| { | |
| "paper_id": "O17-1002", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T07:59:24.551006Z" | |
| }, | |
| "title": "A Study on Voice Activation Detection by Using Neural Networks", | |
| "authors": [ | |
| { | |
| "first": "Yu-Chih", | |
| "middle": [], | |
| "last": "\u9127\u6709\u5fd7", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "National Taipei University", | |
| "location": {} | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Deng", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "National Taipei University", | |
| "location": {} | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Chen-Yu", | |
| "middle": [], | |
| "last": "\u6c5f\u632f\u5b87", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "National Taipei University", | |
| "location": {} | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Chiang", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "National Taipei University", | |
| "location": {} | |
| }, | |
| "email": "[email protected]" | |
| }, | |
| { | |
| "first": "Chen-Ming", | |
| "middle": [], | |
| "last": "\u6f58\u632f\u9298", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "\u4e2d\u83ef\u96fb\u4fe1\u7814\u7a76\u9662 Chunghwa Telecom Laboratories", | |
| "institution": "", | |
| "location": {} | |
| }, | |
| "email": "[email protected]" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Pan", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "\u4e2d\u83ef\u96fb\u4fe1\u7814\u7a76\u9662 Chunghwa Telecom Laboratories", | |
| "institution": "", | |
| "location": {} | |
| }, | |
| "email": "" | |
| } | |
| ], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "\u672c\u7814\u7a76\u4ee5\u6df1\u5c64\u985e\u795e\u7d93\u7db2\u8def (Deep Neural Network, DNN) \u9032\u884c\u8a9e\u97f3\u7aef\u9ede\u5075\u6e2c\uff0c\u8a0e\u8ad6\u4e86\u4ee5 \u4e0b\u5f71\u97ff\u8a9e\u97f3\u7aef\u9ede\u5075\u6e2c\u8868\u73fe\u7684\u5e7e\u500b\u8b8a\u91cf\uff1a(1) \u7279\u5fb5\u53c3\u6578\u62bd\u53d6\u6642\u8003\u91cf\u7684\u5206\u6790\u8996\u7a97\u5927\u5c0f\u3001(2) DNN \u5c64\u6578\u3001(3) \u8a0a\u8e81\u6bd4\u4ee5\u53ca(4) \u80cc\u666f\u74b0\u5883\u985e\u578b\u3002\u5be6\u9a57\u662f\u4f7f\u7528\u53f0\u5317\u5927\u5b78\u96dc\u8a0a\u8a9e\u6599\u5eab (NTPU Noise Corpus)\uff0c\u6b64\u8cc7\u6599\u5eab\u662f\u7531\u667a\u6167\u578b\u624b\u6a5f\u9304\u88fd\u7684\u5404\u7a2e\u80cc\u666f\u96dc\u8a0a\u4ee5\u53ca TCC300 \u8a9e\u6599\u5eab\u6df7\u97f3 \u800c\u6210\uff0c\u80cc\u666f\u74b0\u5883\u5305\u542b\uff1a (1)\u516c\u8eca\u7ad9\u3001 (2)\u6377\u904b\u7ad9\u3001 (3)\u706b\u8eca\u7ad9\u3001 (4)\u9910\u5ef3\uff0c\u800c\u6df7\u97f3\u7684\u8a0a \u8e81\u6bd4\u6709\uff1a10dB\u30015dB\u30010dB \u4ee5\u53ca\u4e7e\u6de8\u8a9e\u97f3\u3002\u7cfb\u7d71\u8a55\u91cf\u7684\u6a19\u6e96\u70ba\u97f3\u6846\u6b63\u78ba\u7387 (frame accuracy) \u4ee5\u53ca equal error rate (EER)\u3002\u5be6\u9a57\u7d50\u679c\u6307\u51fa\u7279\u5fb5\u53c3\u6578\u5206\u6790\u8996\u7a97\u8d8a\u5927\u800c\u5728\u8a13\u7df4\u8207\u767c\u5c55\u96c6\u5408 \u7684\u8868\u73fe\u6709\u660e\u986f\u8b8a\u597d\u7684\u8da8\u52e2\uff0c\u4f46\u5728\u6e2c\u8a66\u96c6\u5408\u5247\u9032\u6b65\u5e45\u5ea6\u8f03\u5c0f\u3002DNN \u5c64\u6578\u5728 2 layer \u6642\u7684 multi-condition \u5176\u8868\u73fe\u8f03\u597d\uff0c\u8a0a\u8e81\u6bd4\u8d8a\u9ad8\u5247\u9032\u6b65\u4e5f\u6bd4\u8f03\u986f\u8457\uff0c\u5c24\u5176\u662f\u5728\u80cc\u666f\u74b0\u5883\u70ba\u9910\u5ef3 \u7684\u60c5\u6cc1\u4e0b\u3002\u6700\u5f8c multi-condition \u8a13\u7df4\u6cd5\u4e2d\u7684\u5404\u500b condition\uff0c\u5728\u6e2c\u8a66\u96c6\u5408\u7684\u8868\u73fe\u7686\u512a\u65bc matched-condition\uff0c\u8b49\u5be6\u4e86 multi-condition \u4e2d\u7684\u5404\u500b condition\uff0c\u5728 hidden layer \u4e2d\u80fd\u5920\u4e92 \u76f8\u7684\u5b78\u7fd2\u3002 5", | |
| "pdf_parse": { | |
| "paper_id": "O17-1002", | |
| "_pdf_hash": "", | |
| "abstract": [ | |
| { | |
| "text": "\u672c\u7814\u7a76\u4ee5\u6df1\u5c64\u985e\u795e\u7d93\u7db2\u8def (Deep Neural Network, DNN) \u9032\u884c\u8a9e\u97f3\u7aef\u9ede\u5075\u6e2c\uff0c\u8a0e\u8ad6\u4e86\u4ee5 \u4e0b\u5f71\u97ff\u8a9e\u97f3\u7aef\u9ede\u5075\u6e2c\u8868\u73fe\u7684\u5e7e\u500b\u8b8a\u91cf\uff1a(1) \u7279\u5fb5\u53c3\u6578\u62bd\u53d6\u6642\u8003\u91cf\u7684\u5206\u6790\u8996\u7a97\u5927\u5c0f\u3001(2) DNN \u5c64\u6578\u3001(3) \u8a0a\u8e81\u6bd4\u4ee5\u53ca(4) \u80cc\u666f\u74b0\u5883\u985e\u578b\u3002\u5be6\u9a57\u662f\u4f7f\u7528\u53f0\u5317\u5927\u5b78\u96dc\u8a0a\u8a9e\u6599\u5eab (NTPU Noise Corpus)\uff0c\u6b64\u8cc7\u6599\u5eab\u662f\u7531\u667a\u6167\u578b\u624b\u6a5f\u9304\u88fd\u7684\u5404\u7a2e\u80cc\u666f\u96dc\u8a0a\u4ee5\u53ca TCC300 \u8a9e\u6599\u5eab\u6df7\u97f3 \u800c\u6210\uff0c\u80cc\u666f\u74b0\u5883\u5305\u542b\uff1a (1)\u516c\u8eca\u7ad9\u3001 (2)\u6377\u904b\u7ad9\u3001 (3)\u706b\u8eca\u7ad9\u3001 (4)\u9910\u5ef3\uff0c\u800c\u6df7\u97f3\u7684\u8a0a \u8e81\u6bd4\u6709\uff1a10dB\u30015dB\u30010dB \u4ee5\u53ca\u4e7e\u6de8\u8a9e\u97f3\u3002\u7cfb\u7d71\u8a55\u91cf\u7684\u6a19\u6e96\u70ba\u97f3\u6846\u6b63\u78ba\u7387 (frame accuracy) \u4ee5\u53ca equal error rate (EER)\u3002\u5be6\u9a57\u7d50\u679c\u6307\u51fa\u7279\u5fb5\u53c3\u6578\u5206\u6790\u8996\u7a97\u8d8a\u5927\u800c\u5728\u8a13\u7df4\u8207\u767c\u5c55\u96c6\u5408 \u7684\u8868\u73fe\u6709\u660e\u986f\u8b8a\u597d\u7684\u8da8\u52e2\uff0c\u4f46\u5728\u6e2c\u8a66\u96c6\u5408\u5247\u9032\u6b65\u5e45\u5ea6\u8f03\u5c0f\u3002DNN \u5c64\u6578\u5728 2 layer \u6642\u7684 multi-condition \u5176\u8868\u73fe\u8f03\u597d\uff0c\u8a0a\u8e81\u6bd4\u8d8a\u9ad8\u5247\u9032\u6b65\u4e5f\u6bd4\u8f03\u986f\u8457\uff0c\u5c24\u5176\u662f\u5728\u80cc\u666f\u74b0\u5883\u70ba\u9910\u5ef3 \u7684\u60c5\u6cc1\u4e0b\u3002\u6700\u5f8c multi-condition \u8a13\u7df4\u6cd5\u4e2d\u7684\u5404\u500b condition\uff0c\u5728\u6e2c\u8a66\u96c6\u5408\u7684\u8868\u73fe\u7686\u512a\u65bc matched-condition\uff0c\u8b49\u5be6\u4e86 multi-condition \u4e2d\u7684\u5404\u500b condition\uff0c\u5728 hidden layer \u4e2d\u80fd\u5920\u4e92 \u76f8\u7684\u5b78\u7fd2\u3002 5", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Abstract", | |
| "sec_num": null | |
| } | |
| ], | |
| "body_text": [ | |
| { | |
