| { | |
| "paper_id": "O17-1003", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T07:59:29.967272Z" | |
| }, | |
| "title": "", | |
| "authors": [], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "", | |
| "pdf_parse": { | |
| "paper_id": "O17-1003", | |
| "_pdf_hash": "", | |
| "abstract": [], | |
| "body_text": [ | |
| { | |
| "text": "[2] 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.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| } | |
| ], | |
| "back_matter": [], | |
| "bib_entries": { | |
| "BIBREF0": { | |
| "ref_id": "b0", | |
| "title": "Music Speech Discrimination", | |
| "authors": [ | |
| { | |
| "first": "Yash", | |
| "middle": [], | |
| "last": "Malviya", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Shiv", | |
| "middle": [], | |
| "last": "Kaul", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Kushaagra", | |
| "middle": [], | |
| "last": "Goyal", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": null, | |
| "venue": "CS 229", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Yash Malviya, Shiv Kaul, Kushaagra Goyal, \"Music Speech Discrimination\",in CS 229", | |
| "links": null | |
| }, | |
| "BIBREF1": { | |
| "ref_id": "b1", | |
| "title": "Machine Learning Final Projects", | |
| "authors": [], | |
| "year": 2016, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Machine Learning Final Projects, Autumn 2016.", | |
| "links": null | |
| }, | |
| "BIBREF2": { | |
| "ref_id": "b2", | |
| "title": "Convolutional neural networks applied to house numbers digit classification", | |
| "authors": [ | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Sermanet", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Chintala", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Lecun", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2012, | |
| "venue": "International Conference on Pattern Recognition", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Sermanet, S. Chintala, and Y. LeCun. Convolutional neural networks applied to house numbers digit classification. In International Conference on Pattern Recognition (ICPR 2012), 2012.", | |
| "links": null | |
| }, | |
| "BIBREF4": { | |
| "ref_id": "b4", | |
| "title": "The Kaldi Speech Recognition Toolkit", | |
| "authors": [ | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Motlicek", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "P", | |
| "middle": [], | |
| "last": "Qian", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Schwarz", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Silovsky", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Stemmer", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Vesely", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2011, | |
| "venue": "IEEE 2011 Workshop on Automatic Speech Recognition and Understanding", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Motlicek, YM Qian, P. Schwarz, J. Silovsky, G. Stemmer, K. Vesely, \"The Kaldi Speech Recognition Toolkit,\" in IEEE 2011 Workshop on Automatic Speech Recognition and Understanding, Big Island, Hawaii, 2011.", | |
| "links": null | |
| }, | |
| "BIBREF5": { | |
| "ref_id": "b5", | |
| "title": "audio segmentation and classification in broadcast news domains", | |
| "authors": [ | |
| { | |
| "first": "Diego", | |
| "middle": [], | |
| "last": "Castan", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "David", | |
| "middle": [], | |
| "last": "Tavarez", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Paula", | |
| "middle": [], | |
| "last": "Lopez-Otero", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Javier", | |
| "middle": [], | |
| "last": "Franco-Pedroso", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Hector", | |
| "middle": [], | |
| "last": "Delgado", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Eva", | |
| "middle": [], | |
| "last": "Navas", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Laura", | |
| "middle": [], | |
| "last": "Docio-Fernandez", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Daniel", | |
| "middle": [], | |
| "last": "Ramos", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Javier", | |
| "middle": [], | |
| "last": "Serrano", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Alfonso", | |
| "middle": [], | |
| "last": "Ortega", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Eduardo", | |
| "middle": [], | |
| "last": "Lleida", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2015, | |
