| { | |
| "paper_id": "O16-1018", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T08:05:09.950053Z" | |
| }, | |
| "title": "A study of enhancing the modulation spectrum of speech signals via nonnegative matrix factorization", | |
| "authors": [ | |
| { | |
| "first": "Xu-Xiang", | |
| "middle": [], | |
| "last": "\u738b\u7dd2\u7fd4", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Wang", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Zhi-Hao", | |
| "middle": [], | |
| "last": "\u3001\u912d\u81f3\u7693", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Zheng", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "\u3001", | |
| "middle": [], | |
| "last": "\u66f9\u6631", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Yu", | |
| "middle": [], | |
| "last": "Tsao", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "\u3001\u6d2a\u5fd7\u5049jhih-Wei", | |
| "middle": [], | |
| "last": "Hong", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "\u4e2d\u7814\u9662\u8cc7\u8a0a\u79d1\u6280\u5275\u65b0\u7814\u7a76\u4e2d\u5fc3", | |
| "middle": [], | |
| "last": "\u6458\u8981", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| } | |
| ], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "In this paper, we propose to enhance the modulation spectrum of the spectrograms for speech signals via the technique of non-negative matrix factorization (NMF). In the training phase, the clean speech and noise in the training set are separately transformed to spectrograms and modulation spectra in turn, and then the magnitude modulation spectra are used to train the NMF-based basis matrices for clean speech and noise, respectively. In the test phase, the test signal is converted to its modulation spectrum, which is then enhanced via NMF with the basis matrices obtained in the training phase. The updated modulation spectrum is finally transformed back to the time domain as the enhanced signal. In addition, we propose two variants for the newly method in order to possess relatively high computation complexity One is to consider the several adjacent acoustic frequencies as a whole for the subsequent processing, and the other is to process the low modulation frequency components. These new methods are validated via a subset of the Aurora-2 noisy connected-digit database. Preliminary experiments have indicated that these methods can achieve better signal quality relative to the baseline results in terms of the Perceptual Evaluation of Speech Quality (PESQ) index, and they outperform some well-known speech enhancement methods including spectral subtraction (SS), Wiener filtering (WF) and minimum mean squared error short-time spectral amplitude estimation (MMSE-STSA).", | |
| "pdf_parse": { | |
| "paper_id": "O16-1018", | |
| "_pdf_hash": "", | |
| "abstract": [ | |
| { | |
| "text": "In this paper, we propose to enhance the modulation spectrum of the spectrograms for speech signals via the technique of non-negative matrix factorization (NMF). In the training phase, the clean speech and noise in the training set are separately transformed to spectrograms and modulation spectra in turn, and then the magnitude modulation spectra are used to train the NMF-based basis matrices for clean speech and noise, respectively. In the test phase, the test signal is converted to its modulation spectrum, which is then enhanced via NMF with the basis matrices obtained in the training phase. The updated modulation spectrum is finally transformed back to the time domain as the enhanced signal. In addition, we propose two variants for the newly method in order to possess relatively high computation complexity One is to consider the several adjacent acoustic frequencies as a whole for the subsequent processing, and the other is to process the low modulation frequency components. These new methods are validated via a subset of the Aurora-2 noisy connected-digit database. Preliminary experiments have indicated that these methods can achieve better signal quality relative to the baseline results in terms of the Perceptual Evaluation of Speech Quality (PESQ) index, and they outperform some well-known speech enhancement methods including spectral subtraction (SS), Wiener filtering (WF) and minimum mean squared error short-time spectral amplitude estimation (MMSE-STSA).", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Abstract", | |