| "text": "= = = \u00d7 \u00e5 \u00e5 (1-3)", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "\u5176\u4e2d \" \u4ee3\u8868\u67d0\u8a9e\u53e5 (TCC300 \u8a9e\u6599\u4e4b\u8a9e\u53e5) \u7684\u7b2c t \u500b sample \u7684 sample value\uff0cT \u4ee3\u8868\u67d0\u8a9e \u53e5\u4ee5 sample \u6578\u70ba\u55ae\u4f4d\u7684\u9577\u5ea6\uff0c\u800c '+,,-. ( )\u4ee3\u8868\u7b2c t \u500b sample \u662f\u5426\u70ba\u8a9e\u97f3\u4fe1\u865f\uff0c\u4e5f\u5c31\u662f 1, if sample is a speech sample ( ) 0, if sample is a non-speech sample ", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "speech t t t d \u00ec = \u00ed \u00ee (1-4) \u800c\u96dc\u8a0a\u7684\u529f\u7387 ) ( \u53ef\u7531\u5f0f(1-5)\u8a08\u7b97\u5f97\u5230 1 2 2 2 2 2 0 1L n t n t g n g L s s - = = \u00d7 = \u00d7 \u00e5 (1-5) \u5176\u4e2d \" \u4ee3\u8868\u67d0\u96dc\u8a0a\u4fe1\u865f\u6bb5\u843d\u7684\u7b2c t \u7684 sample \u4e4b sample value\uff1b L \u4ee3\u8868\u6b64\u96dc\u8a0a\u6bb5\u843d\u7684 sample \u6578\uff1b g \u4ee3\u8868\u96dc\u8a0a\u4fe1\u865f\u7684\u653e\u5927\u500d\u7387(magnification)\uff1b 1 2 2 0 1 L n t t n L s - = = \u00e5", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| } | |
| ], | |
| "back_matter": [], | |
| "bib_entries": { | |
| "BIBREF0": { | |
| "ref_id": "b0", | |
| "title": "Statistical Voice Activity Detection Using Low-Variance Spectrum Estimation and an Adaptive Threshold", | |
| "authors": [ | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Davis", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Nordholm", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "R", | |
| "middle": [], | |
| "last": "Togneri", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2006, | |
| "venue": "IEEE Signal Processing Society", | |
| "volume": "14", | |
| "issue": "2", | |
| "pages": "412--424", | |
| "other_ids": { | |
| "DOI": [ | |
| "10.1109/TSA.2005.855842" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Davis, A., Nordholm, S., & Togneri, R. (2006, February). Statistical Voice Activity Detection Using Low-Variance Spectrum Estimation and an Adaptive Threshold. IEEE Signal Processing Society, 14(2), 412-424. doi:10.1109/TSA.2005.855842.", | |
| "links": null | |
| }, | |
| "BIBREF1": { | |
| "ref_id": "b1", | |
| "title": "Voice Activity Detection Based on Statistical Likelihood Ratio with Adaptive Thresholding. IWAENC", | |
| "authors": [ | |
| { | |
| "first": "X", | |
| "middle": [], | |
| "last": "Li", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "R", | |
| "middle": [], | |
| "last": "Horaud", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "L", | |
| "middle": [], | |
| "last": "Girin", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2016, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Li, X., Horaud, R., & Girin, L. (2016, October). Voice Activity Detection Based on Statistical Likelihood Ratio with Adaptive Thresholding. IWAENC, China.", | |
| "links": null | |
| }, | |
| "BIBREF2": { | |
| "ref_id": "b2", | |
| "title": "Vector Quantization in Speech Coding", | |
| "authors": [ | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Makhoul", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Roucos", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Gish", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1985, | |
| "venue": "Proceedings of the IEEE", | |
| "volume": "73", | |
| "issue": "11", | |
| "pages": "1551--1558", | |
| "other_ids": { | |
| "DOI": [ | |
| "10.1109/PROC.1985.13340" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Makhoul, J., Roucos, S., & Gish, H. (1985, November). Vector Quantization in Speech Coding. Proceedings of the IEEE, 73(11), 1551-1558. doi:10.1109/PROC. 1985.13340.", | |
| "links": null | |
| }, | |
| "BIBREF3": { | |
| "ref_id": "b3", | |
| "title": "Gaussian Mixture Models", | |
| "authors": [ | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Reynolds", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2008, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Reynolds, D. (2008). Gaussian Mixture Models. Springer US.", | |
| "links": null | |
| }, | |
| "BIBREF4": { | |
| "ref_id": "b4", | |
| "title": "Acoustic Model Training for Speech Recognition System", | |
| "authors": [ | |
| { | |
| "first": "N", | |
| "middle": [ | |
| "S" | |
| ], | |
| "last": "Dlamini", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2015, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Dlamini, N. S. (2015, November). Acoustic Model Training for Speech Recognition System, National Taipei University, New Taipei City.", | |
| "links": null | |
| }, | |
| "BIBREF5": { | |
| "ref_id": "b5", | |
| "title": "Voice Activity Detection Based on Sequential Gaussian Mixture Model and with Maximum Likelihood Criterion", | |
| "authors": [ | |
| { | |
| "first": "Z", | |
| "middle": [], | |
| "last": "Shen", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Wei", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Dang", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2016, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Shen, Z., Wei, J., & Dang J. (2016, October). Voice Activity Detection Based on Sequential Gaussian Mixture Model and with Maximum Likelihood Criterion. ISCSLP, China.", | |
| "links": null | |
| }, | |
| "BIBREF6": { | |
| "ref_id": "b6", | |
| "title": "Principles of neurodynamics. perceptrons and the theory of brain mechanisms", | |
| "authors": [ | |
| { | |
| "first": "F", | |
| "middle": [], | |
| "last": "Rosenblatt", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1961, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Rosenblatt, F. (1961). Principles of neurodynamics. perceptrons and the theory of brain mechanisms. Springer, Berlin, Heidelberg.", | |
| "links": null | |
| }, | |
| "BIBREF7": { | |
| "ref_id": "b7", | |
| "title": "Finding structure in time", | |
| "authors": [ | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Elamn", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "L", | |
| "middle": [], | |
| "last": "", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1990, | |
| "venue": "Cognitive Science", | |
| "volume": "14", | |
| "issue": "2", | |
| "pages": "179--211", | |
| "other_ids": { | |
| "DOI": [ | |
| "10.1207/s15516709cog1402_1" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Elamn, J., L. (1990, April). Finding structure in time. Cognitive Science,14(2), 179-211. doi:10.1207/s15516709cog1402_1.", | |
| "links": null | |
| }, | |
| "BIBREF8": { | |
| "ref_id": "b8", | |
| "title": "Serial order: A parallel distributed processing approach", | |
| "authors": [ | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Jordan", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "I", | |
| "middle": [], | |
| "last": "", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1997, | |