| "venue": "EURASIP Journal on Audio Speech and Music Processing", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Diego Castan, David Tavarez, Paula Lopez-Otero, Javier Franco-Pedroso, Hector Delgado,Eva Navas, Laura Docio-Fernandez, Daniel Ramos, Javier Serrano, Alfonso Ortega and Eduardo Lleida, \" audio segmentation and classification in broadcast news domains\",in EURASIP Journal on Audio Speech and Music Processing \u2022 December 2015.", | |
| "links": null | |
| }, | |
| "BIBREF6": { | |
| "ref_id": "b6", | |
| "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 | |
| }, | |
| "BIBREF7": { | |
| "ref_id": "b7", | |
| "title": "MUSAN: A Music, Speech, and Noise Corpus", | |
| "authors": [], | |
| "year": null, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "MUSAN: A Music, Speech, and Noise Corpus,https://scirate.com/arxiv/1510.08484.", | |
| "links": null | |
| } | |
| }, | |
| "ref_entries": { | |
| "TABREF0": { | |
| "text": "\u51fa\u4e4b CNN \u97f3\u8a0a\u5207\u5272\u67b6\u69cb\u78ba\u5be6\u53ef\u589e\u5f37\u6548\u679c\uff0c\u8a31\u591a\u6587\u737b\u4e2d\u4e5f\u4f7f\u7528\u53e6\u4e00\u7a2e ivector \u6c42\u53c3\u6578\uff0c\u5c07 \u4f86\u6703\u8003\u616e\u6539\u7528 ivector \u6c42\u53c3\u6578\uff0c\u8ddf AlexNet, VGG, ResNet \u7b49\uff0c\u505a\u8fd1\u4e00\u6b65\u5be6\u9a57\u6bd4\u8f03\u3002 \u8868 8 GMM \u548c CNN \u7cfb\u7d71\u4e4b\u6b63\u78ba\u7387\u8207 EER 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.", | |
| "html": null, | |
| "num": null, | |
| "type_str": "table", | |
| "content": "<table><tr><td>1. \u7c21\u4ecb \u4e86\u50b3\u7d71 MFCCs \u5916\uff0c\u9084\u6709\u54ea\u4e9b\u7279\u5fb5\u53c3\u6578\u5c0d\u97f3\u8a0a\u4e8b\u4ef6\u5075\u6e2c\u6548\u80fd\u6700\u597d\u3002 2.3. \u6a19\u8a18\u7d50\u679c\u7d71\u8a08 \u6bcf\u500b\u5377\u7a4d\u6838\u6703\u901a\u904e\u4e00\u6ed1\u52d5\u7a97\u53e3(sliding window) \uff0c\u6383\u63cf\u4e0a\u4e00\u7d1a\u8f38\u5165\u7684\u53c3\u6578\uff0c\u9010\u6b65\u8a08\u7b97 3.2.3. \u5168\u9023\u63a5\u5c64 \u9752\u5e74\u6545\u4e8b\u9928 252 34 55 206 0.72 212 \u8868 5 \u8a9e\u97f3/\u97f3\u6a02/\u5176\u5b83\u4e8b\u4ef6\u5075\u6e2c\u6e96\u78ba\u5ea6 4.4.3. \u5176\u5b83(\u7b11\u8072\u3001\u7279\u6548\u8072) \u4e8b\u4ef6\u5075\u6e2c \u6700\u5f8c\uff0c\u6211\u5011\u5373\u53ef\u4ee5\u5229\u7528\u9019\u500b\u6a94\u6848\uff0c\u9032\u4e00\u6b65\u7d44\u7e54\u4e26\u4e14\u5c0d\u5ee3\u64ad\u7bc0\u76ee\u5167\u5bb9\u9032\u884c\u505a\u52a0\u503c\u904b\u7528\u3002</td></tr><tr><td>\u904e\u53bb\u5728\u7db2\u8def(\u5c24\u5176\u662f\u884c\u52d5\u7db2\u8def)\u9084\u4e0d\u5920\u666e\u53ca\u7684\u6642\u5019\uff0c\u7528\u6536\u97f3\u6a5f\u6536\u807d\u5ee3\u64ad\u96fb\u81fa\u7bc0\u76ee\u662f \u4eba\u5011\u4e3b\u8981\u7684\u8cc7\u8a0a\u8207\u5a1b\u6a02\u4f86\u6e90\u3002\u4f46\u73fe\u4eca\u7db2\u8def\u5df2\u7d93\u975e\u5e38\u767c\u9054\uff0c\u8a31\u591a\u4eba\u7686\u6539\u6210\u76f4\u63a5\u7528\u624b\u6a5f\u89c0\u8cde \u96fb\u8996\u8207\u96fb\u5f71\u7b49\u8996\u983b\u985e\u7bc0\u76ee\u3002\u56e0\u6b64\u5927\u90e8\u5206\u7684\u5ee3\u64ad\u96fb\u81fa\u4e5f\u7a4d\u6975\u8f49\u578b\u56e0\u61c9\uff0c\u9664\u8a2d\u7acb\u7db2\u8def\u5ee3\u64ad\u96fb \u81fa\uff0c\u7dda\u4e0a\u5373\u6642(online)\u76f4\u64ad\u7bc0\u76ee\u5167\u5bb9\uff0c\u5438\u5f15\u5e74\u8f15\u807d\u773e\u6536\u807d\u5916\u3002\u66f4\u8a66\u8457\u628a\u64ad\u904e\u7684\u7bc0\u76ee\u5167 \u5bb9\uff0c\u96e2\u7dda(offline)\u653e\u5728\u7db2\u8def\u4e0a\uff0c\u5efa\u7acb\u96fb\u81fa\u7684\u5ee3\u64ad\u7bc0\u76ee\u5178\u85cf\u5eab(archive) \uff0c\u4f9b\u807d\u773e\u81ea\u7531\u5b89 \u6392\u6642\u9593\u6536\u807d\u4efb\u4f55\u7bc0\u76ee\uff0c\u4ee5\u7372\u53d6\u66f4\u591a\u5143\u7684\u807d\u773e\u4f86\u6e90\u3002 \u4f46\u662f\uff0c\u56e0\u50b3\u7d71\u4e0a\u5ee3\u64ad\u7bc0\u76ee\u662f\u807d\u89ba\u5a92\u9ad4\uff0c\u4e14\u5177\u6709\u50b3\u64ad\u901f\u5ea6\u5feb\uff0c\u6642\u6548\u6027\u5f37\u7b49\u7279\u6027\u3002\u76ee\u524d \u591a\u6578\u96fb\u81fa\u7684\u5ee3\u64ad\u7bc0\u76ee\u5178\u85cf\uff0c\u5927\u90fd\u53ea\u6709\u4fdd\u5b58\u7bc0\u76ee\u97f3\u6a94\u672c\u8eab\uff0c\u5c11\u6709\u4ed8\u52a0\u8207\u7bc0\u76ee\u5167\u5bb9\u76f8\u95dc\u7684\u5f8c \u8a2d\u8cc7\u6599(metadata)\u4ee5\u5229\u67e5\u8a62\u3002\u56e0\u6b64\u807d\u773e\u82e5\u60f3\u8981\u67e5\u8a62\u8207\u6aa2\u7d22\u6b77\u53f2\u7bc0\u76ee\u6642\uff0c\u5c31\u76f8\u7576\u4e0d\u5bb9\u6613\uff0c \u5c24\u5176\u662f\u5e38\u5e38\u7121\u6cd5\u76f4\u63a5\u627e\u5230\u807d\u773e\u6700\u6709\u8208\u8da3\u7684\u67d0\u7bc0\u76ee\u4e2d\u7684\u67d0\u6bb5\u91cd\u8981\u7684\u5167\u5bb9\u3002 \u56e0\u6b64\u672c\u8ad6\u6587\u5617\u8a66\u5075\u6e2c\u8207\u5207\u5272\u5ee3\u64ad\u7bc0\u76ee\u4e2d\u7684\u591a\u7a2e\u97f3\u8a0a\u4e8b\u4ef6[1][2]\uff0c\u56e0\u70ba\uff0c\u5ee3\u64ad\u96fb\u81fa\u7bc0 \u76ee\u4e2d\u901a\u5e38\u5305\u542b\u8a9e\u97f3\uff0c\u97f3\u6a02\u8207\u5176\u4ed6\u97f3\u8a0a\u4e8b\u4ef6(\u5982\u7b11\u8072\u6216\u7279\u6548\u8072) \u3002\u82e5\u80fd\u5075\u6e2c\u4e26\u5207\u5272\u9019\u4e9b\u97f3 \u8a0a\u4e8b\u4ef6[3]\uff0c\u5c31\u80fd\u9032\u4e00\u6b65\u7d44\u7e54\u5ee3\u64ad\u7bc0\u76ee\u5167\u5bb9\u9032\u884c\u9032\u4e00\u6b65\u52a0\u503c\u3002\u5c24\u5176\u662f\u82e5\u80fd\u81ea\u52d5\u8f49\u5beb\u6bcf\u4e00\u8a9e \u97f3\u7247\u6bb5\u7684\u9010\u5b57\u7a3f\uff0c\u64f7\u53d6\u51fa\u95dc\u9375\u5b57\u8207\u6458\u8981\uff0c\u6216\u662f\u81ea\u52d5\u8fa8\u8a8d\u51fa\u6bcf\u4e00\u97f3\u6a02\u7247\u6bb5\u7684\u6b4c\u540d\u6216\u66f2\u540d\uff0c \u5c31\u80fd\u8b93\u807d\u773e\u76f4\u63a5\u4ee5\u6587\u5b57\u9032\u884c\u5168\u6587\u6aa2\u7d22\uff0c\u627e\u5230\u76f8\u95dc\u7bc0\u76ee\u5167\u5bb9\uff0c\u6216\u662f\u4ee5\u54fc\u5531\u65b9\u5f0f\u627e\u5230\u60f3\u807d\u7684 \u97f3\u6a02\u6b4c\u66f2\u6bb5\u843d\u3002 \u5716 1 \u97f3\u8a0a\u4e8b\u4ef6\u5075\u6e2c\u7cfb\u7d71\u67b6\u69cb\u5716 \u5176\u4e2d\u5ee3\u64ad\u7bc0\u76ee\u97f3\u8a0a\u4e8b\u4ef6\u8cc7\u6599\u5eab\u5efa\u7acb\uff0c\u5c07 \u8490\u96c6\u4e26\u5c0d\u96fb\u81fa\u7bc0\u76ee\u505a\u5206\u985e\uff0c\u518d\u4ee5\u4eba\u5de5\u9032\u884c\u6a19 \u8a3b\uff0c\u627e\u51fa\u8a9e\u97f3\uff0c\u97f3\u6a02\u8207\u5176\u4ed6\u97f3\u8a0a\u4e8b\u4ef6\u7684\u8d77\u59cb\u8207\u7d50\u675f\u6642\u9593\u3002\u97f3\u8a0a\u4e8b\u4ef6\u6a21\u578b\u8a13\u7df4\uff0c\u5247\u662f\u5229\u7528 \u4eba\u5de5\u5207\u5272\u51fa\u4e4b\u4e0d\u540c\u97f3\u8a0a\u4e8b\u4ef6\u7684\u6a23\u672c\u8207\u7b54\u6848\u96c6\uff0c\u5206\u5225\u5efa\u7acb\u8a9e\u97f3\u8207\u975e\u8a9e\u97f3\uff0c\u97f3\u6a02\u8207\u975e\u97f3\u6a02\uff0c \u5176\u4ed6\u8207\u975e\u5176\u4ed6\u7b49\u97f3\u8a0a\u4e8b\u4ef6\u7684\u6a21\u578b\u3002\u6700\u5f8c\u81ea\u52d5\u5ee3\u64ad\u7bc0\u76ee\u97f3\u8a0a\u4e8b\u4ef6\u5075\u6e2c\u8207\u6a19\u8a18\uff0c\u5c31\u662f\u5229\u7528\u6240 \u8a13\u7df4\u51fa\u4e4b\u4e09\u7a2e\u97f3\u8a0a\u4e8b\u4ef6\u6a21\u578b\uff0c\u5075\u6e2c\u8f38\u5165\u4e4b\u5ee3\u64ad\u7bc0\u76ee\u4e2d\u7684\u5404\u7a2e\u97f3\u8a0a\u4e8b\u4ef6\uff0c\u4e26\u5c07\u5176\u8f49\u6210\u8207\u7bc0 \u76ee\u97f3\u6a94\u76f8\u5c0d\u61c9\uff0c\u5177\u6642\u9593\u8cc7\u8a0a\u7684\u97f3\u8a0a\u4e8b\u4ef6\u6a19\u6ce8\u6a94\u3002 \u8868 2 \u5247\u662f\u9032\u884c\u4eba\u5de5\u6a19\u8a18\u5f8c\uff0c\u5404\u7a2e\u6a19\u8a18\u51fa\u7684\u97f3\u8a0a\u4e8b\u4ef6\u7684\u6642\u9593\u9577\u5ea6\u7d71\u8a08\u8cc7\u6599\u3002\u5176\u4e2d\uff0c\u8a9e \u5176\u8207\u5377\u7a4d\u6838\u7684\u5167\u7a4d\uff0c\u8f38\u51fa\u4e00\u5377\u7a4d\u7279\u5fb5\u53c3\u6578\u5716(feature map) \u3002\u56e0\u6b64\uff0c\u4e00\u500b\u5377\u7a4d\u6838\u5c31\u76f8\u7576 \u6700\u5f8c\u5247\u662f\u5229\u7528\u5168\u9023\u63a5\u5c64\u8fa8\u8a8d\u76ee\u6a19\u8207\u975e\u76ee\u6a19\u4e8b\u4ef6\u3002\u5168\u9023\u63a5\u5c64\u7684\u67b6\u69cb\u5c31\u662f\u4e00\u500b\u50b3\u7d71\u7684 Total 1129 142 359 866 4.35 1074 Accuracy(%) Speech Music Other \u5716 16 \u70ba\u7d93\u5176\u4ed6\u4e8b\u4ef6\u5075\u6e2c\u5668\u8655\u7406\u904e\u7684\u7d50\u679c\u3002\u7bc4\u4f8b\u5716\u4e2d\u53ef\u4ee5\u770b\u5230\u7cfb\u7d71\u7684\u78ba\u80fd\u5920\u6b63\u78ba\u5075 \u5c24\u5176\u662f\u81ea\u52d5\u8f49\u5beb\u6bcf\u4e00\u8a9e\u97f3\u7247\u6bb5\u7684\u9010\u5b57\u7a3f\uff0c\u64f7\u53d6\u51fa\u95dc\u9375\u5b57\u8207\u6458\u8981\uff0c\u6216\u662f\u81ea\u52d5\u8fa8\u8a8d\u51fa\u6bcf\u4e00\u97f3 2. \u5ee3\u64ad\u7bc0\u76ee\u97f3\u8a0a\u4e8b\u4ef6\u8cc7\u6599\u5eab \u97f3\u4e8b\u4ef6\u6700\u591a\uff0c\u7e3d\u6578\u7d04\u70ba 3000 \u5206\u9418\uff0c\u97f3\u6a02\u4e8b\u4ef6\u6b21\u4e4b\uff0c\u7d04\u6709 700 \u5206\u9418\uff0c\u5176\u4ed6\u4e8b\u4ef6\u6700\u5c11\uff0c\u53ea \u65bc\u4e00\u500b\u591a\u7dad\u5ea6\u5339\u914d\u6ffe\u6ce2\u5668\uff0c\u4f46\u5176\u6b0a\u91cd\u53ef\u7d93\u7531\u8a13\u7df4\u81ea\u52d5\u6700\u4f73\u5316\u3002 \u591a\u5c64\u5f0f\u985e\u795e\u7d93\u7db2\u8def(\u901a\u5e38\u53ea\u7528\u5169\u5c64) \uff0c\u4f46\u6700\u5f8c\u662f\u4ee5 softmax \u6fc0\u767c\u51fd\u6578\uff0c\u8a08\u7b97\u76ee\u6a19\u4e8b\u4ef6\u8207 4.2. CNN \u8a2d\u5b9a GMM 97.12 94.88 94.15 CNN \u6e2c\u51fa\u7b11\u8072\u4e8b\u4ef6\u3002 \u6a02\u7247\u6bb5\u7684\u6b4c\u540d\u6216\u66f2\u540d\u3002\u8b93\u807d\u773e\u76f4\u63a5\u4ee5\u6587\u5b57\u9032\u884c\u5168\u6587\u6aa2\u7d22\uff0c\u627e\u5230\u76f8\u95dc\u7bc0\u76ee\u5167\u5bb9\uff0c\u6216\u662f\u4ee5\u54fc 98.46 96.43 97.47 2.1. \u5ee3\u64ad\u7bc0\u76ee\u8cc7\u6599\u641c\u96c6\u8207\u8a2d\u8a08 \u6709\u7d04 25 \u5206\u9418\u3002\u56e0\u6b64\u8cc7\u6599\u5206\u4f48\u76f8\u7576\u4e0d\u5e73\u8861\u3002 \u5716 4 \u9078\u7528\u4e0d\u540c\u7684\u7279\u5fb5\u53c3\u6578\u793a\u610f\u5716 \u975e\u76ee\u6a19\u4e8b\u4ef6\u767c\u751f\u7684\u6a5f\u7387\u503c\uff0c\u7528\u4ee5\u6aa2\u6e2c\u662f\u5426\u6709\u76ee\u6a19\u97f3\u8a0a\u4e8b\u4ef6\u767c\u751f\u3002 \u5531\u65b9\u5f0f\u627e\u5230\u60f3\u807d\u7684\u97f3\u6a02\u6b4c\u66f2\u6bb5\u843d\u3002 \u4e3b\u8981\u662f\u9700\u8981\u8003\u616e\u4f7f\u7528\u7684\u7279\u5fb5\u53c3\u6578\u3001CNN \u7db2\u8def\u7d50\u69cb\u5927\u5c0f\u8207\u9078\u64c7\u9069\u7576\u7684\u512a\u5316\u5668\u3002 \u6211\u5011\u9996\u5148\u5c07\u5ee3\u64ad\u7bc0\u76ee\u985e\u578b\u5206\u70ba\u7d14\u8a9e\u97f3\u3001\u8a9e\u97f3+\u97f3\u6a02\u3001\u8a9e\u97f3+\u8f03\u591a\u97f3\u6a02\u4e09\u5927\u985e[5]\uff0c\u6bcf\u4e00 \u8868 2 \u4eba\u5de5\u6a19\u8a18\u97f3\u8a0a\u5206\u985e\u6642\u9593\u7e3d\u8868(minute) 3.2. \u5377\u7a4d\u985e\u795e\u7d93\u7db2\u8def\u6a21\u578b\u67b6\u69cb 4. \u97f3\u8a0a\u4e8b\u4ef6\u5075\u6e2c\u5be6\u9a57 4.2.1. \u7279\u5fb5\u53c3\u6578\u8a2d\u5b9a\uff1a 4.3.2. \u5be6\u9a57\u4e8c-\u57fa\u65bc\u4e0d\u540c\u53c3\u6578\u4e4b CNN \u6548\u80fd\u6e2c\u8a66 6. \u7d50\u8ad6 \u985e\u7bc0\u76ee\u5404\u6311\u9078\u591a\u96c6\u7bc0\u76ee\u3002\u8868 1 \u70ba\u5ee3\u64ad\u7bc0\u76ee\u985e\u578b\u5206\u985e\uff0c\u8207\u6311\u9078\u51fa\u7684\u7bc0\u76ee\u8207\u5176\u9577\u5ea6\u3002 \u8868 1 \u5ee3\u64ad\u8a9e\u6599\u5eab\u7d71\u8a08\u8cc7\u6599 \u985e\u578b \u7bc0\u76ee\u540d\u7a31 \u96c6\u6578 \u6311\u9078\u6642\u9577 (minute) \u985e\u578b \u7bc0\u76ee\u540d\u7a31 \u96c6\u6578 \u6311\u9078\u6642\u9577 (minute) \u7d14\u8a9e\u97f3 \u5275\u9752\u5b85\u6025\u4fbf 10 384 \u8a9e\u97f3+\u97f3\u6a02 \u5275\u8a2d\u5e02\u96c6 On-Air 10 267 \u81ea\u7136\u6709\u610f\u601d 8 157 \u6559\u80b2\u958b\u8b1b 10 261 \u79d1\u5b78 SoEasy 10 261 \u4eca\u5929\u4e0d\u88dc\u7fd2 10 298 \u7279\u5225\u7684\u611b 10 270 \u5152\u7ae5\u65b0\u805e 10 98 \u591a\u611b\u81ea\u5df1\u4e00\u9ede\u9ede 10 289 \u6587\u6559\u65b0\u805e 10 68 \u570b\u969b\u6559\u80b2\u5fc3\u52d5\u7dda 10 210 \u8a9e\u97f3+\u8f03\u591a\u97f3\u6a02 \u5f9e\u5fc3\u6b78\u96f6 10 540 \u9752\u5e74\u6545\u4e8b\u9928 10 286 \u5716 2 \u97f3\u8a0a\u4e8b\u4ef6\u4eba\u5de5\u6a19\u8a8c\u7684\u793a\u610f\u5716 \u6a19\u6ce8\u7a0b\u5e8f\uff0c\u5247\u662f\u5229\u7528 Praat \u8edf\u9ad4\uff0c\u5148\u5efa\u7acb\u4e09\u8ecc\u6a19\u8a3b\u9762\u677f\uff0c\u518d\u4f9d\u7167\u4e0b\u5217\u898f\u5247\u6a19\u8a18\u3002( 1) \u8a9e\u97f3\u90e8\u5206\uff1a\u53ea\u8981\u97f3\u6a94\u4e2d\u6709\u4eba\u8b1b\u8a71\u7684\u90e8\u5206\uff0c\u4e14\u807d\u5f97\u51fa\u8b1b\u8a71\u5167\u5bb9\uff0c\u7686\u6a19\u8a18\u6210\u70ba\u8a9e\u97f3(\u5982\uff1a\u4e3b \u6301\u4eba\u6216\u4f86\u8cd3\u8b1b\u8a71\u6216\u662f Call in \u7684\u6c11\u773e)\u3002( 2)\u97f3\u6a02\u90e8\u5206\uff1a\u53ea\u8981\u97f3\u6a94\u6709\u97f3\u6a02\uff0c\u4e14\u807d\u5f97\u51fa\u97f3\u6a02 \u5167\u5bb9\u7684\u8a71\uff0c\u7686\u6a19\u70ba\u97f3\u6a02(\u5982\uff1a\u9ad8\u4e2d\u751f\u5408\u5531\u5718\u6f14\u5531\u3001\u6a02\u5718\u8868\u6f14\u4ee5\u53ca\u6d41\u884c\u97f3\u6a02\u2026\u7b49)\u3002( 3)\u5176 Program speech non-speech music non-music other non-other \u5275\u8a2d\u5e02\u96c6 On-Air 222 45 84 180 0.86 264 \u6559\u80b2\u958b\u8b1b 234 27 50 17 0.66 264 \u4eca\u5929\u4e0d\u88dc\u7fd2 246 52 63 6 2 294 \u5152\u7ae5\u65b0\u805e 92 6 7 92 0.4 99 \u6280\u8077\u6700\u524d\u7dda 186 27 72 114 3 150 \u5f9e\u5fc3\u6b78\u96f6 372 168 52 342 7 354 \u7279\u5225\u7684\u611b 270 0 0 300 5 294 \u5275\u9752\u5b85\u6025\u4fbf 384 0 0 384 1 384 \u81ea\u7136\u6709\u610f\u601d 156 1 1 138 1 156 \u79d1\u5b78 SoEasy 246 15 60 198 0.78 204 \u6587\u6559\u65b0\u805e 61 7 9 60 0.66 68 \u50b3\u7d71\u985e\u795e\u7d93\u7db2\u8def\u6a21\u578b\uff0c\u90fd\u662f\u5148\u4ee5\u4eba\u70ba\u65b9\u5f0f\uff0c\u8a2d\u8a08\u597d\u8981\u4f7f\u7528\u7684\u7279\u5fb5\u53c3\u6578\uff0c\u7136\u5f8c\u76f4\u63a5\u63a1 \u7528\u5168\u9023\u63a5\u7684 deep neural networks(DNNs)\u8a13\u7df4\u6a21\u578b\u3002\u4f46\u7531\u65bc\u6211\u5011\u4e0d\u80fd\u78ba\u5b9a\u54ea\u4e00\u7a2e\u7279\u5fb5 \u53c3\u6578\uff0c\u80fd\u5920\u6709\u6700\u597d\u7684\u97f3\u8a0a\u4e8b\u4ef6\u5075\u6e2c\u6548\u679c\uff0c\u6240\u4ee5\u672c\u8ad6\u6587\u6539\u4f7f\u7528\u5377\u7a4d\u795e\u7d93\u7db2\u7d61 (Convolutional Neural Networks, CNNs)\u67b6\u69cb\u4f86\u505a\u97f3\u8a0a\u5075\u6e2c\u5207\u5272\u7cfb\u7d71\uff0c\u5176\u67b6\u69cb\u5982\u5716 5\u3002\u6b64\u5916\uff0c\u7531\u65bc\u4e0d\u540c \u97f3\u8a0a\u4e8b\u4ef6\u53ef\u80fd\u6703\u540c\u6642\u767c\u751f\uff0c\u6240\u4ee5\u6bcf\u7a2e\u97f3\u8a0a\u4e8b\u4ef6(speech\u3001music and other)\uff0c\u90fd\u9700\u8981\u7368\u7acb\u5efa \u7acb\u4e00\u500b\u6a21\u578b\uff0c\u7136\u5f8c\u500b\u5225\u904b\u4f5c\u505a\u5075\u6e2c\u3002 \u5716 6 CNN \u5377\u7a4d\u5c64\u904b\u4f5c\u65b9\u5f0f 3.2.2. \u6c60\u5316\u5c64 \u6c60\u5316\u5c64(Pooling)[6]\u63a5\u5728\u5377\u7a4d\u5c64\u4e4b\u5f8c\uff0c\u5176\u904b\u4f5c\u5982\u5716 7 \u6240\u793a\u3002\u4e3b\u8981\u662f\u5c07 feature map 4.1. \u8a13\u7df4\u8207\u6e2c\u8a66\u8a9e\u6599 \u672c\u5be6\u9a57\u5c07\u6559\u80b2\u96fb\u81fa\u7bc0\u76ee\u5206\u6210\u7d14\u8a9e\u97f3\u3001\u8a9e\u97f3+\u97f3\u6a02\u3001\u8a9e\u97f3+\u8f03\u591a\u97f3\u6a02\u4e09\u7a2e\u985e\u578b\uff0c\u5171\u64f7\u53d6 14 \u500b\u7bc0\u76ee(\u9577\u5ea6\u5171\u7d04 60 \u5c0f\u6642) \uff0c\u7d93\u4eba\u5de5\u6a19\u6ce8\u5f8c\uff0c\u5c07\u5176\u5206\u6210\u8a13\u7df4\u7528\u8207\u6e2c\u8a66\u7528\u5169\u7d44\uff0c\u7528\u4f86\u6bd4 \u8f03 CNN \u7684\u6548\u80fd\u3002\u8a9e\u6599\u5167\u5bb9\u6982\u6cc1\u5982\u8868 3\u3001\u8868 4 \u6240\u793a\u3002\u6b64\u5916\u70ba\u4e86\u5e73\u8861\u4e0d\u540c\u4e8b\u4ef6\u7684\u6a23\u672c\u6578\u91cf\uff0c \u6211\u5011\u5728\u8a13\u7df4\u8a9e\u6599\u7d44\uff0c\u984d\u5916\u518d\u52a0\u4e0a MUSAN[7]\u8a9e\u6599\uff0c\u4ee5\u589e\u52a0\u97f3\u6a02\u8207\u5176\u4ed6\u985e\u5225\u4e8b\u4ef6\u6a23\u672c\u7684\u6578 \u5f9e\u4e0a\u4e00\u500b\u5be6\u9a57\uff0c\u6211\u5011\u5df2\u7d93\u9a57\u8b49\u4e86\uff0c\u5728\u97f3\u8a0a\u4e8b\u4ef6\u6a21\u578b\u88e1\uff0cCNN \u67b6\u69cb\u6703\u6bd4 GMM \u67b6\u69cb \u5728\u7279\u5fb5\u53c3\u6578\u65b9\u9762\uff0c\u6211\u5011\u53d6\u4e86\u4e09\u7a2e\u53c3\u6578\u9032\u884c\u6bd4\u8f03\uff0c\u5305\u62ec(1)MFCCs\uff0c (2)Mel-spectrum \u4f86\u5f97\u6709\u6548\u3002\u5728\u6b64\uff0c\u6211\u5011\u9032\u4e00\u6b65\u6e2c\u8a66\u4e0d\u540c\u53c3\u6578\u7684\u6548\u80fd\u3002\u5be6\u9a57\u7d50\u679c\u5982\u8868 6 \u6240\u793a\uff0c\u53ef\u4ee5\u770b\u51fa\u76f4 \u5716 10 \u97f3\u6a02\u8a13\u7df4\u8cc7\u6599\u7d50\u679c \u672c\u8ad6\u6587\u5148\u5efa\u7acb\u4eba\u5de5\u6a19\u8a18\u4e4b\u5ee3\u64ad\u7bc0\u76ee\u97f3\u8a0a\u4e8b\u4ef6\u8cc7\u6599\u5eab\uff0c\u518d\u4f7f\u7528 CNN \u5be6\u4f5c\u97f3\u8a0a\u5075\u6e2c\u5207 \u5716 11 \u97f3\u6a02\u6e2c\u8a66\u8cc7\u6599\u7d50\u679c (Mels)\u8207(3)Raw Spectrum (Specgram)\u3002 4.2.2. \u5272\u7cfb\u7d71\uff0c\u4e26\u76f4\u63a5\u4f7f\u7528\u983b\u8b5c\uff0c\u907f\u514d\u53c3\u6578\u8a2d\u8a08\u5de5\u7a0b\u554f\u984c\u3002\u6574\u9ad4\u5be6\u9a57\u7d50\u679c\u986f\u793a\uff0c\u5982\u8868 8 \u6240\u793a\uff0c \u63a5\u8f38\u5165 raw spectrum\uff0c\u5c31\u53ef\u4ee5\u5f97\u5230\u6700\u4f73\u7684\u8fa8\u8a8d\u7d50\u679c\u3002 \u7db2\u8def\u914d\u7f6e\uff1a \u4ee5 CNN \u76f4\u63a5\u642d\u914d\u983b\u8b5c\u53c3\u6578\uff0c\u5728\u5075\u6e2c\u8a9e\u97f3\u8207\u975e\u8a9e\u97f3\uff0c\u97f3\u6a02\u8207\u975e\u97f3\u6a02\u6216\u5176\u5b83\u8207\u975e\u5176\u5b83\u97f3\u8a0a \u8868 6 \u4e0d\u540c\u7279\u5fb5\u53c3\u6578\u7684\u97f3\u8a0a\u4e8b\u4ef6\u5075\u6e2c\u6e96\u78ba\u5ea6 \u5728 batch size \u65b9\u9762\uff0c\u6211\u5011\u6e2c\u8a66\u4e86 16\u300132\u300164\u3001128 \u7b46\u6a23\u672c\u56db\u7a2e\u8b8a\u5316\uff0c\u6700\u5f8c\u5c07 batch size \u4e8b\u4ef6\u7b49\u7684\u932f\u8aa4\u7387 EER\uff0c\u5206\u5225\u70ba 2.27%\u300112.52%\u8207 9.51%\uff0c\u7686\u4f4e\u65bc\u50b3\u7d71\u4ee5 GMM \u642d\u914d Mel-Accuracy(%) MFCCs Mels Specgram \u8a2d\u6210 64 \u7b46\u3002\u6211\u5011\u4e5f\u5728\u53c3\u6578\u8f38\u5165\u5c64\u52a0\u5165 dropout\uff0c\u5617\u8a66\u904e\u4e1f\u68c4 2%\u30015%\u8207 10%\u8f38\u5165\u53c3\u6578\u7b49 \u8b8a\u5316\uff0c\u6700\u5f8c\u8a2d\u5b9a\u70ba 2%\uff0c\u5728\u5377\u7a4d\u5c64\u7684 dropout \u5247\u662f\u8a2d\u5b9a\u70ba 25%\u3002\u6700\u5f8c\u5728\u5c64\u6578\u65b9\u9762\uff0c\u6211\u5011 Speech 97.96 97.87 98.46 Music 95.23 96.43 Frequency Cepstral Coefficients(MFCCs)\u7684 3.65%\u300115.68%\u8207 13.25%\u3002\u56e0\u6b64\u672c\u8ad6\u6587\u63d0 \u5716 14 \u8a9e\u97f3\u5075\u6e2c\u7d50\u679c 96.28 \u91cf\u3002 \u8868 3 \u6559\u80b2\u96fb\u81fa\u97f3\u8a0a\u4e8b\u4ef6\u8a13\u7df4\u8a9e\u6599 (minute) Program Other 96.53 96.47 96.22 4.4.2. \u97f3\u6a02\u4e8b\u4ef6\u5075\u6e2c \u6e2c\u8a66\u4e86 2 layers\u30013 layers\u30014 layers \u8207 5 layers\uff0c\u6700\u5f8c\u7686\u8a2d\u6210 4 layers\u3002 4.2.3. Average 96.57 96.92 96.99 \u5f9e\u5716 15 \u53ef\u4ee5\u770b\u5230\u7d93\u97f3\u6a02\u4e8b\u4ef6\u5075\u6e2c\u5668\u8655\u7406\u904e\u7684\u7d50\u679c\u7bc4\u4f8b\u5716\u3002\u5f9e\u5716\u4e2d\u53ef\u77e5\u9053\u975e\u97f3\u6a02\u90e8 \u9078\u64c7\u512a\u5316\u5668\uff1a 4.3.3. \u5be6\u9a57\u4e09-EER \u7d50\u679c speech non-speech music non-music other non-other \u6280\u8077\u6700\u524d\u7dda 10 213 \u5404\u6311\u9078\u591a\u96c6\u7bc0\u76ee\u539f\u56e0\u5728\u65bc\uff0c\u6559\u80b2\u5ee3\u64ad\u96fb\u81fa\u7684\u7bc0\u76ee\u76f8\u7576\u591a\u5143\uff0c\u97f3\u6a94\u53ef\u80fd\u5305\u542b\u8aaa\u8a71\u8072\u3001 \u97f3\u6a02\u548c\u7279\u6548\u7b49\u7b49\u3002\u70ba\u4e86\u80fd\u5920\u8b93\u6240\u6709\u60c5\u6cc1\u90fd\u80fd\u5920\u6536\u96c6\u5230\uff0c\u56e0\u6b64\u6211\u5011\u9700\u8981\u62ff\u591a\u7a2e\u4e0d\u540c\u985e\u578b\u7684 \u7bc0\u76ee\u9032\u884c\u6311\u9078\uff0c\u624d\u80fd\u6db5\u84cb\u6240\u6709\u53ef\u80fd\u60c5\u6cc1\u3002 2.2. \u97f3\u8a0a\u4e8b\u4ef6\u4eba\u5de5\u6a19\u8a18\u898f\u7bc4 \u5b83\u90e8\u5206\uff1a\u5247\u662f\u5728\u97f3\u6a94\u5167\u5bb9\u51fa\u73fe\u7b11\u8072\u6216\u7279\u6548\u8072\uff0c\u7686\u6a19\u793a\u70ba\u5176\u5b83\u7684\u90e8\u5206\u3002\u5716 3 \u70ba\u5be6\u969b\u4eba\u5de5\u6a19 \u8a3b\u7d50\u679c\u7684\u7bc4\u4f8b\u3002 \u591a\u611b\u81ea\u5df1\u4e00\u9ede\u9ede 204 85 234 54 0.86 224 \u570b\u969b\u6559\u80b2\u5fc3\u52d5\u7dda 210 0 0 210 0.33 210 \u9752\u5e74\u6545\u4e8b\u9928 252 34 55 206 0.72 212 Total 3135 467 687 2301 25 3177 3. \u57fa\u65bc CNN \u4e4b\u97f3\u8a0a\u4e8b\u4ef6\u5075\u6e2c\u7cfb\u7d71 \u5283\u5206\u6210\u6578\u500b\u5340\u57df\uff0c\u4e26\u4ee5\u5340\u57df\u70ba\u55ae\u4f4d\uff0c\u5728\u6bcf\u4e00\u5340\u57df\u4ee5\u985e\u4f3c\u6295\u7968\u65b9\u5f0f\uff0c\u53ea\u9078\u53d6\u6b64\u5340\u57df\u4e2d\u8f03 \u5f37\u7684\u5377\u7a4d\u503c\u505a\u8f38\u51fa\uff0c\u4e26\u4e1f\u6389\u8f03\u5fae\u5f31\u7684\u5377\u7a4d\u503c\u3002\u6b64\u904b\u4f5c\u9664\u53ef\u964d\u4f4e\u6578\u64da\u91cf\u3001\u6e1b\u5c0f\u904e\u64ec\u5408\uff0c \u6700\u91cd\u8981\u7684\u662f\uff0c\u53ef\u4ee5\u5bb9\u5fcd\u76ee\u6a19\u4e8b\u4ef6\u5728\u983b\u8b5c\u4e0a\u7684\u4f4d\u7f6e\u8b8a\u7570\u3002 \u5275\u8a2d\u5e02\u96c6 On-Air 222 45 84 180 0.86 264 \u6559\u80b2\u958b\u8b1b 234 27 50 17 0.66 \u5716 12 \u5176\u5b83\u8a13\u7df4\u8cc7\u6599\u7d50\u679c \u5716 13 \u5176\u5b83\u6e2c\u8a66\u8cc7\u6599\u7d50\u679c \u5206\u53ef\u4ee5\u660e\u986f\u88ab\u62ff\u6389\uff0c\u53ea\u4fdd\u7559\u97f3\u6a02\u4e8b\u4ef6\u7684\u6bb5\u843d\u3002\u4e0d\u904e\u6bd4\u8f03\u5716 14 \u8207\u5716 15\uff0c\u53ef\u77e5\u5728\u908a\u754c\u7684\u5730 \u5716 16 \u5176\u5b83\u5075\u6e2c\u7d50\u679c \u5be6\u9a57\u4e2d\u5617\u8a66\u5169\u7a2e\u512a\u5316\u5668\u4f86\u505a\u6e2c\u8a66\uff0c\u4e00\u500b\u662f SGD\uff0c\u53e6\u4e00\u500b\u662f Adadelta\u3002SGD \u662f\u6307 \u5716 8 \u81f3\u5716 13 \u5206\u5225\u70ba\u6240\u8a13\u7df4\u597d\u7684\u8a9e\u97f3\u3001\u97f3\u6a02\u3001\u5176\u4ed6\u97f3\u8a0a\u4e8b\u4ef6\u5075\u6e2c\u5668\uff0c\u5c0d\u8a13\u7df4\u8cc7\u6599\u548c 264 \u4eca\u5929\u4e0d\u88dc\u7fd2 246 52 63 6 2 gradient descent\uff0c\u662f\u6700\u5e38\u898b\u7684\u512a\u5316\u65b9\u6cd5\uff0c\u4f46\u5176\u4e0d\u6703\u81ea\u52d5\u8abf\u6574\u5b78\u7fd2\u7387\uff0c\u9808\u81ea\u884c\u5617\u8a66\u3002\u800c \u5c0d\u6e2c\u8a66\u8cc7\u6599\u8a08\u7b97\u6aa2\u6e2c\u932f\u8aa4\u6b0a\u8861\u66f2\u7dda\u7684\u7d50\u679c\u3002\u5176\u4e2d y \u8ef8\u70ba\u932f\u8aa4\u62d2\u7d55\u7387\uff0cx \u8ef8\u70ba\u932f\u8aa4\u63a5\u53d7\u7387\u3002 \u65b9\u8f03\u5bb9\u6613\u767c\u751f\u932f\u8aa4\uff0c\u5c24\u5176\u8a9e\u97f3\u5bb9\u6613\u88ab\u5224\u5225\u6210\u97f3\u6a02\u3002 5. \u81ea\u52d5\u97f3\u8a0a\u4e8b\u4ef6\u6a19\u8a18\u61c9\u7528 GMM CNN 294 \u5152\u7ae5\u65b0\u805e 92 6 7 92 0.4 99 \u6280\u8077\u6700\u524d\u7dda 186 27 72 114 3 Adadelta \u5247\u6703\u5c0d\u5b78\u7fd2\u7387\u9032\u884c\u81ea\u9069\u61c9\u7d04\u675f\uff0c\u4f7f\u7528\u8d77\u4f86\u5fc5\u8f03\u65b9\u4fbf\u3002\u56e0\u6b64\u6700\u5f8c\u6211\u5011\u9078\u7528 Adadelta \u7531\u8868 7EER \u7684\u7d50\u679c\u4e2d\uff0c\u53ef\u77e5\u8a9e\u97f3\u4e8b\u4ef6\u5075\u6e2c\u7684 EER \u6700\u4f4e\uff0c\u800c\u97f3\u6a02\u8207\u5176\u4ed6\u4e8b\u4ef6\u7684 EER \u90fd\u8f03 \u8868 7 EER \u7d50\u679c EER(%) Train Test Audio event Accuracy(%) EER(%) Accuracy(%) EER(%) \u672c\u8ad6\u6587\u6700\u5f8c\u4f7f\u7528\u524d\u9762\u8a13\u7df4\u51fa\u4f86\u7684\u4e09\u7a2e\u97f3\u8a0a\u4e8b\u4ef6\u5075\u6e2c\u5668\uff0c\u5c07\u5075\u6e2c\u5230\u7684\u97f3\u8a0a\u4e8b\u4ef6\u7684\u8d77 Speech 97.12 3.65 98.46 2.27 150 \u5f9e\u5fc3\u6b78\u96f6 372 168 52 342 7 354 \u512a\u5316\u5668\u3002 \u9ad8\u3002\u9019\u53ef\u80fd\u662f\u97f3\u6a02\u8207\u5176\u4ed6\u4e8b\u4ef6\u7684\u8b8a\u5316\u8f03\u591a\uff0c\u5728\u8a13\u7df4\u8a9e\u6599\u4e2d\u7684\u6a23\u672c\u6db5\u84cb\u7387\uff0c\u9084\u662f\u6bd4\u8f03\u4e0d\u8db3\u3002 Speech 0.91 2.27 \u59cb\u8207\u7d50\u675f\u6642\u9593\u627e\u51fa\u4f86\u3002\u5728\u5408\u4f75\u9019\u4e9b\u8cc7\u8a0a\u505a\u51fa\u5982\u5716 17\uff0c\u985e\u4f3c\u5f71\u7247\u5b57\u5e55\u5177\u6709 timing code \u683c Music 94.88 15.68 96.43 12.52</td></tr><tr><td>\u91dd\u5c0d\u6b64\u554f\u984c\uff0c\u672c\u8ad6\u6587\u7684\u6574\u9ad4\u8655\u7406\u7a0b\u5e8f\u5982\u5716 1 \u6240\u793a\uff0c\u4e3b\u8981\u7684\u60f3\u6cd5\u5305\u542b(1)\u5efa\u7acb\u5ee3\u64ad \u7bc0\u76ee\u97f3\u8a0a\u4e8b\u4ef6\u8cc7\u6599\u5eab\uff0c\u5982\u5716 1(a) \uff1b (2)\u8a13\u7df4\u97f3\u8a0a\u4e8b\u4ef6\u6a21\u578b\uff0c\u5982\u5716 1(b) \uff1b (3)\u9032\u884c\u81ea \u52d5\u5ee3\u64ad\u7bc0\u76ee\u97f3\u8a0a\u4e8b\u4ef6\u5075\u6e2c\u8207\u6a19\u8a18\uff0c\u5982\u5716 1(c) \u3002 \u672c\u8ad6\u6587\u5c07\u4f7f\u7528 Convolutional Neural Network (CNN)[4]\u67b6\u69cb\u4f86\u5b8c\u6210\u6211\u5011\u7684\u97f3\u8a0a\u5075\u6e2c \u5207\u5272\u7cfb\u7d71\uff0c\u4e00\u822c\u97f3\u8a0a\u4e8b\u4ef6\u5075\u6e2c\u7cfb\u7d71\u7af6\u8cfd\u90fd\u662f\u4ee5\u50b3\u7d71 GMM+MFCCs \u67b6\u69cb\u4f86\u8a13\u7df4\uff0c\u5982 DCASE2016 Challenge[5]\uff0c\u6211\u5011\u5c07\u4f7f\u7528 CNN \u67b6\u69cb\u8207\u5178\u578b\u7684 i-vector \u7cfb\u7d71 Kaldi speech recognition toolkit[6]\u4e2d Bn_music_speech \u63d0\u4f9b\u7684\u4e00\u500b GMM \u57fa\u790e\u4f5c\u6cd5\u505a\u6bd4\u8f03\uff0c\u4e3b\u8981\u662f\u56e0 \u70ba CNN \u5177\u6709\u4ee5\u4e0b\u7279\u6027\uff1a (1)\u5c0d\u97f3\u8a0a\u4e8b\u4ef6\u5728\u8f38\u5165\u53c3\u6578\u5e8f\u5217\u4e2d\u7684\u4f4d\u7f6e\uff0c\u5177\u6709\u6642\u9593\u8207\u983b\u5e36\u4e0a \u7684\u5e73\u79fb\u4e0d\u8b8a\u6027\uff0c\u53ef\u4ee5\u5bb9\u5fcd\u97f3\u8a0a\u4e8b\u4ef6\u5728\u6642\u9593\u8207\u983b\u8b5c\u4e0a\u7684\u8b8a\u7570\u3001 (2)\u80fd\u81ea\u6211\u8a13\u7df4\u5982\u4f55\u64f7\u53d6\u6700 \u4f73\u5316\u7684\u97f3\u8a0a\u4e8b\u4ef6\u7279\u5fb5\u53c3\u6578\uff0c\u56e0\u6b64\u53ef\u4ee5\u907f\u514d\u9700\u5c08\u696d\u77e5\u8b58\uff0c\u624d\u80fd\u8a2d\u8a08\u51fa\u9069\u5408\u7684\u97f3\u8a0a\u53c3\u6578\u7684\u53c3 \u6578\u5de5\u7a0b(feature engineering)\u554f\u984c\u3002\u800c\u80fd\u76f4\u63a5\u8f38\u5165\u983b\u8b5c\u53c3\u6578\uff0c\u8b93 CNN \u81ea\u52d5\u53bb\u63a2\u7d22\uff0c\u9664 \u6211\u5011\u8003\u616e\u5728\u4e00\u6bb5\u97f3\u6a94\u88e1\uff0c\u6703\u6709\u591a\u7a2e\u4e0d\u540c\u985e\u578b\u7684\u97f3\u8a0a\u4e8b\u4ef6\uff0c\u4e14\u767c\u751f\u671f\u9593\u53ef\u80fd\u6703\u91cd\u758a\uff0c \u6216\u662f\u591a\u7a2e\u4e8b\u4ef6\u4e00\u4f75\u767c\u751f\u7684\u72c0\u6cc1\u3002\u56e0\u6b64\u5c07\u6a19\u6ce8\u6e96\u5247\u8a2d\u70ba\u4e00\u985e\u4e00\u8ecc\uff0c\u5404\u81ea\u7368\u7acb\u6a19\u6ce8(\u5982\u5716 2 \u4e4b\u898f\u7bc4\u793a\u610f\u5716) \uff0c\u4ee5\u5efa\u7acb\u97f3\u8a0a\u4e8b\u4ef6\u8cc7\u6599\u5eab\u3002 \u5716 3 Praat \u4eba\u5de5\u6a19\u8a18\u97f3\u8a0a\u4e8b\u4ef6\u7d50\u679c\u7bc4\u4f8b\u5716 3.1. \u7279\u5fb5\u53c3\u6578\u9078\u64c7\u8207\u97f3\u8a0a\u6a21\u578b\u8a13\u7df4 \u5728\u97f3\u8a0a\u4e8b\u4ef6\u5075\u6e2c\u7279\u5fb5\u53c3\u6578\u90e8\u5206\uff0c\u50b3\u7d71\u4e0a\u666e\u904d\u4f7f\u7528\u7684\u53c3\u6578\u70ba MFCCs\u3002\u4f46\u5176\u5be6\u5c0d\u97f3\u8a0a \u5716 5 CNN \u97f3\u8a0a\u4e8b\u4ef6\u5075\u6e2c\u7cfb\u7d71 \u5728\u6b64 CNN \u7db2\u8def\u67b6\u69cb\u4e2d\uff0c\u6709\u4e09\u500b\u4e3b\u8981\u7684\u795e\u7d93\u5c64\uff0c\u5305\u62ec(1)\u5377\u7a4d\u5c64\uff0c (2)\u6c60\u5316\u5c64\u8207(3) \u7279\u5225\u7684\u611b 270 0 0 300 5 294 \u5275\u9752\u5b85\u6025\u4fbf 384 0 0 384 1 384 MUSAN_Data 390 816 846 0 408 Music 3.31 12.52 Other 1.22 9.51 Other 94.15 13.25 97.47 9.51 \u5f0f\u7684\u6a94\u6848\u3002 Average 95.38 