| "sec_num": null | |
| } | |
| ], | |
| "body_text": [ | |
| { | |
| "text": "EQUATION", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [ | |
| { | |
| "start": 0, | |
| "end": 8, | |
| "text": "EQUATION", | |
| "ref_id": "EQREF", | |
| "raw_str": "\uf0d8 \u6b50\u5e7e\u91cc\u5fb7\u8ddd\u96e2\u5e73\u65b9 (Squared Euclidean distance)\uff1a 1 = \u2211 ( \u2212 ( ) ) 2 , (2) \uf0d8 KL \u6563\u5ea6 (Kullback-Leibler divergence)\uff1a 2 = \u2211 ( log ( ) \u2212 + ( ) ) ,", | |
| "eq_num": "(3)" | |
| } | |
| ], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "\u63a5\u8457\u4f7f\u7528\u4e58\u6cd5\u6cd5\u5247 (multiplication rule) \u66f4\u65b0 \u8207 \uff0c\u5c07 (2) \u5f0f\u6216 (3) \u5f0f\u6240\u8868\u793a\u7684\u6210\u672c\u51fd \u6578\u9010\u6b65\u7e2e\u5c0f\uff0c\u82e5\u4f7f\u7528 (2) \u5f0f\uff0c\u5247 \u8207 \u7684\u8fed\u4ee3\u66f4\u65b0\u5f0f\u5982\u4e0b:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "EQUATION", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [ | |
| { | |
| "start": 0, | |
| "end": 8, | |
| "text": "EQUATION", | |
| "ref_id": "EQREF", | |
| "raw_str": "\u2190 ( ) ( ) (4) \u2190 ( ) ( )", | |
| "eq_num": "(5)" | |
| } | |
| ], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "\u82e5\u4f7f\u7528 (3) \u5f0f\u5247\u70ba:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "\u2190 \u2211 ( ) \u2044 \u2211 (6) \u2190 \u2211 ( ) \u2044 \u2211 (7)", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "\u7279\u5225\u4e00\u63d0\u7684\u662f\uff0c\u672c\u6587\u5f8c\u7e8c\u6240\u4f7f\u7528\u7684 NMF \u6cd5\uff0c\u5176\u6210\u672c\u51fd\u6578\u56fa\u5b9a\u70ba 2 ", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "EQUATION", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [ | |
| { | |
| "start": 0, | |
| "end": 8, | |
| "text": "EQUATION", | |
| "ref_id": "EQREF", | |
| "raw_str": "[ , ] = \u2211 (\u2113) \u2212 2 \u2113 \u22121 =0 ,", | |
| "eq_num": "(8)" | |
| } | |
| ], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "0 \u2264 \u2264 \u2212 1, 0 \u2264 \u2264 \u2212 1, \u5176\u4e2d\uff0c{ (\u2113), 0 \u2264 \u2113 \u2264 \u2212 1} \u4ee3\u8868\u4e86\u7b2c \u500b\u97f3\u6846\u7684\u6642\u9593\u8a0a\u865f\uff0c \u8207 \u5206\u5225\u70ba\u97f3\u6846\u7e3d\u6578\u8207\u8072 \u5b78\u983b\u7387\u9ede\u6578\u3002\u5f0f(8)\u4e2d\u7684 [ , ]\u901a\u5e38\u7a31\u70ba\u8a9e\u97f3\u7684\u6642\u983b\u5716 (spectrogram)\u3002\u63a5\u8457\uff0c\u6211\u5011\u5c0d\u4efb\u4e00 \u8072\u5b78\u983b\u7387\u9ede \u7684\u8072\u5b78\u983b\u8b5c\u5f37\u5ea6| [ , ]|\u3001\u6cbf\u8457\u97f3\u6846\u6642\u9593\u5e8f\u5217\u8ef8 (\u5373 \u8ef8)\u518d\u505a\u4e00\u6b21\u5085\u7acb\u8449 \u8f49\u63db (Fourier transform)\uff0c\u5373\u53ef\u5f97\u5230\u5404\u8072\u5b78\u983b\u7387\u9ede\u4e4b\u8abf\u8b8a\u983b\u8b5c\uff0c\u5982\u4e0b\u5f0f(9)\u3002 [ , ] = \u2211 | [ , ]| \u2212 2 \u22121 =0 (9) 0 \u2264 \u2264 \u2212 1, 0 \u2264 \u2264 \u2212 1, \u5176\u4e2d \u70ba\u8abf\u8b8a\u983b\u7387\u9ede\u6578\u3002\u5728\u6211\u5011\u6240\u65b0\u63d0\u51fa\u7684 NMF-MSE \u6cd5\u4e2d\uff0c\u5373\u662f\u91dd\u5c0d\u5f0f(9)", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "\u6700\u5f8c\uff0c\u53d6\u5f97\u8fd1\u4f3c\u4e7e\u6de8\u8a9e\u97f3\u983b\u8b5c\u4e4b\u5f37\u5ea6\uff0c\u5982 \u6216 ", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "EQUATION", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [ | |
| { | |
| "start": 0, | |
| "end": 8, | |
| "text": "EQUATION", | |
| "ref_id": "EQREF", | |
| "raw_str": "( ( + ) \u2044 ) \u00d7 ,", | |
| "eq_num": "(11)" | |
| } | |
| ], | |
| "section": "", | |
| "sec_num": null | |
| } | |
| ], | |
| "back_matter": [], | |
| "bib_entries": { | |
| "BIBREF0": { | |
| "ref_id": "b0", | |