| "venue": "Advances in Psychology", | |
| "volume": "121", | |
| "issue": "1", | |
| "pages": "471--495", | |
| "other_ids": { | |
| "DOI": [ | |
| "10.1016/S0166-4115(97)80111-2" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Jordan, M., I. (1997, September). Serial order: A parallel distributed processing approach. Advances in Psychology, 121(1), 471-495. doi:10.1016/S0166-4115(97)80111-2.", | |
| "links": null | |
| }, | |
| "BIBREF11": { | |
| "ref_id": "b11", | |
| "title": "Methods for speech SNR estimation: Evaluation tool and analysis of VAD dependency", | |
| "authors": [ | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Vondasek", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "P", | |
| "middle": [], | |
| "last": "Pollak", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2005, | |
| "venue": "Radioengineering", | |
| "volume": "14", | |
| "issue": "1", | |
| "pages": "6--11", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Vondasek, M., & Pollak, P. (2005). Methods for speech SNR estimation: Evaluation tool and analysis of VAD dependency. Radioengineering, 14(1), 6-11.", | |
| "links": null | |
| }, | |
| "BIBREF12": { | |
| "ref_id": "b12", | |
| "title": "A new self-adaptive audio-mix algorithm's research and realization based on voice energy", | |
| "authors": [ | |
| { | |
| "first": "Z", | |
| "middle": [], | |
| "last": "Wei", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "P", | |
| "middle": [], | |
| "last": "Lou", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2009, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Wei, Z., & Lou, P. (2009, September). A new self-adaptive audio-mix algorithm's research and realization based on voice energy, Beijing University of Posts and Telecommunications, Beijing.", | |
| "links": null | |
| }, | |
| "BIBREF13": { | |
| "ref_id": "b13", | |
| "title": "Audio mixing for centralized conferences in a SIP environment", | |
| "authors": [ | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Hawwa", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2002, | |
| "venue": "ICME", | |
| "volume": "2", | |
| "issue": "1", | |
| "pages": "269--272", | |
| "other_ids": { | |
| "DOI": [ | |
| "10.1109/ICME,2002.1035572" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Hawwa, S. (2002, August). Audio mixing for centralized conferences in a SIP environment. ICME, 2(1), 269-272. doi:10.1109/ICME,2002.1035572.", | |
| "links": null | |
| }, | |
| "BIBREF14": { | |
| "ref_id": "b14", | |
| "title": "A fast learning algorithm for deep belief nets", | |
| "authors": [ | |
| { | |
| "first": "E", | |
| "middle": [ | |
| "H" | |
| ], | |
| "last": "Geoffrey", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Osindero", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Y", | |
| "middle": [ | |
| "W" | |
| ], | |
| "last": "Teh", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2006, | |
| "venue": "Neural computation", | |
| "volume": "18", | |
| "issue": "7", | |
| "pages": "1527--1554", | |
| "other_ids": { | |
| "DOI": [ | |
| "10.1162/neco.2006.18.7.1527" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Geoffrey, E. H., Osindero, S., & Teh, Y. W. (2006, May). A fast learning algorithm for deep belief nets. Neural computation 18(7), 1527-1554. doi:10.1162/neco.2006.18.7.1527.", | |
| "links": null | |
| }, | |
| "BIBREF15": { | |
| "ref_id": "b15", | |
| "title": "Greedy layer-wise training of deep networks", | |
| "authors": [ | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Bengio", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2006, | |
| "venue": "Advances in neural information processing systems, 19(NIPS'06)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "153--160", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Bengio, Y. (2006, January). Greedy layer-wise training of deep networks. Advances in neural information processing systems, 19(NIPS'06), 153-160.", | |
| "links": null | |
| }, | |
| "BIBREF16": { | |
| "ref_id": "b16", | |
| "title": "Probability, Statistics, and Random Process for Electrical Engineering", | |
| "authors": [ | |
| { | |
| "first": "A", | |
| "middle": [ | |
| "L" | |
| ], | |
| "last": "Garcia", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2009, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Garcia, A. L. (2009). Probability, Statistics, and Random Process for Electrical Engineering (3 rd Edition), New Jersey: Pearson Prentice Hall.", | |
| "links": null | |
| }, | |
| "BIBREF17": { | |
| "ref_id": "b17", | |
| "title": "Multimedia Signal Processing, Theory and Applications in Speech, Music and Communications", | |
| "authors": [ | |
| { | |
| "first": "S", | |
| "middle": [ | |
| "V" | |
| ], | |
| "last": "Vaseghi", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2007, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Vaseghi, S. V. (2007). Multimedia Signal Processing, Theory and Applications in Speech, Music and Communications. UK: John Wiley & Sons Ltd.", | |
| "links": null | |
| } | |
| }, | |
| "ref_entries": { | |
| "FIGREF0": { | |
| "num": null, | |
| "type_str": "figure", | |
| "uris": null, | |
| "text": "-based)\u7684\u8a9e\u97f3\u7aef\u9ede\u5075\u6e2c\u6280\u8853(Voice Activation Detection, Activation Detection, VAD)\u76ee\u7684\u662f\u5728\u65bc\u6aa2\u6e2c\u8a9e\u97f3\u8a0a\u865f\u4e2d\uff0c\u8a9e\u97f3\u7247\u6bb5\u7684 \u958b\u59cb\u8207\u7d50\u675f\uff0c\u5c0d\u65bc ASR \u7cfb\u7d71\u662f\u5f88\u91cd\u8981\u7684\u524d\u8655\u7406(Front-end)\u5de5\u4f5c\u4e4b\u4e00\u3002\u56e0\u70ba\u8a9e\u97f3\u7aef\u9ede\u5075\u6e2c \u7684\u6548\u80fd\uff0c\u6703\u76f4\u63a5\u5f71\u97ff ASR \u7cfb\u7d71\u7684\u8fa8\u8b58\u7387\uff0c\u6240\u4ee5\u6b64\u985e\u65b9\u6cd5\u88ab\u61c9\u7528\u65bc\u8a9e\u97f3\u559a\u9192(Voice Trigger, VT)\u3001\u8a9e\u97f3\u6703\u8b70(Audio conference)\u3001\u8a9e\u97f3\u7de8\u78bc(Speech codding)\u3001\u514d\u6301\u901a\u8a71(Hands-free)\u3001 \u8a9e \u97f3 \u964d \u566a (Speech enhancement) \u3001 \u8072 \u97f3 \u5b9a \u4f4d (Sound positioning) \u3001 \u8a9e \u8005 \u8fa8 \u8b58 (Speaker" | |
| }, | |
| "FIGREF1": { | |
| "num": null, | |
| "type_str": "figure", | |
| "uris": null, | |
