| { | |
| "paper_id": "O17-1004", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T07:59:26.967547Z" | |
| }, | |
| "title": "", | |
| "authors": [], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "", | |
| "pdf_parse": { | |
| "paper_id": "O17-1004", | |
| "_pdf_hash": "", | |
| "abstract": [], | |
| "body_text": [], | |
| "back_matter": [], | |
| "bib_entries": { | |
| "BIBREF2": { | |
| "ref_id": "b2", | |
| "title": "Nonliner Acoustic Echo Cancellstion With 2nd Order Adaptive Volterra Filters", | |
| "authors": [ | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Stenger", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "L", | |
| "middle": [], | |
| "last": "Trautmann", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "R", | |
| "middle": [], | |
| "last": "Rabenstein", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1999, | |
| "venue": "IEEE Int. Conf. on Acoustics, Speech & Signal Procrssing(ICASSP)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "A. Stenger, L. Trautmann and R. Rabenstein, Nonliner Acoustic Echo Cancellstion With 2nd Order Adaptive Volterra Filters, IEEE Int. Conf. on Acoustics, Speech & Signal Procrssing(ICASSP), 1999.", | |
| "links": null | |
| }, | |
| "BIBREF4": { | |
| "ref_id": "b4", | |
| "title": "High-resolution frequency-wavenumber spectrum analysis", | |
| "authors": [ | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Capon", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1969, | |
| "venue": "Proc. IEEE", | |
| "volume": "57", | |
| "issue": "8", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "J. Capon, High-resolution frequency-wavenumber spectrum analysis, Proc. IEEE vol.57, no.8, 8,August 1969.", | |
| "links": null | |
| }, | |
| "BIBREF5": { | |
| "ref_id": "b5", | |
| "title": "Suiyang KhooandZhihong Man, Frequency-domain beamformers using conjugate gradient techniques for speech enhancement", | |
| "authors": [ | |
| { | |
| "first": "Shengkui", | |
| "middle": [], | |
| "last": "Zhao", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Douglas", | |
| "middle": [ | |
| "L" | |
| ], | |
| "last": "Jones", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2014, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Shengkui Zhao, Douglas L. Jones,Suiyang KhooandZhihong Man, Frequency-domain beamformers using conjugate gradient techniques for speech enhancement , 2014.", | |
| "links": null | |
| }, | |
| "BIBREF6": { | |
| "ref_id": "b6", | |
| "title": "MMSE STSA Based Techniques for Single channel, Electronics and Communication Department, Parul institute of engineering and technology", | |
| "authors": [ | |
| { | |
| "first": "Simit", | |
| "middle": [], | |
| "last": "Shah", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Roma", | |
| "middle": [], | |
| "last": "Patel", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2015, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Simit Shah and Roma Patel,MMSE STSA Based Techniques for Single channel, Electronics and Communication Department, Parul institute of engineering and technology, Vadodara, Gujarat india, 2015.", | |
| "links": null | |
| }, | |
| "BIBREF7": { | |
| "ref_id": "b7", | |
| "title": "MMSE STSA Based Techniques for Single channel, Electronics and Communication Department, Parul institute of engineering and technology", | |
| "authors": [ | |
| { | |
| "first": "Simit", | |
| "middle": [], | |
| "last": "Shah", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Roma", | |
| "middle": [], | |
| "last": "Patel", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2015, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Simit Shah and Roma Patel,MMSE STSA Based Techniques for Single channel, Electronics and Communication Department, Parul institute of engineering and technology, Vadodara, Gujarat india,2015.", | |
| "links": null | |
| }, | |
| "BIBREF8": { | |
| "ref_id": "b8", | |