10.86 97.45 8.1 4.3. \u5be6\u9a57\u7d50\u679c 0 Total 2396 1141 1174 1435 427.92 2103 \u4ee5\u4e0b\u5206\u5225\u9032\u884c\u4e09\u500b\u5be6\u9a57\uff0c\u5305\u62ec\uff0c (1)\u6bd4\u8f03 GMM \u8207 CNN \u6a21\u578b\u7684\u6548\u80fd\uff0c (2)\u6bd4\u8f03\u4e0d 4.4. \u5be6\u9a57\u5206\u6790\u8207\u8a0e\u8ad6 \u53c3\u8003\u6587\u737b \u4e8b\u4ef6\u5075\u6e2c\uff0cMFCCs \u4e0d\u898b\u5f97\u662f\u6700\u4f73\u7684\u3002\u5c24\u5176\u662f\u5c0d\u97f3\u6a02\u8207\u5176\u4ed6\u4e8b\u4ef6\uff0c\u9084\u6709\u5f88\u591a\u4e0d\u540c\u7684\u53c3\u6578 \u88ab\u63d0\u51fa\u4f86\u3002 \u7531\u65bc\uff0c\u6211\u5011\u4e0d\u80fd\u78ba\u5b9a\u54ea\u4e00\u7a2e\u7279\u5fb5\u53c3\u6578\uff0c\u80fd\u5920\u6709\u6700\u597d\u7684\u97f3\u8a0a\u4e8b\u4ef6\u5075\u6e2c\u6548\u679c\uff0c\u6240\u4ee5\u6211\u5011 \u67b6\u69cb\uff0c\u5c24\u5176\u662f\u6e2c\u8a66\u4f7f\u7528 Spectrum(Specgram)\uff0c Mel-spectrum (Mels)\u8207 MFCCs\uff0c\u5206\u5225\u9032 \u884c\u8a9e\u97f3\u8207\u975e\u8a9e\u97f3\uff0c\u97f3\u6a02\u8207\u975e\u97f3\u6a02\uff0c\u5176\u4ed6\u8207\u975e\u5176\u4ed6\u97f3\u8a0a\u4e8b\u4ef6\u5075\u6e2c\u6a21\u578b\u3002 \u6578\uff0c\u8a13\u7df4 CNN \u6a21\u578b\u3002\u4ee5\u4e0b\u8aaa\u660e\u5377\u5404\u5c64\u7684\u904b\u4f5c\u65b9\u5f0f\u3002 3.2.1. \u5377\u7a4d\u5c64\u53ef\u5305\u542b\u8a31\u591a\u5377\u7a4d\u6838\uff0c\u5176\u904b\u4f5c\u5982\u4e0b\u5716 6 \u6240\u793a(\u4ee5\u4e8c\u7dad\u8f38\u5165\u53c3\u6578\u70ba\u4f8b) \uff0c\u4e3b\u8981\u662f \u570b\u969b\u6559\u80b2\u5fc3\u52d5\u7dda 210 0 0 210 0.33 210 \u5716 15 \u97f3\u6a02\u5075\u6e2c\u7d50\u679c \u5716 17 \u8a9e\u97f3/\u97f3\u6a02/\u5176\u5b83\u5207\u5272\u7d50\u679c\u6a19\u8a3b\u6a94\u683c\u5f0f \u5f9e\u8868 5 \u7684\u5be6\u9a57\u7d50\u679c\u53ef\u77e5\uff0cCNN \u7cfb\u7d71\u5728\u97f3\u8a0a\u4e8b\u4ef6\u5075\u6e2c\u7684\u6548\u679c\u8f03\u4f73\u3002 \u5377\u7a4d\u5c64 \u5716 7 \u6c60\u5316\u5c64 \u6587\u6559\u65b0\u805e 61 7 9 60 0.66 \u591a\u611b\u81ea\u5df1\u4e00\u9ede\u9ede 204 85 234 54 0.86 224 \u5716 8 \u8a9e\u97f3\u8a13\u7df4\u8cc7\u6599\u7d50\u679c \u56e0\u6b64\u975e\u8a9e\u97f3\u90e8\u5206\u660e\u986f\u53ef\u4ee5\u88ab\u6b63\u78ba\u5075\u6e2c\u4e26\u62ff\u6389\uff0c\u53ea\u4fdd\u7559\u8a9e\u97f3\u7684\u90e8\u5206\u3002 \u5716 9 \u8a9e\u97f3\u6e2c\u8a66\u8cc7\u6599\u7d50\u679c 68 \u4e0b\u5c64\u5247\u662f\u4eba\u5de5\u6240\u6a19\u8a18\u7684\u97f3\u8a0a\u4e8b\u4ef6\u6a19\u6e96\u7b54\u6848\u3002\u5f9e\u5716\u4e2d\u53ef\u77e5\u9053\u8a9e\u97f3\u8207\u97f3\u6a02\u983b\u8b5c\u7279\u6027\u76f8\u7576\u4e0d\u540c\uff0c \u6c42\u53d6\u4e26\u6e2c\u8a66\u5404\u7a2e\u4e0d\u540c\u53c3\u6578\u7684\u97f3\u8a0a\u4e8b\u4ef6\u5075\u6e2c\u6548\u80fd\u3002\u56e0\u6b64\u5728\u672c\u8ad6\u6587\u4e2d\uff0c\u5c07\u5617\u8a66\u5982\u5716 4 \u6240\u793a\u7684 \u5168\u9023\u63a5\u5c64\u3002\u5176\u4e2d\u5377\u7a4d\u5c64\u8207\u6c60\u5316\u5c64\u53ef\u91cd\u8907\u6578\u6b21\u3002\u4f7f\u7528 CNN \u7684\u597d\u8655\u662f\uff0cCNN \u80fd\u5229\u7528\u5377\u7a4d\u5c64\uff0c \u81ea\u52d5\u5b78\u7fd2\u5982\u4f55\u6c42\u53d6\u6700\u4f73\u53c3\u6578\uff0c\u8207\u5229\u7528\u6c60\u5316\u5c64\uff0c\u5bb9\u5fcd\u76ee\u6a19\u4e8b\u4ef6\u5728\u983b\u8b5c\u4e0a\u7684\u4f4d\u7f6e\u8b8a\u7570\u3002\u6240\u4ee5 \u79d1\u5b78 SoEasy 246 15 60 198 0.78 204 4.3.1. \u6211\u5011\u53ef\u4ee5\u81ea\u7531\u5617\u8a66\u8a31\u591a\u4e0d\u540c\u7684\u7279\u5fb5\u53c3\u6578\uff0c\u5c24\u5176\u662f\u53ef\u4ee5\u4e0d\u7d93\u904e\u4eba\u70ba\u8a2d\u8a08\uff0c\u76f4\u63a5\u8f38\u5165\u983b\u8b5c\u53c3 \u8868 4 \u6559\u80b2\u96fb\u81fa\u97f3\u8a0a\u4e8b\u4ef6\u6e2c\u8a66\u8a9e\u6599(minute) Program speech non-speech music non-music other non-other \u81ea\u7136\u6709\u610f\u601d 156 1 1 138 1 \u4e8c\u5c64\u70ba\u983b\u8b5c\u3001\u7b2c\u4e09\u5c64\u70ba\u53ea\u4fdd\u7559\u5075\u6e2c\u51fa\u7684\u8a9e\u97f3\u4e8b\u4ef6\u90e8\u5206\u7684\u97f3\u6a94\u6ce2\u5f62\u3001\u7b2c\u56db\u5c64\u70ba\u5176\u983b\u8b5c\u3001\u6700 156 \u8a9e\u97f3\u4e8b\u4ef6\u5075\u6e2c 4.4.1. \u540c\u53c3\u6578\u7684\u5f71\u97ff\u8207(3)\u6c42\u53d6\u6aa2\u6e2c\u932f\u8aa4\u6b0a\u8861\u66f2\u7dda(detection error tradeoff curves\uff0cDETs)\u4ee5 \u8a08\u7b97 equal error rate(EER) \uff0c\u5206\u5225\u8aaa\u660e\u5982\u4e0b\uff1a \u5716 14 \u662f\u7d93\u8a9e\u97f3\u4e8b\u4ef6\u5075\u6e2c\u5668\u8655\u7406\u904e\u7684\u7d50\u679c\u7bc4\u4f8b\u5716\u3002\u5716\u4e2d\u7b2c\u4e00\u5c64\u70ba\u539f\u59cb\u97f3\u6a94\u6ce2\u5f62\u3001\u7b2c [1]</td></tr></table>" | |
| } | |
| } | |
| } | |
| } |