| "text": "\u70ba\u8a13\u7df4\u8cc7\u6599\u7684\u7e3d\u6578\u800c m \u5247\u4ee3\u8868\u5176 index\u3002\u5c0d\u65bc category \u7684\u60c5\u6cc1\u4f86\u8aaa\uff0c\u5047\u8a2d y \u662f\u4e00\u500b\u6a5f \u7387\u5206\u5e03\u3001C \u8868\u793a\u985e\u5225\u6578\u91cf\u800c i \u70ba\u5176 index\uff0c\u5247 ML \u6e96\u5247\u70ba \u8868\u793a\u8a13\u7df4 samples\u3001e \u70ba Learning rate\u3002\u8f38\u51fa\u5c64\u6b0a\u91cd\u77e9\u9663\u76f8\u5c0d\u65bc\u8a13\u7df4\u6e96\u5247\u7684\u68af\u5ea6\u53d6\u6c7a\u65bc\u5176 \u8a13\u7df4\u6e96\u5247\uff0c\u5728 Category \u7684\u60c5\u6cc1\u4e0b\u5247\u4f7f\u7528 CE \u8a13\u7df4\u6e96\u5247(4-4)\u5f0f\u548c softmax \u8f38\u51fa\u5c64 \u5728\u6b64\u5c0d\u65bc DNN \u7684 back propagation \u6f14\u7b97\u6cd5\u95dc\u9375\u6b65\u9a5f\u505a\u8aaa\u660e\u3002" | |
| }, | |
| "TABREF0": { | |
| "num": null, | |
| "html": null, | |
| "text": "Recognition)\u53ca\u8a9e\u97f3\u8fa8\u8b58(Speech Recognition)\u88e1\u3002\u6211\u5011\u53ef\u4ee5\u5c07\u773e\u591a VAD \u7684\u6f14\u7b97\u6cd5\u5206\u6210\u4ee5 \u4e0b\u56db\u7a2e\u65b9\u6cd5\uff0c\u4ee5 Energy-based[1][2]\u3001Statistical-based[3][4]\u3001GMM-based[5][6][7][8]\u8207 NNbased[9][10][11][12]\u4e4b\u6aa2\u6e2c\u6cd5\u3002 \u5728 Energy-based \u6aa2\u6e2c\u6cd5\u4e2d\u8a9e\u97f3\u90e8\u5206\u7684 energy \u660e\u986f\u6bd4\u96dc\u8a0a energy \u5927\u3002\u6240\u4ee5\u6211\u5011\u53ef\u4ee5 \u5728\u6642\u57df\u4e0a\u5b9a\u7fa9\u4e00\u500b\u7c21\u55ae\u7684 Threshold \u4f86\u5c0d\u65bc Energy\u3001Zero Crossing Rate (ZCR)\u548c Pitch \u505a \u5224\u65b7\uff0c\u4e26\u4f7f\u7528 VAD state machine \u4f86\u63cf\u8ff0\u8a9e\u97f3\u7684\u958b\u59cb\u8207\u7d50\u675f\u3002 Statistical-based \u6aa2\u6e2c\u6cd5\uff0c\u6b64\u65b9\u6cd5\u5728\u904e\u53bb\u7684\u7814\u7a76\u662f\u89c0\u5bdf\u6642\u57df\u4e0a\u67d0\u983b\u5e36\u4e0a\u8a9e\u97f3\u8a0a\u865f\u9577\u671f \u7a69\u5b9a\u7684\u8b8a\u5316\u4ee5\u53ca\u5728\u983b\u57df\u4e0a\u89c0\u5bdf\u5176\u5e73\u5766\u5ea6\uff0c\u5229\u7528\u9019\u4e9b\u65b9\u6cd5\u4f86\u5224\u5225\u51fa\u5404\u8a72\u7247\u6bb5\u70ba\u8a9e\u97f3\u9084\u662f\u975e \u8a9e\u97f3\u7684\u8b8a\u5316\u3002\u8fd1\u671f\u7684 Statistical-based VAD \u7814\u7a76\u5247\u662f\u8a66\u5716\u53bb\u6700\u4f73\u5316\u6aa2\u6e2c\u96dc\u8a0a\u7684\u5b58\u5728\uff0c\u4f8b \u5982\u662f\u4f7f\u7528 low-variance spectrum \u4f30\u8a08\u6cd5\u4e26\u4e14\u914d\u5408\u7d71\u8a08\u6aa2\u6e2c\u6a5f\u5236\u4f86\u78ba\u5b9a\u6700\u4f73\u7684 Threshold\uff0c \u4e26\u4e14\u642d\u914d Hangover state machine \u4f86\u907f\u514d\u8a9e\u97f3\u5feb\u7d50\u675f\u6642\u7684\u8a9e\u53e5\uff0c\u5728 low-energy \u7684\u8a9e\u97f3\u7247 NN-based \u6aa2\u6e2c\u6cd5\u4e2d\u50b3\u7d71\u662f\u4f7f\u7528 MLP \u7684\u67b6\u69cb\uff0c\u4f46\u8fd1\u5e74\u4f86\u8a31\u591a\u5b78\u8005\u5c0d\u65bc NN \u6709\u7a81 \u7834\u6027\u7684\u7814\u7a76\u6210\u679c\uff0c\u6240\u4ee5\u9010\u6f38\u6709 DNN\u3001RNN \u751a\u81f3\u662f LSTM\u3001GRU \u7b49\u67b6\u69cb\u51fa\u73fe\u3002DNN \u6539 \u5584\u4e86\u50b3\u7d71 MLP \u53ea\u6709\u4e09\u5c64\u4e4b\u67b6\u69cb(input layer\u3001hidden layer \u8207 output layer)\uff0c\u589e\u52a0\u4e86 MLP \u5728 hidden layer \u7684\u6578\u76ee\u3001hidden layer \u88e1\u9762 node \u7684\u6578\u76ee\uff0c\u4f7f\u5f97\u6574\u500b network \u8b8a\u5f97\u53c8\u5bec\u4e14 \u6df1\u3002\u4e26\u4e14 DNN \u52a0\u5165\u4e86 dropout \u53ca mini-batch \u5728\u50b3\u7d71\u7684 MLP \u8a13\u7df4\u904e\u7a0b\u4e2d\uff0c\u5c0d\u65bc neural network \u7684 unit \u4f9d\u7167\u4e00\u5b9a\u6bd4\u4f8b\u66ab\u6642\u6027\u96a8\u6a5f\u7684\u4e1f\u68c4\uff0c\u5176\u512a\u9ede\uff1a\u662f\u5728\u65bc\u8a13\u7df4\u6578\u64da\u8f03\u5c11\u6642\uff0c\u5247 \u5176\u4e2d \" \u70ba TCC300 \u7684\u8a9e\u97f3\u90e8\u5206(clean speech) \" \u3001\u5247\u662f\u96dc\u8a0a\u7684\u90e8\u5206(noisy data)\uff0c\u7136\u800c g \u70ba \u96dc\u8a0a\u90e8\u5206 \" \u6b32\u5408\u6210\u51fa\u8cc7\u6599 Noise speech \" \u6240\u9700\u8981\u4e58\u4e0a\u4e4b\u500d\u7387\u4e26\u52a0\u4e0a\u8a9e\u97f3\u90e8\u5206 \" \u3002\u4e00\u822c\u4f86", | |
| "type_str": "table", | |
| "content": "<table><tr><td colspan=\"7\">\u8868\u4e00\uff1aTCC300 \u8a9e\u6599\u5eab\u8cc7\u8a0a\u7d71\u8a08\u8868 \u8a9e\u8005\u7e3d\u6578 \u7e3d\u97f3\u7bc0\u6578 \u7537 152 \u7537 193,167 \u5973 151 \u5973 197,296 \u7e3d\u8a08 303 \u7e3d\u8a08 390,463 (\u4e8c) \u3001\u667a\u6167\u578b\u624b\u6a5f\u96dc\u8a0a\u8cc7\u6599\u5eab \u8a9e\u6599\u5eab \u6587\u7ae0\u5c6c\u6027 TCC300 \u9577\u77ed\u53e5 \u8868\u4e09\uff1a\u53f0\u5317\u5927\u5b78\u96dc\u8a0a\u8a9e\u6599\u5eab \u8a9e\u6599\u5eab NTPU Additive Noise Corpus \u53d6\u6a23\u983b\u7387 16kHz \u53d6\u6a23\u7de8\u78bc Lin 16 \u8072\u9053 1 \u8a9e\u97f3\u5167\u5bb9 \u4e2d\u7814\u9662 500 \u842c\u8a5e\u8a5e\u985e\u6a19\u8a18\u8a9e\u6599\u5eab \u7537 \u5973 \u7e3d\u8a08 \u6a94\u6848\u7e3d\u6578 4,614 4,265 8,879 \u8a9e\u97f3\u9577\u5ea6 \u9577\u53e5+\u77ed\u53e5(TCC300) \u6a21\u5f0f\u7a2e\u985e Clean Multiple Additive Noise \u5716\u4e00\uff1aNoisy speech \u8cc7\u6599\u5eab\u5efa\u7acb\u65b9\u6cd5\u5716 Noise speech \" \u5efa\u7acb\u4e4b\u6578\u5b78\u5f0f\u5982\u4e0b\u5f0f\u6240\u793a t t t x s g n = + \u00d7 (1-1) \u5728 \u53ef\u4ee5\u7528\u65bc\u907f\u514d over-fit\uff1b\u4f46\u662f\u7f3a\u9ede\uff1a\u5247\u6703\u4f7f\u8a13\u7df4\u6642\u9593\u52a0\u9577\uff0c\u4f46\u4e0d\u5f71\u97ff\u5176\u6e2c\u8a66\u7684\u6642\u9593\uff0c\u4e14\u6bcf \u4e00\u500b mini-batch \u90fd\u5728\u8a13\u7df4\u4e0d\u540c\u7684 network\u3002 \u7121\u52a0\u6210\u6027\u566a\u97f3 \u52a0\u6210\u6027\u566a\u97f3: \u9910\u5ef3(\u53f0\u5317\u5927\u5b78)\u3001\u706b\u8eca\u7ad9(\u677f\u6a4b)\u3001\u6377\u904b\u7ad9 \u96a8\u8457\u79d1\u6280\u7684\u9032\u6b65\uff0c\u9304\u97f3\u88dd\u7f6e\u5df2\u7d93\u4e0d\u50c5\u9650\u65bc\u50b3\u7d71\u7684\u5916\u63a5\u9ea5\u514b\u98a8\uff0c\u50cf\u662f\u7b46\u8a18\u578b\u96fb\u8166\u3001\u9304\u97f3\u7b46\u3001 SNR: 0\u30015\u300110 dB \u5e73\u677f\u3001\u667a\u6167\u578b\u624b\u6a5f\u7b49\u90fd\u5177\u6709\u9ea5\u514b\u98a8\u9304\u97f3\u88dd\u7f6e\u3002\u4f46\u5728\u9019\u4e9b\u88dd\u7f6e\u4e2d\u4e26\u7121\u6cd5\u638c\u63a7\u5176\u9304\u97f3\u7684\u54c1\u8cea\uff0c Training Mode (\u677f\u6a4b)\u3001\u516c\u8eca\u7ad9\u724c(\u677f\u6a4b) \u8aaa\u6a19\u6e96\u5df2\u77e5\u5b9a\u7fa9\u7684\u8a9e\u53e5\u4e4b Global SNR \u7b97\u6cd5[14]\u5982\u5f0f(1-2)\u6240\u793a</td></tr><tr><td colspan=\"7\">(\u4e09) \u3001\u7814\u7a76\u65b9\u5411 \u9304\u97f3\u88dd\u7f6e\u53ef\u80fd\u56e0\u70ba\u5e74\u4e45\u4f7f\u7528\u4e0b\u800c\u9020\u6210\u5668\u6750\u7684\u640d\u8017\uff0c\u7136\u800c\u9019\u985e\u7684\u640d\u8017\u4ee5\u4e0d\u5f71\u97ff\u4eba\u8033\u80fd\u5920\u8b58 \u5225\u7684\u689d\u4ef6\u4e0b\u4e26\u4e0d\u6703\u88ab\u66f4\u63db\uff0c\u70ba\u4e86\u80fd\u5920\u4fdd\u6709\u9304\u97f3\u7684\u4fbf\u6377\u6027\u53ca\u5176\u771f\u5be6\u6027\uff0c\u5247\u4f7f\u7528\u73fe\u4ee3\u4eba\u90fd\u5177 