| "title": "Improving Single Frequency Filtering based Voice Activity Detection (VAD) using Spectral Subtraction based Noise Cancellation", | |
| "authors": [], | |
| "year": 2016, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Improving Single Frequency Filtering based Voice Activity Detection (VAD) using Spectral Subtraction based Noise Cancellation ,Department of Electronics and Communications NMAM Institute of Technology, Karnataka State, India,2016", | |
| "links": null | |
| } | |
| }, | |
| "ref_entries": { | |
| "TABREF0": { | |
| "text": "\u7136\u800c\uff0c\u5728\u89c0\u770b\u96fb\u8996\u7bc0\u6642\uff0c\u56e0\u96fb\u8996\u7bc0\u76ee\u662f\u6301\u7e8c\u64ad\u653e\u4e14\u8072\u97f3\u901a\u5e38\u958b\u5f88\u5927\u8072\uff0c\u56de\u8072\u7d93\u904e\u91cd \u91cd\u53cd\u5c04\u5f8c\uff0c\u6b98\u97ff\u6301\u7e8c\u6642\u9593\u901a\u5e38\u5f88\u4e45\u3002\u800c\u4e14\uff0c\u96fb\u8996\u7bc0\u76ee\u7684\u539f\u59cb\u8072\u97f3\u8a0a\u865f\uff0c\u8ddf\u7d93\u904e\u5587\u53ed\u64ad\u653e \u51fa\u4f86\u5f8c\uff0c\u518d\u88ab\u9ea5\u514b\u98a8\u9663\u5217\u6536\u9304\u9032\u4f86\u7684\u8072\u97f3\u8a0a\u865f\u9593\uff0c\u901a\u5e38\u5448\u975e\u7dda\u6027\u95dc\u4fc2\u3002\u4f46\u662f\u50b3\u7d71 NLMS \u6ffe\u6ce2\u5668\u57fa\u672c\u4e0a\u662f\u7dda\u6027\u7cfb\u7d71\uff0c\u53ea\u6703\u8003\u616e\u8f38\u5165\u8ddf\u8f38\u51fa\u8a0a\u865f\u9593\u7684\u7dda\u6027\u95dc\u4fc2\uff0c\u7121\u6cd5\u8655\u7406\u975e\u7dda\u6027\u8a0a \u865f\u640d\u8017\u3002\u6b64\u5916\uff0c\u56e0\u70ba NLMS \u662f\u524d\u994b\u5f0f\u7cfb\u7d71\uff0c\u82e5\u7cfb\u7d71\u8981\u8003\u616e\u5f88\u9577\u7684\u6b98\u97ff\u6642\u9593\uff0c\u6703\u9700\u8981\u5f88 \u9577\u6703\u9700\u8981\u5f88\u9577\u7684\u7cfb\u7d71\u53c3\u6578\uff0c\u6240\u4ee5\u5e38\u6703\u5c0e\u81f4\u904b\u7b97\u904e\u4e45\uff0c\u4e0d\u5bb9\u6613\u6536\u6582\u3002 \u5c24\u5176\u662f\u7576 NLMS \u5728 time domain \u904b\u4f5c\uff0c\u4ee5 raw sample \u70ba\u8655\u7406\u55ae\u4f4d\u7684\u60c5\u6cc1\u4e0b\uff0c\u9019\u500b\u554f \u984c\u6703\u7279\u5225\u56b4\u91cd\u3002\u4f8b\u5982\u5728 16 kHz \u53d6\u6a23\u983b\u7387\u4e0b\uff0c\u82e5\u8981\u6db5\u84cb 0.25 \u79d2\u7684\u96fb\u8996\u56de\u97f3\uff0c\u5c31\u9700\u8981\u4e00\u500b \u6709 4096 \u500b taps \u7684\u9069\u61c9\u6027\u6ffe\u6ce2\u5668\u3002 \u56e0\u6b64\uff0c\u70ba\u4e86\u89e3\u6c7a\u9019\u554f\u984c\uff0c\u6240\u4ee5\u6211\u5011\u6539\u7528\u542b\u6709\u56de\u6388\u529f\u80fd\u7684\u6df1\u5ea6\u905e\u8ff4\u985e\u795e\u7d93\u7db2\u8def(deep recurrent neural networks, RNNs)[4]\uff0c\u56e0\u70ba RNN \u80fd\u628a\u8f38\u51fa\u518d\u56de\u994b\u56de\u4f86\u7576\u8f38\u5165\u53c3\u6578\uff0c\u56e0\u6b64 \u80fd\u4ee5\u8f03\u5c11\u7684\u53c3\u6578\uff0c\u8a18\u61b6\u8f03\u9577\u7684\u72c0\u614b\u6b77\u53f2\u8cc7\u8a0a\uff0c\u5e6b\u52a9\u8655\u7406\u96fb\u8996\u56de\u97f3\u554f\u984c\u5c07\u7cfb\u7d71\u6539\u6210\u800c\u4e14\uff0c \u5176\u4e2d\uff0c\u56e0\u9ea5\u514b\u98a8\u9663\u5217\u53d6\u6a23\u983b\u7387\u70ba 16Khz\uff0c\u6211\u5011\u5c07 FFT \u8a2d\u70ba 256\uff0c\u800c\u4e14\u7531\u65bc FFT \u8f49 \u63db\u5c0d\u5be6\u6578\u8a0a\u865f\u5177\u6709\u5c0d\u7a31\u6027\uff0c\u6240\u4ee5 RNN \u7684\u8f38\u5165\u53ea\u9700\u8981 128 \u9ede\u7684\u8cc7\u6599\u3002\u5b78\u7fd2\u5b8c\u7684\u983b\u57df\u8f38\u51fa \u8a0a\u865f\u53ef\u7528 IFFT \u53cd\u8f49\u70ba\u6642\u57df\u8a0a\u865f\uff0c\u518d\u4f7f\u7528 Overlap-Add \u4e00\u4e00\u758a\u52a0\u8d77\u4f86\uff0c\u5c31\u80fd\u9084\u539f\u6210\u4e00\u6574 \u5be6\u9a57\u4e00\uff0c\u6642\u57df NLMS \u8207 RNN \u96fb\u8996\u56de\u8072\u6d88\u9664\u5be6\u9a57 \u5716\u4e03\u3001\u5be6\u9a57\u4e00 NLMS \u8207\u6642\u57df RNN \u96fb\u8996\u56de\u8072\u6d88\u9664\u7cfb\u7d71\u7684 ERLE \u8207 MSE \u7d50\u679c\u5716 \u5be6\u9a57\u4e09\uff0cMVDR \u52a0 RNN \u96fb\u8996\u56de\u8072\u6d88\u9664\u5be6\u9a57 \u5be6\u9a57\u4e09\u5247\u5be6\u9a57\u5982\u5716\u4e8c\u6240\u793a\u4e4b\u5169\u7a2e\u4e0d\u540c MVDR \u8207 RNN \u6574\u5408\u67b6\u69cb\uff0c\u4e00\u70ba\u6e2c\u8a66\u6536\u97f3\u5f8c \u7d93\u904e MVDR \u5148\u6291\u5236\u96fb\u8996\u56de\u8072\uff0c\u518d\u7d93\u904e\u983b\u57df RNN \u904e\u6ffe\u96fb\u8996\u56de\u8072\uff0c\u53e6\u4e00\u70ba\u6536\u97f3\u5f8c\u5148\u7531\u983b \u57df RNN \u6d88\u9664\u56de\u8072\uff0c\u518d\u7528 MVDR \u589e\u5f37\u4eba\u8072\u3002\u5be6\u9a57\u6bd4\u8f03\u52a0\u5165 MVDR \u5f8c\u7684\u56de\u8072\u6d88\u9664\u80fd\u5426\u66f4 \u70ba\u512a\u8d8a\uff0c\u8ddf\u5169\u7a2e\u4e0d\u540c\u6574\u5408\u67b6\u69cb\uff0c\u90a3\u4e00\u500b\u66f4\u7a81\u51fa\u3002\u5be6\u9a57\u7d50\u679c\u5982\u5716\u4e5d\u6240\u793a\uff1a \u5716\u4e5d\u3001\u5be6\u9a57\u4e09 \u5404\u7a2e MVDR+RNN \u96fb\u8996\u56de\u8072\u6d88\u9664\u7cfb\u7d71\u7684 ERLE \u8207 MSE \u7d50\u679c\u5716 \u7531\u5716\u4e5d\u53ef\u4ee5\u770b\u51fa\u5148\u7528 MVDR \u6291\u5236\u96fb\u8996\u56de\u8072\uff0c\u518d\u7d93\u904e\u983b\u57df RNN \u904e\u6ffe\u96fb\u8996\u56de\u8072\uff0c\u5728 \u5404\u7a2e\u985e\u578b\u96fb\u8996\u7bc0\u76ee\u7684\u6548\u80fd\uff0c\u90fd\u6bd4\u5148\u7528\u983b\u57df RNN \u505a\u6d88\u9664\u96fb\u8996\u56de\u8072\uff0c\u518d\u7528 MVDR \u589e\u5f37\u4eba \u8072\u7684\u6548\u679c\u597d\u3002 4. \u5be6\u9a57\u56db\uff0cMVDR \u52a0\u983b\u57df RNN \u5728\u4e0d\u540c\u89d2\u5ea6\u96fb\u8996\u56de\u8072\u6d88\u9664\u5be6\u9a57 \u7531\u5be6\u9a57\u4e09\u53ef\u5f97\u77e5\uff0cMVDR+\u983b\u57df RNN \u5728\u5404\u7a2e\u56de\u8072\u6d88\u9664\u5be6\u73fe\u6bd4\u8f03\u4e2d\uff0c\u6709\u6700\u597d\u7684\u6548\u80fd\uff0c \u6211\u5011\u5c31\u4ee5 MVDR+\u983b\u57df RNN \u6e2c\u8a66\u4f7f\u7528\u8005\u5728\u4e0d\u540c\u89d2\u5ea6\u7684\u56de\u8072\u6d88\u9664\u6548\u80fd\uff0c\u4ee5\u4fbf\u5f97\u77e5\u6211\u5011\u63d0 \u51fa\u7684\u7cfb\u7d71\u80fd\u5426\u53ef\u4ee5\u5728\u4f7f\u7528\u8005\u5728\u4e0d\u540c\u4f4d\u7f6e\u6642\uff0c\u4e00\u6a23\u80fd\u63a5\u53d7\u5230\u4e7e\u6de8\u7684\u8a9e\u97f3\u3002\u5be6\u9a57\u6e2c\u8a66\u4f7f\u7528\u8005 \u5728 90 \u5ea6\u300160 \u5ea6\u300130 \u5ea6\u60c5\u5f62\uff0c\u5be6\u9a57\u7d50\u679c\u5982\u5716\u5341\u6240\u793a \u5716\u5341\u3001\u5be6\u9a57\u56db \u983b\u57df RNN \u5728\u4e0d\u540c\u89d2\u5ea6\u7684\u96fb\u8996\u56de\u8072\u6d88\u9664 ERLE \u8207 MSE \u6bd4\u8f03\u5716 \u7531\u5716\u5341\u7684 ERLE \u548c MSE \u6bd4\u8f03\u5716\uff0c\u53ef\u4ee5\u5f97\u77e5\u4e0d\u7ba1\u4f7f\u7528\u8005\u5728\u54ea\u4e00\u500b\u65b9\u4f4d\uff0c\u7cfb\u7d71\u90fd\u80fd\u8f38 \u51fa\u8f03\u70ba\u4e7e\u6de8\u4eba\u8072\u8a9e\u97f3\uff0c\u4e5f\u4ee3\u8868\u5b83\u80fd\u6307\u5411\u4f7f\u7528\u8005\u4f4d\u7f6e\u505a\u8a9e\u97f3\u589e\u5f37\u3002\u800c\u4e14\uff0c\u5f9e\u5716\u5341\u6211\u5011\u53ef\u4ee5 \u767c\u73fe\u5230\u5728 30 \u5ea6\u548c 60 \u5ea6\u6642\u6548\u679c\u6703\u6bd4\u539f\u672c\u7684\u6b63\u524d\u65b9\u9084\u8981\u597d\u3002\u6b64\u5be6\u9a57\u53ef\u8b49\u5be6\u6211\u5011\u63d0\u51fa\u7684\u65b9 \u6cd5\uff0c\u80fd\u5728\u4f7f\u7528\u8005\u5728\u4e0d\u540c\u89d2\u5ea6\u6642\u4f9d\u7136\u6709\u5f88\u597d\u7684\u6548\u7387\u3002 5. \u5be6\u9a57\u4e94\uff0cMVDR \u52a0\u983b\u57df RNN \u518d\u52a0\u4e0a MMSE \u5f8c\u8655\u7406\u96fb\u8996\u56de\u8072\u6d88\u9664\u5be6\u9a57 \u7531\u5be6\u9a57\u4e09\u53ef\u5f97\u77e5\uff0cMVDR+\u983b\u57df RNN \u5728\u5404\u7a2e\u56de\u8072\u6d88\u9664\u5be6\u73fe\u6bd4\u8f03\u4e2d\uff0c\u6709\u6700\u597d\u7684\u6548\u80fd\uff0c \u6240\u4ee5\u6211\u5011\u76f4\u63a5\u4ee5\u8aaa\u8a71\u8005\u5728\u6b63\u524d\u9762\u7684 MVDR+\u983b\u57df RNN \u8207 MVDR+\u983b\u57df RNN \u591a\u52a0\u4e00\u500b \u5f8c\u8655\u7406\u7684\u505a\u6bd4\u8f03\u3002\u5f8c\u8655\u7684\u65b9\u6cd5\u4f7f\u7528 MMSE-log-STSA \u65b9\u6cd5\uff0c\u770b\u80fd\u4e0d\u80fd\u66f4\u9032\u4e00\u6b65\u6ffe\u9664\u6389\u5269 \u9918\u7684\u96fb\u8996\u7bc0\u76ee\u56de\u8072\u3002\u5be6\u9a57\u7d50\u679c\u5982\u5716\u5341\u4e00\u6240\u793a\uff1a \u5716\u5341\u4e00\u3001\u5be6\u9a57\u4e94 \u52a0\u5165\u5f8c\u8655\u7406\u5f8c\u7684\u96fb\u8996\u56de\u8072\u6d88\u9664\u7cfb\u7d71\u7684 ERLE \u8207 MSE \u7d50\u679c\u5716 \u7531\u5716\u5341\u4e00\u5de6\u908a\u53ef\u4ee5\u770b\u5230\uff0c\u5c0d\u5404\u7a2e\u985e\u578b\u7684\u96fb\u8996\u7bc0\u76ee\u7684 ERLE \u503c\uff0c\u4f7f\u7528 MVDR+\u983b\u57df RNN\uff0bMMSE \u7684\u5f8c\u8655\u7406\u7e3d\u5e73\u5747\u70ba 11.754\uff0c\u6bd4\u6c92\u7d93\u5f8c\u8655\u7406\u7684 MVDR+\u983b\u57df RNN \u7684 11.023 \u9084\u8981\u597d\u3002", | |