| "title": "Phoneme-dependent NMF for speech enhancement in monaural mixtures", | |
| "authors": [ | |
| { | |
| "first": "B", | |
| "middle": [], | |
| "last": "Raj", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "R", | |
| "middle": [], | |
| "last": "Singh", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Virtanen", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2011, | |
| "venue": "Proceedings of Interspeech", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "B. Raj, R. Singh, T. Virtanen, \"Phoneme-dependent NMF for speech enhancement in monaural mixtures\" in Proceedings of Interspeech, 2011.", | |
| "links": null | |
| }, | |
| "BIBREF1": { | |
| "ref_id": "b1", | |
| "title": "Codebook driven short-term predictor parameter estimation for speech enhancement", | |
| "authors": [ | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Srinivasan", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Samuelsson", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "W", | |
| "middle": [], | |
| "last": "Kleijn", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2006, | |
| "venue": "IEEE Transactions on Audio, Speech, and Language Processing", | |
| "volume": "14", | |
| "issue": "1", | |
| "pages": "163--176", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "S. Srinivasan, J. Samuelsson, and W. Kleijn, \"Codebook driven short-term predictor parameter estimation for speech enhancement,\" IEEE Transactions on Audio, Speech, and Language Processing, 14(1), pp. 163-176, 2006.", | |
| "links": null | |
| }, | |
| "BIBREF2": { | |
| "ref_id": "b2", | |
| "title": "Supervised and Unsupervised Speech Enhancement Using Nonnegative Matrix Factorization", | |
| "authors": [ | |
| { | |
| "first": "N", | |
| "middle": [], | |
| "last": "Mohammadiha", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2013, | |
| "venue": "IEEE Trans on Audio, Speech, and Language Processing", | |
| "volume": "21", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "N. Mohammadiha, \"Supervised and Unsupervised Speech Enhancement Using Nonnegative Matrix Factorization,\" IEEE Trans on Audio, Speech, and Language Processing, vol 21, Oct. 2013", | |
| "links": null | |
| }, | |
| "BIBREF3": { | |
| "ref_id": "b3", | |
| "title": "Speech enhancement using convolutive nonnegative matrix factorization with cosparsity regularization", | |
| "authors": [ | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Mirbagheri", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Xu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Akram", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": ",", | |
| "middle": [ | |
| "S" | |
| ], | |
| "last": "Shamma", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2013, | |
| "venue": "Proceedings of Interspeech", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "M. Mirbagheri, Y. Xu, S. Akram and, S. Shamma, \"Speech enhancement using convolutive nonnegative matrix factorization with cosparsity regularization\" in Proceedings of Interspeech, 2013", | |
| "links": null | |
| }, | |
| "BIBREF4": { | |
| "ref_id": "b4", | |
| "title": "Suppression of acoustic noise in speech using spectral subtraction", | |
| "authors": [ | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Boll", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1979, | |
| "venue": "IEEE Transactions on Acoustics, Speech, and Signal Processing", | |
| "volume": "27", | |
| "issue": "2", | |
| "pages": "113--120", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "S. Boll, \"Suppression of acoustic noise in speech using spectral subtraction,\" IEEE Transactions on Acoustics, Speech, and Signal Processing, 27(2), pp. 113-120, 1979.", | |
| "links": null | |
| }, | |
| "BIBREF5": { | |
| "ref_id": "b5", | |
| "title": "The Extrapolation, Interpolation, and Smoothing of Stationary Time Series with Engineering Applications", | |
| "authors": [ | |
| { | |
| "first": "N", | |
| "middle": [], | |
