\u5206\u6a19\u6e96\u6709\u5169\u7a2e\u4e00\u7a2e\u662f\u932f\u8aa4\u7387(error rate)\uff0c\u53e6\u4e00\u500b\u5247\u662f F1 \u5206\u6578\u3002\u932f\u8aa4\u7387\u7684\u8a08\u7b97\u65b9 \u5f0f\u70ba\uff1a (4) \u5176\u4e2d\u7684 N \u4ee3\u8868\u6b63\u78ba\u7b54\u6848\u7684\u4e8b\u4ef6\u767c\u751f\u6578\u3002\u4e09\u7d44\u6578\u64da\u5206\u5225\u70ba\uff1a\u63d2\u5165\u932f\u8aa4(Insertion,I)\u3001\u53d6\u4ee3\u932f \u8aa4(Substitution,S)\u53ca\u522a\u9664\u932f\u8aa4(Deletion,D)\u3002F1 \u5206\u6578\u7684\u516c\u5f0f\u5982\u4e0b\uff1a F1= (5) \u5176\u4e2d\u7684 precision \u548c recall \u70ba \uff0c (6) \u9a57\u4e8c\u70ba\u5c45\u5bb6\u74b0\u5883\u97f3\u8a0a\u4e8b\u4ef6\u5075\u6e2c\uff0c\u8981\u5728\u5c45\u5bb6\u74b0\u5883\u4e2d\u8981\u5075\u6e2c 11 \u7a2e\u97f3\u8a0a\u4e8b\u4ef6\u3002\u5b50\u5be6\u9a57\u4e09\u5247\u70ba \u6236\u5916\u97f3\u8a0a\u4e8b\u4ef6\u5075\u6e2c\uff0c\u8981\u5075\u6e2c 7 \u7a2e\u4e0d\u540c\u7684\u97f3\u8a0a\u4e8b\u4ef6\u3002 \u9996\u5148\u5728\u5b50\u5be6\u9a57\u4e00\u5834\u666f\u6e2c\u8a66\u5be6\u9a57\u65b9\u9762\uff0c\u5f9e\u8868\u4e09\u7684\u5be6\u9a57\u7d50\u679c\u4f86\u770b\uff0cDNN \u7cfb\u7d71\u7684\u5e73\u5747\u7e3d\u932f\u8aa4 \u7387 0.\u56e0\u6b64\uff0c\u4ee5\u5be6\u9a57\u7d50\u679c\u4f86\u770b\uff0cDNN \u7684\u932f\u8aa4\u7387\u8207 F1 \u90fd\u662f\u6700\u4f73\u3002\u4f46\u662f\u4ee5\u932f\u8aa4\u7387\u4f86\u770b\uff0cDNN \u7cfb \u7d71\u9084\u6709\u9032\u6b65\u7684\u7a7a\u9593\uff0c\u8a73\u7d30\u7d50\u679c\u5982\u8868\u4e09\u6240\u793a\u3002 \u8868\u4e09\u3001Performance of Scene Recognition GMM DNN #. of layers 1 2 Scene ER F1 ER F1 ER F1 home 0.97 15.40% 0.93 13.20% 0.82 31.90% residential 0.86 31.50% 0.95 11.50% 0.90 21.70% Average 0.91 23.40% 0.94 12.30% 0.86 26.80% \u8868\u56db\u70ba\u5b50\u5be6\u9a57\u4e8c\u5c45\u5bb6\u97f3\u8a0a\u4e8b\u4ef6\u5075\u6e2c\u7684\u5be6\u9a57\u7d50\u679c\u3002\u5f9e\u5e73\u5747\u932f\u8aa4\u7387\u4f86\u770b\uff0cGMM \u70ba 1.06\uff0c\u800c DNN \u5247\u70ba 0.86\uff0c\u4ee5 F1 \u5206\u6578\u4f86\u770b\uff0cGMM \u70ba 8.90%\uff0c\u800c DNN \u5247\u70ba 27.70%\uff0c\u7e3d\u9ad4\u800c\u8a00\u5728 \u5ba4\u5167\u74b0\u5883\u4f7f\u7528 DNN \u97f3\u8a0a\u4e8b\u4ef6\u5075\u6e2c\u7cfb\u7d71\u6bd4\u8f03\u597d\u3002 object_snapping 1.00 0.00% 1.00 0.00% 1.00 0.00% people_walking 1.10 14.80% 1.00 0.00% 1.00 0.00% washing_dishes 1.08 20.30% 0.96 25.90% 0.92 29.70% water_tap_running 0.83 34.10% 0.79 39.60% 0.54 66.70% Average 1.06 8.90% 0.97 9.30% 0.86 27.70% \u6700\u5f8c\u8868\u4e94\u70ba\u5b50\u5be6\u9a57\u4e09\u6236\u5916\u74b0\u5883\u7684\u97f3\u8a0a\u4e8b\u4ef6\u5075\u6e2c\u5be6\u9a57\u7d50\u679c\u3002\u5f9e\u5e73\u5747\u932f\u8aa4\u7387\u4f86\u770b\uff0cGMM \u70ba 1.DNN \u9084\u662f\u6bd4 GMM \u597d\u3002 \u8868\u4e94\u3001\u6236\u5916\u74b0\u5883\u97f3\u8a0a\u4e8b\u4ef6\u5075\u6e2c\u932f\u8aa4\u7387\u8207 F1 \u5206\u6578 GMM DNN #. of layers 1 2 Event ER F1 ER F1 ER F1 bird_singing 0.87 30.10% 1.04 3.60% 0.97 31.60% car_passing_by 0.71 54.50% 0.77 37.70% 0.95 24.20% children_shouting 1.07 0.00% 1.00 0.00% 1.00 0.00% object_banging 1.00 0.00% 1.00 0.00% 0.82 34.00% people_speaking 0.89 25.00% 1.00 0.00% 1.00 0.00% people_walking 1.15 1.70% 1.00 0.00% 1.00 0.00% wind_blowing 1.53 11.80% 1.01 2.20% 1.00 0.00% Average \u81f4\u8b1d \u672c\u7814\u7a76\u611f\u8b1d\u6559\u80b2\u90e8\u300e\u5927\u5b78\u4ee5\u793e\u6559\u6a5f\u69cb\u70ba\u57fa\u5730\u4e4b\u6578\u4f4d\u4eba\u6587\u8a08\u756b\u300f (A36 \u865f)\u8207\u79d1\u6280\u90e8\u5c08\u984c 1.03 17.\u6cd5\u8981\u597d\u3002\u6240\u4ee5\u6211\u5011\u63d0\u51fa\u7684 DNN \u67b6\u69cb\u78ba\u5be6\u662f\u6709\u6548\u53ef\u884c\u7684\u3002 \u8a08\u756b(MOST 104-2221-E-027-079, 105-2221-E-027-119 and 103-2218-E-027-006-MY3)</td></tr></table>", | |
| "num": null, | |
| "html": null | |
| } | |
| } | |
| } | |
| } |