| "html": null, | |
| "content": "<table><tr><td>\u8996\u56de\u8072\u6d88\u9664\u67b6\u69cb\u8207(\u4e09)MMSE-log-STSA \u5f8c\u8655\u7406\u65b9\u6cd5\u3002 \u6700\u4f73\u89e3\u3002 (\u4e09) MMSE log-STSA \u6700\u5c0f\u5747\u65b9\u8aa4\u5dee\u77ed\u6642\u8b5c\u5e45\u5ea6\u4f30\u8a08 (\u4e8c) \u5be6\u9a57\u60c5\u5883 ERLE =</td><td>(3.1)</td></tr><tr><td colspan=\"2\">\u5f8c\u8655\u7406\u7684\u90e8\u5206\u4f7f\u7528 MMSE[10]\u4f86\u505a\u66f4\u9032\u4e00\u6b65\u7684\u7684\u6d88\u96dc\u8a0a\u52d5\u4f5c\u3002\u9996\u5148\u5229\u7528\u8a9e\u97f3\u6d3b\u6027\u6aa2 \u70ba\u4e86\u80fd\u5920\u6a21\u64ec\u771f\u5be6\u7684\u8072\u5b78\u56de\u8072\u6d88\u9664\u7cfb\u7d71\u60c5\u5f62\uff0c\u6211\u5011\u5728\u4e00\u500b\u985e\u4f3c\u5ba2\u5ef3\u7684\u623f\u9593\u6a21\u64ec\u9060\u8ddd</td></tr><tr><td colspan=\"2\">RNNs \u7684\u8f38\u51fa\u5be6\u969b\u4e0a\u662f\u4e00\u975e\u7dda\u6027\u51fd\u6578\uff0c\u56e0\u6b64\u53ef\u4ee5\u7528\u4f86\u4f30\u8a08\u5587\u53ed\uff0c\u9ea5\u514b\u98a8\u9663\u5217\u8207\u623f\u9593\u6b98\u97ff \u9593\u7684\u975e\u7dda\u6027\u95dc\u4fc2\u3002\u5982\u4e0b\u5716\u4e00-B \u6240\u793a\u3002 \u6b64\u5916\uff0c\u5728\u4e00\u524d\u7f6e\u5be6\u9a57\u4e2d\uff0c\u6211\u5011\u767c\u73fe\u82e5\u5728\u6642\u57df\u5be6\u73fe RNNs \u96fb\u8996\u56de\u97f3\u6d88\u9664\u67b6\u69cb\uff0c\u6709\u6642\u5019 \u5716\u4e00-A\u3001 \u5716\u4e00-B\u3001 \u50b3\u7d71 NLMS \u96fb\u8996\u56de\u8072\u6d88\u9664\u7cfb\u7d71\u67b6\u69cb\u5716 \u983b\u57df RNN \u96fb\u8996\u56de\u8072\u6d88\u9664\u7cfb\u7d71\u67b6\u69cb\u5716 \u4e8c\u3001\u57fa\u65bc\u6b21\u983b\u5e36 RNN \u4e4b\u9ea5\u514b\u98a8\u9663\u5217\u96fb\u8996\u56de\u8072\u6d88\u9664\u7cfb\u7d71 \u672c\u8ad6\u6587\u4f7f\u7528\u591a\u901a\u9053\u9ea5\u514b\u98a8\u9663\u5217\u70ba\u7cfb\u7d71\u7684\u8f38\u5165[5]\uff0c\u70ba\u4e86\u5728\u6536\u97f3\u6642\uff0c\u80fd\u5920\u6307\u5411\u4f7f\u7528\u8005 \u8072\u6e90\uff0c\u5148\u4f7f\u7528\u6700\u5c0f\u8b8a\u7570\u7121\u5931\u771f\u97ff\u61c9\u6ce2\u675f\u5f62\u6210\u5668(Minimum Variance Distortionless Response Beamformer,MVDR)[5][6]\uff0c\u5229\u7528\u7a7a\u9593\u8cc7\u8a0a\u9032\u884c\u6307\u5411\u6027\u8655\u7406\u3002\u518d\u7528\u6709\u975e\u7dda\u6027\u8655\u7406\u80fd\u529b\u53ca\u80fd \u88dc\u6349\u9577\u6642\u9593\u8cc7\u6599\u7684\u905e\u8ff4\u985e\u795e\u7d93\u7db2\u8def RNN[4]\uff0c\u4f5c\u70ba\u6838\u5fc3\u4f86\u505a\u56de\u8072\u6d88\u9664\uff0c\u800c\u4e14\u70ba\u4e86\u8b93 RNN \u5728\u56de\u8072\u8def\u5f91\u4e0a\u80fd\u5920\u9810\u4f30\u5f97\u66f4\u7d30\u5fae\uff0c\u6211\u5011\u628a\u6642\u57df\u8a0a\u865f\u8f49\u63db\u5230\u983b\u57df\uff0c\u4e26\u5c07\u8a0a\u865f\u5206\u6210\u4e0d\u540c\u983b\u5e36 \u9032\u884c\u4e0d\u540c\u8655\u7406\u3002\u6700\u5f8c\u5247\u4f7f\u7528 Ephraim's Minimum Mean Square Error log-Shor time Spectral Amplitude(MMSE-log-STSA)[9]\u983b\u8b5c\u6d88\u53bb\u6cd5\u505a\u5f8c\u8655\u7406\uff0c\u9032\u4e00\u6b65\u5229\u7528\u6642\u9593\u8cc7\u8a0a\uff0c\u6d88\u9664\u6b98 \u5716\u4e8c-A\u3001MVDR+\u983b\u57df RNN \u4e4b\u96fb\u8996\u56de\u8072\u6d88\u9664\u67b6\u69cb\u5716 \u5716\u4e8c-B\u3001\u983b\u57df RNN+MVDR \u96fb\u8996\u56de\u8072\u6d88\u9664\u67b6\u69cb\u5716 \u5716\u4e09\u3001MVDR \u67b6\u69cb\u5716[7] \u5176\u4e2d\u9ea5\u514b\u98a8\u63a5\u6536\u8072\u97f3\u8a0a\u865f \u4e2d\u542b\u6709\u76ee\u6a19\u8072\u6e90\u8a0a\u865f \u8ddf\u5e72\u64fe\u52a0\u566a\u97f3\u8a0a\u865f \u3002 \u901a\u904e\u5c0d\u6642\u57df \u9ea5\u514b\u98a8\u8f38 \u5165 \u8a0a\u865f\u61c9\u7528 \u5feb\u901f\u5085\u91cc \u8449\u8b8a\u63db( FFT)\uff0c \u5247 \u9ea5 \u514b\u98a8 \u983b\u57df\u4fe1 \u865f\u70ba { \uff0c =1,2\u2026,4}\u3002\u5176\u8a0a\u865f\u6210\u5206\u53ef\u8868\u793a\u70ba (2.1) \u5176\u4e2d \u662f\u983b\u7387\u7d22\u5f15\uff0c \u662f\u8f38\u5165\u7684\u77ed\u6642\u6bb5\u7d22\u5f15\uff1b \u8868\u793a\u7b2c m \u500b\u9ea5\u514b\u98a8\u5c0d\u76ee\u6a19\u4f86\u6e90 \u7684\u5230\u9054\u65b9\u5411(DOA)\uff0c \u662f\u76ee\u6a19\u4f86\u6e90\uff0c \u548c \u5206\u5225\u8868\u793a\u8b8a\u63db\u5f8c\u7684\u8a9e\u97f3\u4fe1\u865f\u548c\u4e7e \u64fe\u52a0\u566a\u8072\u4fe1\u865f\u3002\u5047\u8a2d\u8a9e\u97f3\u8207\u5e72\u64fe\u52a0\u566a\u8072\u4fe1\u865f\u4e0d\u76f8\u95dc\uff0c\u5247 \u7684\u7a7a\u9593\u983b\u8b5c\u76f8\u95dc\u77e9\u9663\u70ba (2.2) = 1 (2.4) \u5176\u4e2d \u70ba\u62d3\u64b2\u5411\u91cf\uff0c\u9019\u88e1\u5047\u8a2d\u96dc\u8a0a\u8207\u76ee\u6a19\u8a0a\u865f\u7121\u95dc\u806f\u4e14\u5e73\u5747\u70ba 0\uff0c\u5c07 \u5c0d \u505a\u77e9\u9663\u8655\u7406\u3002\u5247\u5728\u7121\u5931\u771f\u6e96\u5247\u7121\u96dc\u8a0a\u7684\u60c5\u6cc1\u4e0b\uff0c\u5c0d\u4efb\u610f \u4e4b MVDR \u6700\u4f73\u89e3\u70ba (\u4e8c) \u983b\u57df RNN \u9069\u61c9\u6027\u96fb\u8996\u56de\u8072\u6d88\u9664 RNN \u9069\u61c9\u6027\u96fb\u8996\u56de\u8072\u6d88\u9664\u7684\u505a\u6cd5\uff0c\u4e3b\u8981\u662f\u5148\u5c07\u96fb\u8996\u539f\u59cb\u8a0a\u865f\u8f38\u5165 RNN\uff0c\u8b93 RNN \u9810\u6e2c\u9ea5\u514b\u98a8\u6536\u5230\u7684\u96fb\u8996\u56de\u8072\u8a0a\u865f\u3002\u518d \u8207\u771f\u6b63\u6536\u5230\u7684\u4eba\u8072\u52a0\u96fb\u8996\u56de\u8072\u8a0a\u865f\u76f8\u6e1b\u5f8c\uff0c\u6d88\u6389\u96fb \u8996\u56de\u8072\uff0c\u5f97\u5230\u8f03\u4e0d\u53d7\u96fb\u8996\u56de\u8072\u5f71\u97ff\u7684\u4eba\u8072\u8a0a\u865f\u3002\u4e0d\u904e\u6211\u5011\u9032\u4e00\u6b65\u5c07\u6574\u500b\u904b\u4f5c\u5f9e\u6642\u57df\u8f49\u5230 \u983b\u57df\uff0c\u8b8a\u6210\u6b21\u983b\u5e36 RNN \u9069\u61c9\u6027\u96fb\u8996\u56de\u8072\u6d88\u9664\u6cd5\u3002\u6b64\u983b\u57df RNN \u67b6\u69cb\u8207\u4e00\u822c\u6642\u57df RNN \u76f8 \u540c\u3002\u4e0d\u540c\u7684\u662f\u5c07\u8f38\u5165\u505a\u6539\u8b8a\uff0c\u539f\u672c\u6642\u57df RNN \u7684\u8f38\u5165\u662f\u4e00\u6574\u6bb5\u7684\u6642\u57df\u8a9e\u97f3\u8a0a\u865f\uff0c\u800c\u983b\u57df RNN \u505a\u6cd5\u662f\u5c07\u8a0a\u865f\u5207\u6210\u97f3\u6846\uff0c\u5c07\u97f3\u6846\u5f9e\u6642\u57df\u7528 FFT \u8f49\u5230\u983b\u57df\uff0c\u6240\u4ee5\u8f38\u5165\u5c31\u6539\u70ba\u9019\u4e00\u5c0f \u6bb5\u97f3\u6846\u7684\u983b\u8b5c\u8a0a\u865f\u3002\u4e0b\u5716\u56db\u70ba\u6b21\u983b\u5e36 RNN \u9069\u61c9\u6027\u96fb\u8996\u56de\u8072\u6d88\u9664\u6d41\u7a0b\u5716\u3002 (2.5) \u6bb5\u7684\u6642\u57df\u8f38\u51fa\u8a0a\u865f \u672c\u8ad6\u6587\u6240\u63d0\u51fa\u7684\u983b\u57df RNN \u96fb\u8996\u56de\u8072\u6d88\u9664\u6cd5\u505a\u6cd5\u5982\u4e0b\u3002\u9996\u5148\u5c07\u539f\u59cb\u7684\u96fb\u8996\u8072\u7d93\u904e RNN \u8a08\u7b97\u3002\u516c\u5f0f\u5982\u4e0b(\u5047\u8a2d\u53ea\u6709\u4e00\u5c64\u96b1\u85cf\u5c64\u7684\u60c5\u6cc1\u4e0b)\uff0c\u5176\u4e2d \u70ba\u96b1\u85cf\u5c64\u6b0a\u91cd\u503c\uff0c \u70ba\u8f38\u51fa\u5c64\u6b0a\u91cd\u503c\uff1a \u662f\u9810\u6e2c\u51fa\u4f86\u7684\u8a0a\u865f\uff0c\u4e26\u8207\u9ea5\u514b\u98a8\u6536\u5230\u4eba\u8072\u548c\u96fb\u8996\u8072\u7684\u6df7\u5408\u8a0a\u865f\u76f8\u6e1b\uff0c\u5f97\u5230\u67d0\u4e00\u6642 \u9593\u9ede\u7684\u8aa4\u5dee\u8a0a\u865f\u3002\u518d\u4ee5\u5747\u65b9\u5dee\u8a08\u7b97\u6210\u672c\u51fd\u6578\uff0c\u516c\u5f0f\u5982\u4e0b(\u5176\u4e2d \u70ba\u9ea5\u514b\u98a8\u6536\u5230\u4eba\u8072\u52a0 \u96fb\u8996\u8072\u7684\u8a0a\u865f)\uff1a (2.6) (2.7) \u6e2c(voice activity detection\uff0cVAD)[11]\uff0c\u5c0d\u8f38\u5165\u4fe1\u865f\u7684\u4e00\u500b\u5340\u584a\u63d0\u53d6\u7279\u5fb5\u7136\u5f8c\u5c0d\u9019\u500b\u5340\u584a \u9032\u884c\u5206\u985e\uff0c\u5207\u51fa\u8a9e\u97f3\u8a0a\u865f\uff0c\u4f30\u8a08\u80cc\u666f\u96dc\u8a0a\u983b\u8b5c\u3002\u7136\u5f8c\u518d\u5229\u7528\u983b\u8b5c\u6d88\u53bb\u6cd5\uff0c\u53bb\u6389\u80cc\u666f\u96dc\u8a0a\uff0c \u8b93\u8f38\u5165\u8a0a\u865f\u900f\u904e\u983b\u8b5c\u76f8\u6e1b\u9054\u5230\u964d\u566a\u7684\u6548\u679c\u3002 \u4e09\u3001\u5be6\u9a57\u7d50\u679c \u672c\u8ad6\u6587\u7684\u5be6\u9a57\u8a9e\u6599\uff0c\u5305\u542b 8 \u500b\u8a9e\u8005\u7684 TCC300 \u8a9e\u6599\uff0c\u4ee5\u53ca\u8003\u616e 4 \u7a2e\u96fb\u8996\u7bc0\u76ee\u985e\u578b\uff0c \u6bcf\u4e00\u985e\u9078 10 \u500b\u96fb\u8996\u7bc0\u76ee\u8072\u3002SNR \u8a2d\u5b9a\u5247\u70ba\u4eba\u8072\u8207\u96fb\u8996\u8072\u4e00\u6a23(0dB)\u3001\u4eba\u8072\u6bd4\u96fb\u8996\u8072\u5c0f (-6dB)\u3001\u4eba\u8072\u6bd4\u96fb\u8996\u8072\u5927(+6dB)\u4e09\u7a2e\u3002\u5176\u4e2d\u55ae\u901a\u9053\u5be6\u9a57\u53d6\u9760\u8fd1\u9ea5\u514b\u98a8\u4e2d\u5fc3\u9ede\u6700\u8fd1\u7684\u7b2c\u4e8c \u8072\u9053\u505a\u70ba\u6e2c\u8a66\u3002\u591a\u901a\u9053\u5247\u4f7f\u7528 4 \u500b\u9ea5\u514b\u98a8\u3002\u983b\u57df\u8655\u7406\u97f3\u6846\u5927\u5c0f\u8a2d\u70ba 256\uff0c\u4f7f\u7528\u6f22\u660e\u7a97\uff0c RNN \u6f14\u7b97\u6cd5\u8a2d\u5b9a\u70ba\u5169\u5c64\u96b1\u85cf\u5c64\uff0c\u6bcf\u5c64\u795e\u7d93\u5143\u70ba 100\u3002 (\u4e00) \u8a9e\u6599\u8aaa\u660e \u6e2c\u8a66\u4eba\u8072\u9304\u97f3\u8a9e\u6599\u5f9e TCC300 \u8a9e\u6599\u5eab\u4e2d\uff0c\u9078\u64c7\u4e86 4 \u7537 4 \u5973\u7684\u97f3\u6a94\uff0c\u4e14\u96a8\u6a5f\u64f7\u53d6\u5341\u79d2 \u9418\u7247\u6bb5\u8aaa\u8a71\u8072\uff1b\u800c\u6240\u56de\u9304\u7684\u96fb\u8996\u7bc0\u76ee\u8072\u70ba\u56db\u5927\u985e\uff0c\u6bcf\u985e\u70ba\u6709 10 \u500b\u9577\u5ea6\u5927\u7d04\u534a\u5c0f\u6642\u7684\u7bc0 \u76ee\uff0c\u5f9e\u4e2d\u96a8\u6a5f\u64f7\u53d6\u5341\u79d2\u9418\u7247\u6bb5\u7bc0\u76ee\u8072\u3002\u6240\u4ee5\u6bcf\u4eba\u5171\u6709 40 \u500b\u6e2c\u8a66\u96fb\u8996\u7bc0\u76ee\u8072\u3002\u4ee5\u4e0b\u8868 1 \u70ba\u8a9e\u6599\u7684\u683c\u5f0f\u8a2d\u5b9a\u3002 \u96e2\u6536\u97f3\u60c5\u5883\u3002\u9996\u5148\uff0c\u628a Kinect for Xbox one \u5145\u7576\u63a5\u6536\u7aef\u9ea5\u514b\u98a8\u9663\u5217\u3002\u4e26\u5728\u87a2\u5e55\u4f4d\u7f6e\u5de6\u53f3 \u5169\u65c1\u5e73\u884c\u653e\u4e0a\u5169\u9846\u4e3b\u52d5\u5f0f\u76e3\u807d\u5587\u53ed\uff0c\u64ad\u653e\u96fb\u8996\u7bc0\u76ee\u7684\u8072\u97f3\u3002\u53e6\u5916\uff0c\u5728\u87a2\u5e55\u7684\u6b63\u524d\u65b9\u8ddd\u96e2 2m \u8655\uff0c\u64fa\u4e0a\u53e6\u4e00\u9846\u4e3b\u52d5\u5f0f\u76e3\u807d\u5587\u53ed\u64ad\u653e\u4eba\u8072\uff0c\u6a21\u64ec\u4f7f\u7528\u8005\u6b63\u5728\u8b1b\u8a71\u3002\u5982\u6b64\u4e00\u4f86\uff0c\u7576\u64ad \u653e\u51fa\u4eba\u8072\u6642\uff0c\u5f71\u7247\u7bc0\u76ee\u8072\u4e5f\u540c\u6642\u6df7\u9032\u758a\u52a0\u5176\u4e2d\uff0c\u4e00\u8d77\u88ab\u9ea5\u514b\u98a8\u9663\u5217\u6240\u63a5\u6536\uff0c\u4e26\u9304\u97f3\u8d77\u4f86 \u7576\u4f5c\u6211\u5011\u8072\u5b78\u56de\u8072\u6d88\u9664\u7cfb\u7d71\u7684\u8a9e\u6599\u3002\u5be6\u969b\u4e0a\u6211\u5011\u5171\u9304\u88fd\u8aaa\u8a71\u8005\u5728 90\uff0c60 \u8207 30 \u5ea6\u89d2\u4f4d\u7f6e \u7684\u4eba\u8072\uff0c\u6574\u9ad4\u5be6\u9a57\u74b0\u5883\u64fa\u8a2d\u5982\u4e0b\u5716\u4e94\u3001\u516d\u3002 \u5716\u4e94\u3001\u89d2\u5ea6\u64fa\u8a2d\u5716 \u5176\u4e2d\uff0cm(n)\u70ba TCC300 \u6e2c\u8a66\u4eba\u8072\u6df7\u96fb\u8996\u7bc0\u76ee\u8072\u7684\u9304\u97f3\u8a0a\u865f\u8072\u3002s(n)\u70ba\u6d88\u9664\u56de\u8072\u5f8c\u7684 \u4f30\u8a08\u8a9e\u97f3\u8a0a\u865f\uff1b\u5373\u7d93\u56de\u8072\u6d88\u9664\u7cfb\u7d71\u6d88\u9664\u5f71\u7247\u7bc0\u76ee\u8072\u5f8c\u5f97\u5230\u7684 TCC300 \u6e2c\u8a66\u4eba\u8072\u3002\u85c9\u7531\u539f \u59cb\u7684\u8a0a\u865f\u8072 m(n)\u8207\u7d93\u56de\u8072\u6d88\u9664\u5f8c\u5f97\u5230\u7684 TCC300 \u6e2c\u8a66\u4eba\u8072 s(n)\u5169\u8005\u76f8\u4e92\u53bb\u6bd4\u8f03\uff0c\u7576\u5206\u5b50 s(n)\u8d8a\u5c0f\u6642\uff0c\u6b64\u6642 ERLE \u503c\u5c31\u6108\u5927\uff0c\u4ee3\u8868\u6d88\u9664\u6027\u80fd\u6108\u597d\uff1b\u4e5f\u8868\u793a\u5f97\u5230\u6108\u6e05\u6670\u7684 TCC300 \u6e2c\u8a66\u4eba\u8072\u3002 (\u56db) \u5be6\u9a57\u7d50\u679c \u5728\u4ee5\u4e0b\u5be6\u9a57\u4e2d\uff0c\u5148\u8b93\u7cfb\u7d71\u85c9\u7531\u524d 5 \u79d2\u7d14\u96fb\u8996\u56de\u8072\u7684\u9069\u61c9\u5b78\u7fd2\uff0c\u518d\u7528\u4f86\u9810\u6e2c\u5230\u4e0b\u4e94\u79d2 \u7684\u96fb\u8996\u56de\u8072\uff0c\u6e2c\u8a66\u4eba\u8072\u8207\u96fb\u8996\u7bc0\u76ee\u8072\u6df7\u5408\u7684\u6bd4\u4f8b\u5148\u4fdd\u6301\u4e00\u6a23\u5927\u8072\uff0c\u5373\u6a21\u64ec SNR = 0dB \u7684\u60c5\u5f62\u3002\u5be6\u9a57\u7d50\u679c\u4e3b\u8981\u4ee5\u5f8c\u4e94\u79d2\u70ba\u5224\u65b7\u7cfb\u7d71\u597d\u58de\u6e96\u5247\u3002\u4ee5\u4e0b\u4ecb\u7d39\u4e94\u7a2e\u5be6\u9a57\uff1a \uf06c \u5be6\u9a57\u4e00\uff1a\u6642\u57dfNLMS\u8207\u6642\u57dfRNN\u96fb\u8996\u56de\u8072\u6d88\u9664\u5be6\u9a57\u3002 \uf06c \u5be6\u9a57\u4e8c\uff1a\u6642\u57df\u8207\u983b\u57dfRNN\u96fb\u8996\u56de\u8072\u6d88\u9664\u5be6\u9a57\u3002 \uf06c \u5be6\u9a57\u4e09\uff1aMVDR\u52a0\u983b\u57dfRNN\u96fb\u8996\u56de\u8072\u6d88\u9664\u5be6\u9a57\u3002 \uf06c \u5be6\u9a57\u56db\uff1aMVDR\u52a0\u983b\u57dfRNN\u5728\u4e0d\u540c\u89d2\u5ea6\u96fb\u8996\u56de\u8072\u6d88\u9664\u5be6\u9a57\u3002 \uf06c \u5be6\u9a57\u4e94\uff1aMVDR\u52a0\u983b\u57dfRNN\u518d\u52a0\u4e0aMMSE\u5f8c\u8655\u7406\u96fb\u8996\u56de\u8072\u6d88\u9664\u5be6\u9a57\u3002 2. \u5be6\u9a57\u4e8c\uff0c\u6642\u57df\u8207\u983b\u57dfRNN\u96fb\u8996\u56de\u8072\u6d88\u9664\u5be6\u9a57 \u5be6\u9a57\u4e8c\u8003\u616e\u5230\u5229\u7528\u591a\u901a\u9053\u4f86\u589e\u5f37\u4eba\u8072\uff0c\u6240\u4ee5\u5148\u5047\u8a2d\u8aaa\u8a71\u8005\u5728\u6b63\u524d\u65b9\uff0c\u7528\u9ea5\u514b\u98a8\u9663\u5217 \u6536\u96c6\u7684 4 \u500b\u8072\u9053\u97f3\u8a0a\uff0c\u5206\u5225\u505a 4 \u6b21\u7684\u6df1\u5c64\u905e\u8ff4\u5f0f\u795e\u7d93\u7db2\u8def\uff0c\u5728\u52a0\u7e3d\u589e\u5f37\u8a9e\u97f3\uff0c\u4e26\u6bd4\u8f03 RNN \u5728\u6642\u57df\u8207\u983b\u57df\u6d88\u9664\u96fb\u8996\u56de\u8072\u7684\u6548\u80fd\u3002\u7d50\u679c\u5982\u4e0b\u5716\u516b\u6240\u793a\uff1a \u5716\u516b\u3001\u5be6\u9a57\u4e8c RNN \u8207\u983b\u57df RNN \u96fb\u8996\u56de\u8072\u6d88\u9664\u7cfb\u7d71\u7684 ERLE \u8207 MSE \u7d50\u679c\u5716 \u5f9e\u5716\u516b\u7684\u5be6\u9a57\u7d50\u679c\uff0c\u53ef\u4ee5\u770b\u51fa\u591a\u901a\u9053\u983b\u57df RNN \u5728\u5404\u7a2e\u985e\u578b\u7684\u96fb\u8996\u7bc0\u76ee\u90fd\u512a\u8d8a\u65bc\u591a 1. RNN \u6548\u679c\u660e\u986f\u512a\u65bc NLMS\u3002 \u901a\u9053\u6642\u57df RNN\u3002 \u5728\u67d0\u4e9b\u983b\u5e36\u5e38\u5e38\u6d88\u9664\u5f97\u4e0d\u5920\u597d\uff0c\u5c24\u5176\u662f\u5728\u9ad8\u983b\u90e8\u5206\u3002\u9019\u662f\u56e0\u70ba\u96fb\u8996\u7bc0\u76ee\u56de\u8072\u662f\u7d93\u904e\u5f88\u591a \u53cd\u5c04\u56de\u4f86\u7684\u8072\u97f3\u758a\u52a0\u8d77\u4f86\uff0c\u53ef\u4ee5\u8996\u70ba\u539f\u59cb\u8072\u97f3\u8ddf\u591a\u91cd\u8def\u5f91\u623f\u9593\u97ff\u61c9\u7684\u5377\u7a4d\u758a\u52a0\u7684\u7d50\u679c\uff0c \u5f80\u5f80\u5728\u4e0d\u540c\u7684\u983b\u5e36\u4e0a\u6703\u53d7\u5230\u4e0d\u540c\u7a0b\u5ea6\u7684\u5e72\u64fe\u3002\u5982\u679c\u6211\u5011\u53ea\u4f7f\u7528\u6642\u57df RNNs \u67b6\u69cb\uff0c\u5e38\u6703\u7121 \u6cd5\u7167\u9867\u4e0d\u540c\u983b\u5e36\u7684\u8b8a\u5316\uff0c\u5c0e\u81f4\u7121\u6cd5\u6709\u6548\u6d88\u9664\u96dc\u8a0a\u3002\u6240\u4ee5\u6211\u5011\u5728\u6b64\u8ad6\u6587\u4e2d\uff0c\u9032\u4e00\u6b65\u63d0\u51fa\u983b \u57df RNN \u96fb\u8996\u56de\u8072\u6d88\u9664\u7cfb\u7d71\uff0c\u5c31\u53ef\u4ee5\u91dd\u5c0d\u4e0d\u540c\u983b\u5e36\u7684\u8a0a\u865f\uff0c\u61c9\u7528\u4e0d\u540c\u7684\u8abf\u6574\u53c3\u6578\u3002 \u9918\u7684\u56de\u8072\u96dc\u8a0a\u3002 \u6b64\u5916\uff0c\u6211\u5011\u8003\u616e\u5169\u7a2e\u4e0d\u540c\u7684 MVDR \u8207 RNNs \u7d44\u5408\u67b6\u69cb\uff0c\u5206\u5225\u5982\u4e0b\u5716\u4e8c-A \u8207\u4e8c-B \u6240 \u793a\u3002\u5176\u4e2d\u4e0b\u5716\u4e8c-A\uff0c\u662f\u5148\u7d93\u904e MVDR \u505a\u6307\u5411\u6027\u8655\u7406\uff0c\u518d\u7528\u983b\u57df RNN \u505a\u9810\u4f30\u8207\u62b5\u6d88\u3002 \u6700\u5f8c\u5247\u5229\u7528\u53cd\u5411\u50b3\u64ad\u7b97\u6cd5(back propagation through time\uff0cBPTT)\u8abf\u6574 RNN \u6b0a\u91cd\u3002 \u8868 1 \u8a9e\u6599\u683c\u5f0f\u8a2d\u5b9a \u5716\u516d\u3001\u5be6\u969b\u96fb\u8996\u5587\u53ed\u8207\u6536\u97f3\u9ea5\u514b\u98a8\u8207\u5be6\u969b\u6a21\u64ec\u8aaa\u8a71\u8005\u4e4b\u97f3\u6e90\u5587\u53ed\u64fa\u8a2d (\u4e00) MVDR \u6700\u5c0f\u8b8a\u7570\u5931\u771f\u97ff\u61c9 \u5047\u8a2d\u9ea5\u514b\u98a8\u9663\u5217\u6703\u6536\u5230\u76ee\u6a19\u8a9e\u97f3\uff0c\u4ee5\u53ca\u4f86\u81ea\u5176\u4ed6\u65b9\u5411\u7684\u5e72\u64fe\u96dc\u8a0a\u8207\u975e\u76ee\u6a19\u8a9e\u97f3\u4fe1 \u865f\u3002\u6211\u5011\u4f7f\u7528 MVDR Beamformer[7][8]\uff0c\u3002\u56e0\u70ba\u5728\u6bcf\u652f\u9ea5\u514b\u98a8\u5728\u4e0d\u540c\u89d2\u5ea6\u7684\u95dc\u4fc2\uff0c\u4f7f\u5f97 \u8a9e\u97f3\u5230\u9054\u6bcf\u652f\u9ea5\u514b\u98a8\u90fd\u6703\u6709\u5ef6\u9072\u3002\u5982\u679c\u80fd\u900f\u904e\u4f7f\u7528\u8005\u8072\u6e90\u65b9\u4f4d\uff0c\u8a08\u7b97\u8072\u97f3\u50b3\u81f3\u9ea5\u514b\u98a8\u7684 \u5176\u4e2d \u662f\u8868\u793a\u5171\u8edb\u8f49\u7f6e\uff0c \u548c \u662f\u7a7a\u9593\u8b5c\u76f8\u95dc\u77e9\u9663\uff0c \u662f\u8a9e\u97f3\u529f\u7387 \u8b5c\u5bc6\u5ea6\uff0c\u6ce2\u675f\u5f62\u6210\u5668\u7684\u8f38\u51fa\u8a0a\u865f \u53ef\u4ee5\u5beb\u6210\u4e0b\u5f0f\u3002\u6b64\u5916 \u8449\u8f49\u63db\uff0c\u7372\u5f97\u6700\u5f8c\u8f49\u6a94\u6216\u6642\u57df\u7684\u9ea5\u514b\u98a8\u8f38\u51fa\u8a0a\u865f\u3002 (2.8) Sample Rate 44.1K Hz Bit Resolution 16-bit 16-bit \u56de\u8072\u6d88\u9664\u6210\u6548\u9664\u4e86\u4e3b\u89c0\u7684\u7531\u8033\u6735\u807d\u53d6\u8072\u97f3\u5916\uff0c\u9084\u53ef\u4ee5\u7528\u5e73\u5747\u8aa4\u5dee\u503c(Mean Squared 16K Hz \u8a0a\u865f\u53ef\u4ee5\u518d\u7528\u53cd\u5085\u7acb \u96fb\u8996\u7bc0\u76ee\u8072 \u8a9e\u8005\u4eba\u8072 BPTT \u5148\u7531\u6700\u5f8c\u6642\u9593\u9ede\uff0c\u958b\u59cb\u5c0d\u65bc\u6210\u672c\u51fd\u6578\u4f5c\u504f\u5fae\u5206\uff0c\u518d\u5f80\u524d\u7b97\u5230\u4e00\u958b\u59cb\u6642\u9593\u9ede\u7684\u504f\u5fae Format Wave PCM Wave PCM (\u4e09) \u56de\u8072\u6d88\u9664\u6548\u80fd\u8a55\u4f30 \u5206\u503c\u3002RNN \u7684\u8abf\u9069\u516c\u5f0f\u5982\u4e0b\uff1a Source Youtube TCC300 3.</td></tr><tr><td colspan=\"2\">\u4e0b\u5716\u4e8c-B \u5247\u662f\u5148\u5c07\u6bcf\u500b\u901a\u9053\u9ea5\u514b\u98a8\u6536\u5230\u7684\u8072\u97f3\u5148\u7d93\u904e RNN \u6d88\u9664\u56de\u8072\uff0c\u8b93\u6bcf\u500b\u9ea5\u514b\u98a8\u7684 \u6642\u9593\u5dee[5]\uff0c\u4e26\u900f\u904e\u6700\u5c0f\u8f38\u51fa\u65b9\u5dee\u6e96\u5247\u627e\u5230\u983b\u57df\u6ffe\u6ce2\u5668\u4fc2\u6578\u7684\u6700\u4f73\u89e3\uff0c\u9032\u884c filter-and-sum (2.3) Vocal Track stereo mono Error\uff0cMSE)\uff0c\u8207\u56de\u8072\u8fd4\u56de\u640d\u5931\u5f37\u5316(Echo Return Loss Enhancement\uff0cERLE)[5]\u6578\u503c\u5316\u3002</td></tr><tr><td colspan=\"2\">\u8aa4\u5dee\u9054\u5230\u6700\u5c0f\u4e4b\u5f8c\uff0c\u518d\u505a MVDR \u7684\u52d5\u4f5c\u6703\u4e0d\u6703\u6bd4\u8f03\u597d\u3002 \u88dc\u511f\uff0c\u5c31\u53ef\u5c07\u4f7f\u7528\u8005\u8aaa\u8a71\u8072\u589e\u5f37\u3002\u5982\u4e0b\u5716\u4e09\u6240\u793a\u3002 \u800c\u70ba\u4e86\u627e\u5230\u6700\u4f73\u6b0a\u91cd\u5411\u91cf\uff0cMVDR \u6ce2\u675f\u5f62\u6210\u5668\u5c07\u6bcf\u500b\u983b\u7387\u7684\u7e3d\u8f38\u51fa\u4fe1\u865f\u529f\u7387\u6700\u5c0f ERLE \u8a55\u4f30\u51fd\u5f0f\u5f0f\u5982\u4e0b\u5f0f\u6240\u793a\uff1a</td></tr><tr><td colspan=\"2\">\u6700\u5f8c\uff0c\u5728\u4ee5\u4e0b\u7ae0\u7bc0\uff0c\u5c07\u8a73\u7d30\u4ecb\u7d39\u6211\u5011\u6240\u63d0\u51fa\u7684 RNNs \u96fb\u8996\u56de\u97f3\u6d88\u9664\u67b6\u69cb\u8207\u5176\u8a13\u7df4\u65b9\u6cd5\u3002 \u4ee5\u4e0b\u9032\u4e00\u6b65\u8a73\u7d30\u6b64\u5169\u67b6\u69cb\u4e2d\u7684\u5404\u5b50\u6a21\u7d44\uff0c\u5305\u62ec(\u4e00)MVDR\uff0c (\u4e8c)\u983b\u57df RNNs \u96fb \u5316\uff0c\u540c\u6642\u9650\u5236\u6ce2\u675f\u5f62\u6210\u5668\u7684\u6b0a\u91cd\u5927\u5c0f\uff0c\u8b93\u9ea5\u514b\u98a8\u6307\u5411\u7684\u65b9\u5411\u7684\u97ff\u61c9\u70ba 1\uff0c\u5373\u4f9d\u4e0b\u5f0f\u5c0b\u627e \u5716\u56db\u3001\u983b\u57df RNN \u67b6\u69cb\u5716</td></tr></table>", | |
| "type_str": "table", | |
| "num": null | |
| } | |
| } | |
| } | |
| } |