Testing Mode \u7121\u52a0\u6210\u6027\u566a\u97f3 \u52a0\u6210\u6027\u566a\u97f3: \u9910\u5ef3(\u53f0\u5317\u5927\u5b78)\u3001\u706b\u8eca\u7ad9(\u677f\u6a4b)\u3001\u6377\u904b\u7ad9 (\u677f\u6a4b)\u3001\u516c\u8eca\u7ad9\u724c(\u677f\u6a4b) SNR: 0\u30015\u300110 dB 2 10 2 10log s n GSNR s s ae \u00f6 = \u00e8 \u00f8 \u00e7 \u00f7 (1-2)</td></tr><tr><td colspan=\"7\">\u672c\u7814\u7a76\u8003\u616e\u8a9e\u97f3\u7aef\u9ede\u5075\u6e2c\u5c0d\u65bc ASR \u7cfb\u7d71\u7684\u5f71\u97ff\uff0c\u63d0\u51fa\u4e86\u4ee5 NN-based \u7684\u6aa2\u6e2c\u6cd5\u61c9\u7528\u65bc\u7814 \u5099\u7684\u667a\u6167\u578b\u624b\u6a5f\u4f86\u9304\u88fd\u5176\u8a9e\u6599\u5eab\uff0c\u4ee5\u9032\u884c\u5728\u591a\u74b0\u5883\u4e0b\u7684\u8a9e\u97f3\u7aef\u9ede\u6e2c\u8a66\u3002 Development Mode \u7121\u52a0\u6210\u6027\u566a\u97f3 \u52a0\u6210\u6027\u566a\u97f3: \u9910\u5ef3(\u53f0\u5317\u5927\u5b78)\u3001\u706b\u8eca\u7ad9(\u677f\u6a4b)\u3001\u6377\u904b\u7ad9 (\u677f\u6a4b)\u3001\u516c\u8eca\u7ad9\u724c(\u677f\u6a4b) \u5176\u4e2d ' ( \u4ee5\u53ca ) ( \u53ef\u5206\u5225\u70ba\u8a9e\u97f3\u8a0a\u865f\u7684\u529f\u7387\u4ee5\u53ca\u96dc\u8a0a\u7684\u529f\u7387\uff0c ' ( \u53ef\u7531\u5f0f(1-3)\u8a08\u7b97\uff1a</td></tr><tr><td colspan=\"7\">\u8868\u56db\uff1a\u53f0\u5317\u5927\u5b78\u96dc\u8a0a\u8a9e\u6599\u5eab\u7d30\u9805\u8aaa\u660e 0 0 t t \u985e\uff0c\u76f8\u95dc\u8aaa\u660e\u5982\u4e0b: \u5229\u7528 MLP\u3001DNN \u985e\u795e\u7d93\u7db2\u8def\uff0c\u5c0d\u53f0\u5317\u5927\u5b78\u96dc\u8a0a\u8a9e\u6599\u5eab\u505a\u8a13\u7df4\u53ca\u6e2c 2\u3001\u9304\u97f3\u8a08\u756b\u53ca\u5167\u5bb9 SNR: 0\u30015\u300110 dB 2 2 ( ) ( ) s t s p e e c h s p e e c h s t t s d d \u7a76\u4e2d\uff0c\u4e26\u81ea\u884c\u5efa\u7acb\u53f0\u5317\u5927\u5b78\u96dc\u8a0a\u8a9e\u6599\u5eab\u3002\u6b64\u8a13\u7df4\u6cd5\u53ef\u7528\u4f86\u5c0d\u8a13\u7df4\u8cc7\u6599\u505a\u6a21\u578b\u7684\u5efa\u7acb\u53ca\u5206 1 1 T T --</td></tr><tr><td colspan=\"7\">\u8a66\uff0c\u5404\u5225\u5efa\u7acb\u8a9e\u97f3\u4ee5\u53ca\u975e\u8a9e\u97f3\u6a21\u578b\uff0c\u5c07\u500b\u5225\u985e\u795e\u7d93\u7db2\u8def\u8f38\u51fa\u4e4b\u7d50\u679c\uff0c\u7528\u4f86\u63a2\u8a0e\u8a9e\u97f3\u7aef\u9ede \u53f0\u5317\u5927\u5b78\u96dc\u8a0a\u8a9e\u6599\u5eab\u6240\u9304\u88fd\u7684\u5167\u5bb9\u70ba\u591a\u7a2e\u74b0\u5883\u4e0b\u7684\u96dc\u8a0a\u8a9e\u6599\u5eab\uff0c\u6240\u6709\u8a9e\u6599\u6a94\u6848\u5747\u4ee5 \u96dc\u8a0a\u7a2e\u985e SNR \u7a2e\u985e \u8a9e\u8005\u7e3d\u6578 \u7537\u5973\u6bd4 Utterance Utterance length \u6a94\u6848\u7e3d\u6578</td></tr><tr><td colspan=\"7\">\u5075\u6e2c\u7684\u8868\u73fe\u3002\u5f9e ASR \u7cfb\u7d71\u8207\u79d1\u5b78\u7684\u89d2\u5ea6\u51fa\u767c\uff0c\u5c0d\u61c9\u7528\u65bc ASR \u7cfb\u7d71\u4e4b\u8a9e\u97f3\u7aef\u9ede\u5075\u6e2c\u5206 Restaurant 131 \u4eba 65 : 66 131 \u53e5 50:48.96 393 \u7b46 Sampling Rate: 16kHz\u3001Sound Encoding: Lin16 \u53ca Channel: 1 \u7684 PCM \u683c\u5f0f\u8a2d\u5b9a\u9032\u884c\u9304\u88fd\uff0c 0\u30015\u300110 Train Station 157 \u4eba 78 : 79 157 \u53e5 1:00:22.40 471 \u7b46</td></tr><tr><td colspan=\"7\">\u6790\uff0c\u4e26\u8a0e\u8ad6\u54ea\u7a2e\u795e\u7d93\u7db2\u8def\u4e4b\u67b6\u69cb\u6216\u65b9\u6cd5\u66f4\u9069\u5408\u904b\u7528\u81f3 ASR \u7cfb\u7d71\u3002 \u4e26\u5c07\u97f3\u6a94\u5132\u5b58\u6210*.wav \u6a94\u6848\u683c\u5f0f\u3002\u9304\u97f3\u88dd\u7f6e\u4f7f\u7528 HTC Desire \u4e26\u5229\u7528\u5be6\u9a57\u5ba4\u7684 Android \u9304 dB MRT 135 \u4eba 67 : 68 135 \u53e5 52:12.54 405 \u7b46</td></tr><tr><td colspan=\"7\">\u5728\u904e\u53bb\u7684\u7814\u7a76\u4e2d\uff0c\u767c\u73fe\u4ee5 Energy-based \u8207 GMM-based \u4e4b\u8a9e\u97f3\u7aef\u9ede\u5075\u6e2c\uff0c\u5c0d\u65bc ASR Bus Stop 160 \u4eba 80 : 80 160 \u53e5 1:01:08.35 480 \u7b46 \u97f3\u7a0b\u5f0f\u9032\u884c\u9304\u97f3\u3002 Clean \u221e dB 160 \u4eba 98 : 62 160 \u53e5 52:45.04 160 \u7b46</td></tr><tr><td colspan=\"7\">\u7cfb\u7d71\u7684\u8868\u73fe\u5176\u6548\u679c\u6709\u9650\u3002\u672c\u8ad6\u6587\u4ee5 NN-based \u7684\u65b9\u6cd5\uff0c\u63d0\u51fa\u4e00\u8a9e\u97f3\u7aef\u9ede\u5075\u6e2c\u4e4b\u6280\u8853\u4e26\u63a2 \u96dc\u8a0a\u8cc7\u6599\u5eab\u5171\u5206\u6210 4 \u500b\u985e\u5225\u70ba\u53f0\u5317\u5927\u5b78\u5b78\u6821\u9910\u5ef3\u3001\u677f\u6a4b\u706b\u8eca\u7ad9\u3001\u677f\u6a4b\u6377\u904b\u7ad9\u3001\u677f\u6a4b</td></tr><tr><td colspan=\"7\">\u8a0e\u4e0d\u540c\u7a2e\u985e\u7684\u8f38\u5165\u8cc7\u6599\u5c0d\u65bc\u8a9e\u97f3\u7aef\u9ede\u5075\u6e2c\u4e4b\u5f71\u97ff\u3002 \u516c\u8eca\u7b49\u5019\u4ead\u4f86\u9032\u884c\u9304\u88fd\u3002\u6bcf\u500b\u985e\u5225\u7686\u6703\u9304\u88fd\u8fd1 60 \u5206\u9418\u9577\u5ea6\u7684\u97f3\u6a94\uff0c\u8a9e\u6599\u5eab\u9304\u88fd\u8005\u70ba 4\u3001Noise Speech \u4e4b\u5efa\u7acb\u65b9\u6cd5</td></tr><tr><td>ycdeng\u3002</td><td/><td/><td/><td/><td/><td/></tr><tr><td colspan=\"2\">\u4e8c\u3001\u8a9e\u6599\u5eab\u7c21\u4ecb</td><td colspan=\"3\">\u8868\u4e8c\uff1a\u53f0\u5317\u5927\u5b78\u96dc\u8a0a\u8a9e\u6599\u5eab\u8cc7\u8a0a\u7d71\u8a08\u8868</td><td/><td/></tr><tr><td>\u96dc\u8a0a\u7a2e\u985e</td><td>\u5730\u9ede</td><td>\u65e5\u671f</td><td>\u6642\u9593</td><td>\u9304\u88fd\u8005</td><td>\u88dd\u7f6e</td><td>\u96dc\u8a0a\u9577\u5ea6</td></tr><tr><td colspan=\"7\">\u672c\u8ad6\u6587\u4f7f\u7528\u4e86 TCC300 \u8a9e\u6599\u5eab[13]\u3002TCC300 \u8a9e\u6599\u5eab\u662f\u7531\u570b\u7acb\u53f0\u7063\u5927\u5b78\u3001\u570b\u7acb\u4ea4\u901a\u5927\u5b78\u3001 \u6377\u904b\u7ad9 \u6377\u904b\u677f\u5357\u7dda 12/26/2016 20:11~21:03 ycdeng HTC Desire 52:12.54 \u6bb5\u4e2d\u51fa\u73fe\u8aa4\u5224(false reject)\u3002 (\u4e00) \u3001TCC300 \u8a9e\u6599\u5eab \u5b78\u6821\u9910\u5ef3 \u53f0\u5317\u5927\u5b78 12/27/2016 11:54~12:44 ycdeng HTC Desire 50:48.96 \u706b\u8eca\u7ad9 \u677f\u6a4b\u706b\u8eca\u7ad9 12/26/2016 19:02~20:02 ycdeng HTC Desire 1:00:22.40 \u8a9e\u6599\u5eab\u97f3\u6a94\u7684\u8a9e\u97f3\u53ca\u975e\u8a9e\u97f3\u4e4b\u6bb5\u843d\uff0c\u4ee5\u4fbf\u65bc\u7528\u4f86\u8a08\u7b97\u4f7f\u7528\u65bc\u6df7\u97f3\u6240\u9700\u8981\u7684 SNR \u8cc7\u8a0a\u3002</td></tr><tr><td colspan=\"7\">GMM-based \u6aa2\u6e2c\u6cd5\uff0c\u6b64\u65b9\u6cd5\u4e3b\u8981\u662f\u4f9d\u64da\u8a9e\u97f3\u5167\u5bb9\u70ba\u57fa\u790e\u7684\u975e\u76e3\u7763\u5f0f\u8a13\u7df4\u6cd5\uff0c\u9700\u8981 \u570b\u7acb\u6210\u529f\u5927\u5b78\u5404\u81ea\u64c1\u6709\u4e4b\u8a9e\u6599\u5eab\u96c6\u5408\u800c\u6210\uff0c\u5404\u6821\u9304\u88fd\u4e4b\u8a9e\u6599\u5eab\u7686\u5c6c\u65bc\u9ea5\u514b\u98a8\u662f\u6717\u8b80\u8a9e\u97f3\u3002 \u516c\u8eca\u7ad9\u724c \u677f\u6a4b\u516c\u8eca\u7ad9 12/26/2016 17:47~18:48 ycdeng HTC Desire 1:01:08.35 t s</td></tr><tr><td colspan=\"7\">\u5229\u7528 Threshold \u4f86\u5c0d\u8a9e\u97f3\u53ca\u975e\u8a9e\u97f3\u5efa\u7acb\u6a21\u578b\u5f8c\u505a\u5224\u65b7\u3002\u6211\u5011\u5c07\u6b64\u6aa2\u6e2c\u6cd5\u5229\u7528 TCC300 \u4e7e \u5176\u4e2d\u53f0\u5927\u8a9e\u6599\u5eab\u4e3b\u8981\u5305\u542b\u8a5e\u8a9e\u77ed\u53e5\uff0c\u6587\u672c\u7d93\u904e\u4ed4\u7d30\u8a2d\u8a08\u4e26\u8003\u616e\u4e86\u97f3\u7bc0\u53ca\u5176\u76f8\u9023\u51fa\u73fe\u6a5f\u7387\uff0c 3\u3001\u8a9e\u6599\u5eab\u4f7f\u7528 t x</td></tr><tr><td colspan=\"7\">\u6de8\u7684\u8a9e\u6599\u5eab\u5be6\u9a57\u5f8c\u53ef\u4ee5\u767c\u73fe\uff0c\u5728\u5be6\u9a57\u7d50\u679c\u7684 ROC (Receiver Operating Characteristics) \u7531 100 \u4eba\u9304\u88fd\u800c\u6210\u3002\u4ea4\u5927\u53ca\u6210\u5927\u8a9e\u6599\u5eab\u4e3b\u8981\u5305\u542b\u9577\u6587\u8a9e\u6599\uff0c\u6587\u7ae0\u7531\u4e2d\u7814\u9662\u63d0\u4f9b\u4e4b 500 \u842c</td></tr><tr><td colspan=\"7\">curve \u4e0a\u5176 EER (Equal Error Rate)\u7684\u8868\u73fe\u7d50\u679c\u4e26\u4e0d\u5982\u9810\u671f\u3002\u6545\u6b64\u6211\u5011\u5247\u4f7f\u7528 NN-based \u6aa2 \u8a5e\u8a5e\u985e\u6a19\u793a\u8a9e\u6599\u5eab\u6240\u9078\u53d6\uff0c\u6bcf\u7bc7\u6587\u7ae0\u5305\u542b\u6578\u767e\u5b57\u5728\u5207\u5272\u6210 3-4 \u6bb5\uff0c\u6bcf\u6bb5\u5305\u542b\u81f3\u5c11 231 \u5b57 \u5c07\u9304\u88fd\u5b8c\u7684\u96dc\u8a0a\u8cc7\u6599\u5eab\u8207 TCC300 \u8a9e\u6599\u5eab\u505a\u7d50\u5408\u4e26\u5f62\u6210\u52a0\u6210\u6027\u96dc\u8a0a\u3002\u5c07\u5176\u96a8\u6a5f\u5206\u6210 7:2:1 t n</td></tr><tr><td colspan=\"7\">\u6e2c\u6cd5\u65bc\u5be6\u9a57\u4e2d\uff0c\u5e0c\u671b\u80fd\u5f97\u5230\u66f4\u4f73\u7684\u7d50\u679c\u3002 \u7531 200 \u4eba\u6717\u8b80\u9304\u88fd\uff0c\u6bcf\u4eba\u6717\u8b80\u6587\u7ae0\u7686\u4e0d\u76f8\u540c\u3002 \u7684\u8a13\u7df4\u96c6\u3001\u6e2c\u8a66\u96c6\u8207\u767c\u5c55\u96c6\u4e09\u90e8\u5206\u9032\u884c\u5be6\u9a57\uff0c\u4e14\u6bcf\u4e00\u90e8\u5206\u7684\u52a0\u6210\u6027\u96dc\u8a0a\u74b0\u5883\u6bd4\u4f8b\u7686\u76f8\u540c\u3002 g</td></tr></table>" | |
| }, | |
| "TABREF1": { | |
| "num": null, | |
| "html": null, | |
| "text": "\u3002\u7136\u800c DNN \u5247\u662f\u5c07\u5176 hidden layer \u7684\u6578\u76ee\u589e\u52a0\uff0chidden layer \u5167\u7684 node \u6578\u76ee\u4e5f\u589e \u52a0\uff0c\u76ee\u7684\u662f\u8981\u8b93\u6574\u500b Neural network \u5f88\u6df1\u4e14\u5f88\u5bec\u3002\u5728\u6b64\u8655\u5247\u4e0d\u662f\u4ee5\u4e00\u822c\u4ecb\u7d39 DNN \u7684\u793a \u610f\u5716\u4f86\u505a\u8868\u793a\uff0c\u5982\u5716\u4e8c\u6240\u793a\u5247\u662f\u4f7f\u7528 Signal flow graph \u4f86\u63cf\u8ff0\u9019\u6a23\u7684\u7cfb\u7d71\u3002 \u5716\u4e8c\uff1aDNN signal flow graph \u5176\u4e2d", | |
| "type_str": "table", | |
| "content": "<table><tr><td colspan=\"2\">\u5716 2 \u4e4b\u6df7\u97f3\u8a08\u7b97\u6cd5\u662f\u4f7f\u7528\u4e86\u6539\u826f\u5f0f\u7684\u6df7\u97f3\u8a08\u7b97\u6cd5[15]\u3002\u76f8\u8f03\u65bc\u4ee5\u5f80\u7684\u6df7\u97f3\u6f14\u7b97\u6cd5\uff0c\u5247\u662f</td></tr><tr><td colspan=\"2\">\u5c07\u591a\u500b\u8f38\u5165\u6578\u64da\u505a\u7dda\u6027\u758a\u52a0\u7684\u65b9\u5f0f\uff0c\u53ef\u4ee5\u5f88\u660e\u986f\u5730\u807d\u5230\u80cc\u666f\u96dc\u8a0a\u3001\u6ce2\u5f62\u6703\u7a81\u8b8a\u5931\u771f\u4e26\u4e14 \u51fa\u73fe\u6bd4\u8f03\u8f15\u5fae\u5f97\u7206\u97f3\uff0c\u9020\u6210\u5c11\u6578\u7684\u8a9e\u97f3\u4e26\u7121\u6cd5\u8fa8\u8b58\u4e26\u4e14\u6703\u7522\u751f\u6ea2\u51fa\u7684\u554f\u984c\u3002\u6539\u5584\u65b9\u6cd5\uff1b \u5c0d\u65bc\u4ee5\u5f80\u7684\u6df7\u97f3\u8a08\u7b97\u6cd5\u4f86\u8aaa\u662f\u4f7f\u7528\u66f4\u591a\u7684\u4f4d\u5143\u6578\u4f86\u8868\u793a\u5176\u97f3\u6a94\u7684\u4e00\u500b sample\uff0c\u5728\u6df7\u97f3 \u5f8c\u964d\u4f4e\u5176\u632f\u5e45\u4e26\u4f7f\u5176\u5206\u5e03\u5728 16bit \u6240\u80fd\u8868\u5f0f\u7684\u7bc4\u570d\u5167\uff0c\u6b64\u7a2e\u65b9\u6cd5\u70ba Normalize \u505a\u6cd5[16]\uff0c \u4f46\u7f3a\u9ede\u5247\u662f\u6df7\u97f3\u5f8c\u7684\u8072\u97f3\u975e\u5e38\u5c0f\u4e14\u5176\u6548\u679c\u4e0d\u898b\u7406\u60f3\u3002\u4f46\u6539\u826f\u5f0f\u6df7\u97f3\u7b97\u6cd5\u89e3\u6c7a\u6ea2\u51fa\u7684\u65b9\u6cd5 \u5247\u662f\u7b9d\u4f4d (clamping)\uff0c\u7b9d\u4f4d\u4ee5\u4e0a\u7684\u503c\u70ba\u6240\u80fd\u8868\u5f0f\u7684\u6700\u5927\u503c\uff0c\u7576\u767c\u751f\u4e0b\u6ea2\u4f4d\u6642\u5247\u7b9d\u4f4d\u5e73\u79fb \u5f8c\u70ba\u6240\u80fd\u8868\u5f0f\u7684\u6700\u5c0f\u503c\u5982\u4e0b\u5f0f , , , t t t t MAX x MAX x MIN x MIN x otherwise > \u00ec \u00ef \u00ac < \u00ed \u00ef \u00ee (1-7) \u4e09\u3001NN-based VAD \u65b9\u6cd5 (\u4e00) \u3001Deep Neural Network (DNN) Deep Learning \u7684\u6982\u5ff5\u662f\u53ef\u4ee5\u8b93\u5404\u500b\u6a21\u7d44\u51fd\u6578\u7d93\u904e\u7dda\u6027\u6216\u662f\u975e\u7dda\u6027\u7684\u7d44\u5408\u5f8c\u80fd\u6709\u5177\u6709 end-to-end global optimization \u7279\u6027\uff0c\u5176\u4e2d\u6700\u5177\u4ee3\u8868\u6027\u7684 Deep Learning \u662f\u5176\u63a8\u5c0e\u51fa\u7684 Deep Neural Network(DNN)[9][17][18] \u3002 \u82e5 \u5f9e \u5176 \u67b6 \u69cb \u4f86 \u770b \uff0c DNN \u8207 \u50b3 \u7d71 \u7684 Multilayer Perceptron(MLP)\u662f\u76f8\u540c\u7684\uff0c\u4f46\u662f\u50b3\u7d71\u7684 MLP \u5927\u591a\u5c31\u53ea\u6709\u4f7f\u7528\u5230\u4e09\u5c64\u7684\u67b6\u69cb\u4f86\u9032\u884c\uff0c\u9019 \u4e09\u5c64\u5206\u5225\u662f\u4e00\u500b\u8f38\u5165\u5c64(input layer x)\u3001\u4e00\u500b\u96b1\u85cf\u5c64(hidden layer)\u4ee5\u53ca\u4e00\u500b\u8f38\u51fa\u5c64(output 1 1 1 1 1 1 1 1 1 [ ] ( [ ]) ( [ ] ) [ ] ( [ ]) ( [ ] ), 2[ ] ( [ ]) ( [ ] ) K k k K y y y h y k k k k k k k h h xh n F n n n n n k K n n n s s s s s --\u00ec = = + \u00ef \u00ef = = + = \u00ed \u00ef = = + \u00ef \u00ee y x W h I b h z W h Ib h z W x Ib (4-1) (4-1)\u5f0f\u4e2d\u7684\u6fc0\u767c\u51fd\u6578 ( ) s \u00d7 \u53ef\u4ee5\u662f element-wise \u7684 Sigmoid\u3001Hyperbolic\u3001Linear\u3001Rectified linear functions\uff0c\u800c\u8a13\u7df4\u6574\u500b DNN \u7684 criterion \u53ef\u4ee5\u662f Minimum mean squared error(MMSE) \u6216\u662f Maximum likelihood(ML)\uff1b\u5176\u4e2d ML \u7684 criterion \u5728\u9810\u4f30\u76ee\u6a19\u662f\u4ee5 category \u7684\u60c5\u6cc1\u4e0b \u5c31\u7b49\u540c\u65bc Minimum cross entropy(MCE)\u7684\u689d\u4ef6\uff0c\u6839\u64da\u4ee5\u4e0a\u7684\u689d\u4ef6\uff0c\u53ef\u4ee5\u5229\u7528\u4ee5\u4e0b\u7684\u6578\u5b78 \u5f0f\u4f86\u8868\u793a DNN \u7684\u8a13\u7df4\u904e\u7a0b * * , 2 1 0 2 1 0 , arg min ( , ) [ ] ( [ ]) for MMSE ( , )( [ ]) log( ( [ ])) for MCE N n N T n J n F n J n F n -= -= = \u00ec -\u00ef = \u00ed \u00ef-\u00ee \u00e5 \u00e5 W b W b W b y x W b y x (4-2) layer y)\u7684\u6578\u5b78\u95dc\u4fc2\u5f0f \u5176\u4e2d\u02c6[ ] y n \u4ee3\u8868\u662f\u7b2c n \u500b\u7684 input sample [ ] x n \u6240\u5c0d\u61c9\u5230\u7684\u6b63\u78ba\u7b54\u6848(Reference)\uff0c\u800c\u9019\u6a23\u7684</td></tr><tr><td/><td>\u4ee3\u8868\u539f\u59cb\u96dc\u8a0a\u6bb5\u843d</td></tr><tr><td colspan=\"2\">\u7684 noise power\u3002\u70ba\u4e86\u8981\u4f7f\u6df7\u97f3\u7684 noisy speech \u4e4b GSNR \u7b26\u5408\u5be6\u9a57\u7684\u8981\u6c42\u503c\uff0c\u6211\u5011\u5fc5\u9808\u8abf</td></tr><tr><td>\u6574 g \u7684\u503c\u5982\u5f0f(1-6)\u6240\u793a</td><td/></tr><tr><td>g</td><td>20 SNR 10\u011c ae \u00e7 \u00e8 =\u00b4(1-6) s n s s \u00f6 \u00f7 \u00f8</td></tr></table>" | |
| }, | |
| "TABREF2": { | |
| "num": null, | |
| "html": null, | |
| "text": "(\u4e8c) \u3001NN-based VAD \u5be6\u9a57\u8a2d\u5b9a \u672c\u5be6\u9a57\u6240\u4f7f\u7528\u7684\u8a9e\u6599\u5eab\u70ba\u8868 1.3 \u6240\u793a\uff0c\u81ea\u884c\u9304\u88fd\u96dc\u8a0a\u53ca\u6df7\u97f3\u7684\u53f0\u5317\u5927\u5b78\u96dc\u8a0a\u8a9e\u6599\u5eab\u3002\u5176 \u7279\u5fb5\u53c3\u6578\u4f7f\u7528\u4e86 12 \u7dad\u5ea6\u7684 MFCC \u518d\u52a0\u4e0a 1 \u7dad\u5ea6\u4e4b energy\uff0cMFCC \u8a2d\u5b9a\u88e1\u5247\u662f\u4f7f\u7528\u4e86 24 \u500b filter bank \u4e26\u4e14\u5728 cepstrum \u88e1\u53d6\u524d 12 \u500b cosine \u4f86\u63cf\u8ff0\u5176\u6ce2\u5cf0\u7279\u6027\uff0cMFCC \u7279\u5fb5\u53c3\u6578 \u7684\u8a9e\u97f3\u7279\u5fb5\u53c3\u6578\u4f86\u9810\u6e2c frame n \u7684 VAD \u72c0\u614b\uff0cn \u4ee3\u8868\u76ee\u524d\u9810\u4f30 VAD \u72c0\u614b\u7684 frame index\uff0cd \u8868\u793a\u70ba\u5f9e\u76ee\u524d\u6642\u9593\u9ede n \u6240 delay \u7684 frame \u6578\u76ee\uff0c\u800c l \u8868\u793a\u70ba\u5f9e\u76ee\u524d \u6642\u9593\u9ede n \u6240\u63d0\u524d\u7684 frame \u6578\u76ee\u3002NN-based \u5be6\u9a57\u53c3\u6578\u8a2d\u5b9a\u5982\u8868\u516d\u6240\u793a\u3002 \u7a2e\u554f\u984c\uff0c\u5982: feature frames\u3001layer \u6578\u76ee\u3001matchedcondition \u8207 multi-condition \u548c delay decision \u7684\u554f\u984c\u9032\u884c\u4e3b\u89c0\u8a0e\u8ad6\u3002NN-based \u7684 VAD \u7814 \u7a76\u65b9\u6cd5\u5be6\u505a\u65bc Tensorflow \u5e73\u53f0\u4e0a\uff0c\u5728\u7d66\u5b9a\u8f38\u5165\u4ee5\u53ca\u8f38\u51fa\u4e4b\u7b54\u6848\u5f8c\uff0c\u4f9d\u7167\u4e0d\u540c\u7684 NN \u67b6\u69cb \u4f86\u6c7a\u5b9a\u6bcf\u500b frame \u662f\u8a9e\u97f3\u9084\u662f\u975e\u8a9e\u97f3\u3002\u70ba\u4e86\u8981\u8207\u4e0d\u540c NN \u65b9\u6cd5\u505a\u6bd4\u8f03\uff0c\u672c\u7814\u7a76\u6311\u9078\u51fa\u8f03 \u5177\u4ee3\u8868\u6027\u4e4b\u60c5\u6cc1\u4f86\u63a2\u8a0e\u3002 1\u3001DNN \u4e0b feature frames \u7684\u8a0e\u8ad6 \u5716\u56db\u8868\u793a\u5169\u6975\u7aef\u60c5\u6cc1\u4e0b\uff0cDNN \u7684 feature frames \u7d50\u679c\u3002\u6a6b\u8ef8\u70ba feature frames \u6578\u76ee\u3001\u7e31\u8ef8 \u5206\u5225\u662f accuracy (Acc.)\u4ee5\u53ca EER\u3002\u5716\u56db\uff0c\u96a8\u8457 feature frames \u6578\u76ee\u7684\u4e0a\u5347\u5176 Acc.\u4ee5\u53ca EER \u5728 training set \u8207 validation set \u4e2d\u7684\u8868\u73fe\u6709\u660e\u986f\u8b8a\u597d\u7684\u8da8\u52e2\u3002\u4f46\u662f\u5728 outside test \u4e2d\u5176 EER \u9032\u6b65\u5e45\u5ea6\u8f03\u5c0f\u3002\u6545\u6b64\u63a8\u8ad6\u8aaa\uff0c\u7576 feature frames=8 \u6642\uff0c\u5c31\u5df2 over-trained\u3002 \u5716\u56db\uff1a\u8f03\u5177\u4ee3\u8868\u6027\u60c5\u6cc1\u4e0b DNN \u4e4b feature frames \u7d50\u679c 2\u3001DNN \u4e0b layer \u6578\u76ee\u7684\u8a0e\u8ad6 \u5716\u4e94\u8868\u793a\u5728 multi-condition \u4e2d\u5404\u500b\u60c5\u6cc1\u4e0b\uff0cDNN \u7684 layer \u6578\u76ee\u7d50\u679c\u3002\u6a6b\u8ef8\u70ba\u6bcf\u500b condition\u3001 \u7e31\u8ef8\u5206\u5225\u662f Acc.\u4ee5\u53ca EER\uff0c\u5176\u4e2d\u6a6b\u8ef8\u7684\u6bcf\u500b condition \u4f9d\u5e8f\u5206\u5225\u662f\uff1aClean \u70ba\u4e7e\u6de8\u3001BusStop \u70ba\u516c\u8eca\u7ad9\u3001MRT \u70ba\u6377\u904b\u7ad9\u3001TrainStat \u70ba\u706b\u8eca\u7ad9\u3001Rest.\u70ba\u9910\u5ef3\u3001Multi.\u70ba multi-condition \u7684\u60c5\u6cc1\uff0c\u800c\u5728 BusStop\u3001MRT\u3001TrainStat \u8207 Rest.\u7684\u60c5\u6cc1\u7576\u4e2d\u53c8\u5305\u542b\u6709 snr=0, 5, 10 (dB)\u4e4b \u7d50\u679c\u3002\u5206\u6790\u5b8c\u5716\u4e94\u5f8c\u53ef\u4ee5\u5f97\u5230\u7684\u7d50\u8ad6\u662f\uff0c\u5c07\u6bcf\u500b condition \u7684\u8cc7\u6599\u5408\u4f75\u6210 multi-condition \u5f8c\uff0c\u89e3\u6c7a\u4e86\u8cc7\u6599\u91cf\u4e0d\u8db3\u7684\u60c5\u6cc1\u3002\u6211\u5011\u5148\u5f9e DNN \u7684 layer \u6578\u76ee\u4f86\u89c0\u5bdf\u6bcf\u500b layer \u9593\u5f7c\u6b64\u7684 \u95dc\u4fc2\uff0c\u53ef\u4ee5\u5f97\u5230\u7684\u7d50\u679c\u662f\u96a8\u500b layer \u6578\u76ee\u7684\u4e0a\u5347\uff0c\u5176 Acc.\u8207 EER \u5728 training set\u3001validation set \u8207 outside test \u4e2d\u7684\u8868\u73fe\u6709\u8b8a\u597d\u4e4b\u8da8\u52e2\uff0c\u5c24\u5176\u662f\u5728\u5404\u500b condition \u4e2d snr=0 (dB)\u7684\u6642\u5019\u3002 \u6545\u6b64\u63a8\u8ad6\u8aaa\u5728 hidden layer \u8d8a\u6df1\u6642\uff0c\u6bcf\u500b condition \u53ef\u4ee5\u4e92\u76f8\u5b78\u7fd2\u5404\u500b condition \u9593\u5171\u540c \u7684\u7279\u6027\u3002\u4f46\u662f\u5728 2 layers \u8207 3 layers \u6642\u7684 Acc.\u8207 EER \u9032\u6b65\u5e45\u5ea6\u8f03\u5c0f\uff0c\u5176\u539f\u56e0\u662f\u5df2 over-trained\u3002 \u5716\u4e94\uff1amulti-condition \u8207 multi-condition \u4e2d\u5404\u500b\u60c5\u6cc1\u4e4b DNN layer \u6578\u76ee\u7d50\u679c\uff0c\u5176\u4e2d BusStop \u70ba\u516c\u8eca\u7ad9\u3001MRT \u70ba\u6377\u904b\u7ad9\u3001TrainStat \u70ba\u706b\u8eca\u7ad9\u3001Rest.\u70ba\u9910\u5ef3\u3001Multi.\u70ba multi-condition\uff0c \u800c\u5728 BusStop\u3001MRT\u3001TrainStat\u3001Rest.\u60c5\u6cc1\u7576\u4e2d\u53c8\u5305\u542b\u6709 snr=0, 5, 10 (dB)\u7684\u7d50\u679c 3\u3001matched-condition \u8207 multi-condition \u7684\u8a0e\u8ad6 \u5716\u516d\u8868\u793a matched-condition \u8207 multi-condition \u7684\u7d50\u679c\uff0c\u662f\u5728\u6bcf\u500b\u60c5\u6cc1\u4e0b\u6311\u9078\u51fa\u6700\u597d\u7684 layer \u6578\u76ee\u4f86\u505a\u8a0e\u8ad6\u3002\u6a6b\u8ef8\u70ba\u6bcf\u500b condition\u3001\u7e31\u8ef8\u5206\u5225\u662f Acc.\u4ee5\u53ca EER\uff0c\u5176\u4e2d\u6a6b\u8ef8\u7684\u6bcf \u500b condition \u4f9d\u5e8f\u5206\u5225\u662f\uff1aClean \u70ba\u4e7e\u6de8\u3001BusStop \u70ba\u516c\u8eca\u7ad9\u3001MRT \u70ba\u6377\u904b\u7ad9\u3001TrainStat \u70ba\u706b\u8eca\u7ad9\u3001Rest.\u70ba\u9910\u5ef3\u3001Multi.\u70ba multi-condition \u7684\u60c5\u6cc1\uff0c\u800c\u5728 BusStop\u3001MRT\u3001TrainStat \u8207 Rest.\u7684\u60c5\u6cc1\u7576\u4e2d\u53c8\u5305\u542b\u6709 snr=0, 5, 10 (dB)\u4e4b\u7d50\u679c\u3002\u7531\u5716\u516d\u53ef\u4ee5\u5f97\u5230\u4ee5\u4e0b\u4e4b\u7d50\u8ad6\uff0c\u5728 training set \u8207 validation set \u4f86\u89c0\u5bdf Acc.\u8207 EER \u6642\uff0c\u53ef\u4ee5\u767c\u73fe\u5728\u6700\u597d\u8a2d\u5b9a\u4e0b overall \u7684 multi-condition \u7d50\u679c\uff0c\u662f\u512a\u65bc\u5728 matched-condition \u4e0b\u6700\u7cdf\u60c5\u6cc1\u7684 Rest.\u7d50\u679c\u3002\u4f46\u662f\u5f9e outside test \u4f86\u89c0\u5bdf multi-condition \u88e1\u5404\u500b condition \u7684\u7d50\u679c\u5f8c\uff0c\u53ef\u4ee5\u767c\u73fe\u5176\u7d50\u679c\u662f\u512a\u65bc matchedcondition \u4e4b\u7d50\u679c\uff0c\u5c24\u5176\u662f\u5728 TrainStat snr=0 (dB)\u6642\u9032\u6b65\u5e45\u5ea6\u6700\u70ba\u660e\u986f\u3002\u6545\u6b64\u53ef\u4ee5\u63a8\u8ad6\u8aaa\uff0c \u56e0\u70ba\u5728 multi-condition \u88e1\u7684 hidden layer \u80fd\u5920\u5b78\u7fd2\u5230\u4e0d\u540c condition \u7684\u7279\u6027\uff0c\u6240\u4ee5\u5c0d\u65bc\u4e0d DNN \u7684 frames \u554f\u984c\u6703\u96a8\u8457 layer \u6578\u76ee\u7684\u589e\u52a0\uff0c\u800c\u4f7f Acc.\u8207 EER \u7684\u8868\u73fe\u6709\u8b8a\u597d\u3001(2)DNN \u7684 layer \u6578\u76ee\u554f\u984c\uff0c\u5728 matched-condition \u7684\u7d50\u679c\u4e26\u672a\u96a8\u8457 layer \u6578\u76ee\u7684\u4e0a\u5347\u800c\u4f7f Acc.\u53ca EER \u6709\u8b8a\u597d\u4e4b\u8da8\u52e2\uff0c\u6545\u63a8\u8ad6\u5176\u539f\u56e0\u662f\u5728\u65bc\u8a13\u7df4\u7684\u8cc7\u6599\u91cf\u4e0d\u8db3\u6240\u9020\u6210\uff0c\u6216\u662f\u6709\u4e9b\u8a9e\u97f3\u4e2d\u91cd \u8981\u4e4b\u7279\u6027\u5728\u8f49\u63db\u6210 MFCC \u53c3\u6578\u7684\u904e\u7a0b\u4e2d\uff0c\u5c31\u88ab\u5ffd\u7565\u4e86\u3002\u5728 multi-condition \u7684\u7d50\u679c\u4e2d\u53ef\u4ee5 \u767c\u73fe\u96a8\u8457 layer \u6578\u76ee\u7684\u4e0a\u5347\uff0c\u5176 Acc.\u8207 EER \u7684\u8868\u73fe\u5728\u5404\u500b set \u4e2d\u6709\u8b8a\u597d\u7684\u8da8\u52e2\uff0c\u6240\u4ee5\u53ef \u4ee5\u63a8\u8ad6\u5176\u539f\u56e0\u662f\u96a8\u8457 hidden layer \u6578\u76ee\u8d8a\u6df1\u6642\uff0c\u6bcf\u500b condition \u53ef\u4ee5\u4e92\u76f8\u5b78\u7fd2\u5404\u500b condition \u9593\u5171\u540c\u7684\u7279\u6027\u3001(3)matched-condition \u8207 multi-condition \u7684\u554f\u984c\uff0c\u5728 multi-condition \u4e4b performance \u512a\u65bc matched-condition (MRT \u8207 Rest. condition) \uff0c\u6240\u4ee5\u7531\u6b64\u53ef\u63a8\u8ad6\u51fa\u5728 multicondition \u4e2d\u7684 hidden layer \u80fd\u5920\u5b78\u7fd2\u5230\u4e0d\u540c condition \u4e4b\u7279\u6027\uff0c\u660e\u986f\u5c55\u73fe\u4e86\u6df1\u5ea6\u5b78\u7fd2\u7684\u512a \u52e2\u3002 \u4e94\u3001\u53c3\u8003\u8cc7\u6599 [1] Deng, C. Z. (2007, September). Voice Activity Detection and Keyword Spotting System on Embedded Platform, National Chiao Tung University, Hsinchu.", | |
| "type_str": "table", | |
| "content": "<table><tr><td colspan=\"3\">\u62bd\u53d6\u4e4b\u53c3\u6578\u8a2d\u5b9a\u5982\u8868\u4e94\u6240\u793a\u3002\u97f3\u6a94\u4e2d\u8a9e\u97f3\u53ca\u975e\u8a9e\u97f3\u4e4b label \u662f\u4f7f\u7528 HTK (Hidden Markov</td></tr><tr><td colspan=\"2\">Model Toolkit)\u4f86\u505a\u6a19\u8a18\u3002</td><td/></tr><tr><td/><td colspan=\"2\">\u8868\u4e94\uff1aMFCC \u7279\u5fb5\u53c3\u6578\u62bd\u53d6\u8a2d\u5b9a Config of MFCC Feature Extraction \u5716\u516d\uff1amatched-condition \u8207 multi-condition \u7684\u7d50\u679c</td></tr><tr><td/><td>SOURCEFORMAT</td><td>Alien</td></tr><tr><td/><td>HEADERSIZE</td><td>0</td></tr><tr><td>\u56db\u3001\u7d50\u8ad6</td><td>SOURCERATE</td><td>625.0</td></tr><tr><td/><td>TARGETKIND</td><td>MFCC_E</td></tr><tr><td colspan=\"3\">TARGETRATE WINDOWSIZE \u672c\u8ad6\u6587\u63a2\u8a0e\u5c07\u985e\u795e\u7d93\u7db2\u8def\u61c9\u7528\u65bc\u8a9e\u97f3\u7aef\u9ede\u5075\u6e2c\u4e2d\uff0c\u4f7f\u7528\u667a\u6167\u578b\u624b\u6a5f\u4f86\u9304\u88fd\u4e0d\u540c\u7a2e\u985e\u7684\u96dc 100000.0 320000.0 USEHAMMING T PREEMCOEF 0.97 NUMCHANS 24 CEPLIFTER 22 NUMCEPS 12 ENORMALISE F ZMEANSOURCE T \u70ba\u4e86\u8981\u6a21\u64ec\u5be6\u969b\u7cfb\u7d71\u7684\u8a9e\u97f3\u7aef\u9ede\u5075\u6e2c\uff0c\u6240\u4ee5\u672c\u8ad6\u6587\u4f7f\u7528\u5c07\u7279\u5fb5\u53c3\u6578 delay \u7684\u65b9\u5f0f\u3002\u5982\u5716 \u4e09\u6240\u793a\uff0c\u662f\u56e0\u70ba\u5728\u67d0\u500b\u6642\u9593\u9ede\u4e0b\u7684\u8a9e\u97f3\u80fd\u91cf(Energy)\u4e0a\u5347\u6216\u662f\u4e0b\u964d\u4e26\u7121\u6cd5\u7576\u4e0b\u5c31\u6c7a\u5b9a\u51fa \u662f\u5426\u70ba\u8a9e\u97f3\u9084\u662f\u96dc\u8a0a\uff0c\u9700\u8981\u5f80\u5f8c\u6216\u662f\u5f80\u524d\u591a\u770b\u5e7e\u500b frame \u4f86\u5224\u65b7\u7576\u4e0b\u7684 frame \u70ba\u8a9e\u97f3\u9084 \u662f\u96dc\u8a0a\uff1b\u63db\u53e5\u8a71\u8aaa\uff0c\u5247\u662f\u7531 frame n-l \u5230 frame n+d \u7684 (d-l+1)\u500b frame (\u4e5f\u5c31\u662f window \u8868\u516d\uff1aNN-based VAD \u5be6\u9a57\u8a2d\u5b9a NN \u7a2e\u985e DNN Windows size of feature frames (d+l+1) 1\u30013\u30015\u30017\u30019\u300111 -frame \u5be6\u9a57\u8a2d\u5b9a Optimizer Adam Batch_size 64 Nb_epoch 1500 Data set Train (7) : Validation (2) : Test (1) Earlystopping patience 50 Activation function ReLU Loss function categorical crossentropy Node size Dropout 0.3 Output layer function Softmax (\u4e09) \u3001NN-based VAD \u5be6\u9a57\u7d50\u679c\u8207\u5206\u6790 \u672c\u5c0f\u7bc0\u5c07\u4e0d\u540c\u7a2e\u985e NN-based VAD \u4e4b 5 \u540c\u7684\u74b0\u5883\u8ddf\u60c5\u6cc1\u4e0b\u80fd\u5920\u66f4\u52a0\u5f37\u5065(Robustness)\u3002 256 \u8a0a\uff0c\u4e26\u4e14\u81ea\u884c\u6df7\u97f3\u51fa\u7279\u5b9a\u4e4b SNR \u7a2e\u985e\uff0c\u5728\u7531\u4e0d\u540c\u67b6\u69cb\u7684\u985e\u795e\u7d93\u7db2\u8def\u4f86\u505a\u5b78\u7fd2\u3002 \u7d93\u904e\u7814\u7a76\u8207\u5206\u6790\uff0c\u672c\u8ad6\u6587\u5728\u4e0d\u540c\u5be6\u9a57\u4e0b\u7684\u985e\u795e\u7d93\u7db2\u8def\u7d50\u679c\uff0c\u53ef\u4ee5\u5f97\u5230\u4ee5\u4e0b\u4e4b\u7d50\u8ad6\u662f\uff1a size of feature frame) \u5716\u4e09\uff1aWindows size of feature frames \u793a\u610f\u5716\uff0c(\u4e0a) \u8a9e\u97f3\u958b\u59cb\u3001(\u4e0b) \u8a9e\u97f3\u7d50\u675f \u8f38\u5165\u8cc7\u6599 NTPU Additive Noise Corpus (1)</td></tr></table>" | |
| } | |
| } | |
| } | |
| } |