| "last": "Wiener", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1949, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "N. Wiener, \"The Extrapolation, Interpolation, and Smoothing of Stationary Time Series with Engineering Applications,\" New York: Wiley, 1949.", | |
| "links": null | |
| }, | |
| "BIBREF6": { | |
| "ref_id": "b6", | |
| "title": "On causal algorithms for speech enhancement", | |
| "authors": [ | |
| { | |
| "first": "V", | |
| "middle": [], | |
| "last": "Grancharov", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [ | |
| "S B" | |
| ], | |
| "last": "Kleijn", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2006, | |
| "venue": "IEEE Transactions on Audio, Speech, and Language Processing", | |
| "volume": "14", | |
| "issue": "3", | |
| "pages": "764--773", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "V. Grancharov and J. S. B. Kleijn, \"On causal algorithms for speech enhancement,\" IEEE Transactions on Audio, Speech, and Language Processing, 14(3), pp. 764-773, 2006.", | |
| "links": null | |
| }, | |
| "BIBREF7": { | |
| "ref_id": "b7", | |
| "title": "A Laplacian-based MMSE estimator for speech enhancement", | |
| "authors": [ | |
| { | |
| "first": "B", | |
| "middle": [], | |
| "last": "Chen", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "P", | |
| "middle": [ | |
| "C" | |
| ], | |
| "last": "Loizou", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2007, | |
| "venue": "Speech Communication", | |
| "volume": "49", | |
| "issue": "2", | |
| "pages": "134--143", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "B. Chen and P. C. Loizou, \"A Laplacian-based MMSE estimator for speech enhancement,\" Speech Communication, vol. 49, no. 2, pp. 134-143,Feb. 2007.", | |
| "links": null | |
| }, | |
| "BIBREF8": { | |
| "ref_id": "b8", | |
| "title": "Enhancement of single-channel periodic signals in the time-domain", | |
| "authors": [ | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Jensen", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Benesty", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Christensen", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Jensen", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2012, | |
| "venue": "IEEE Trans on Audio, Speech, and Language Processing", | |
| "volume": "20", | |
| "issue": "7", | |
| "pages": "1948--1963", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "J. Jensen, J. Benesty, M. Christensen, and S. Jensen, \"Enhancement of single-channel periodic signals in the time-domain,\" IEEE Trans on Audio, Speech, and Language Processing, vol. 20, no. 7, pp. 1948-1963, 2012.", | |
| "links": null | |
| }, | |
| "BIBREF9": { | |
| "ref_id": "b9", | |
| "title": "Compressive speech enhancement", | |
| "authors": [ | |
| { | |
| "first": "S", | |
| "middle": [ | |
| "Y" | |
| ], | |
| "last": "Low", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "D", | |
| "middle": [ | |
| "S" | |
| ], | |
| "last": "Pham", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Venkatesh", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2013, | |
| "venue": "Speech Communication", | |
| "volume": "55", | |
| "issue": "6", | |
| "pages": "757--768", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "S. Y. Low, D. S. Pham, and S. Venkatesh, \"Compressive speech enhancement,\" Speech Communication, vol. 55, no. 6, pp. 757-768, July 2013.", | |
| "links": null | |
| }, | |
| "BIBREF10": { | |
| "ref_id": "b10", | |
| "title": "Mixtures of local dictionaries for unsupervised speech enhancement", | |
| "authors": [ | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Kim", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "P", | |
| "middle": [], | |
| "last": "Smaragdis", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2015, | |
| "venue": "IEEE Signal processing letters", | |
| "volume": "22", | |
| "issue": "3", | |
| "pages": "293--297", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "M. Kim, P. Smaragdis, \"Mixtures of local dictionaries for unsupervised speech enhancement, \"IEEE Signal processing letters, vol. 22, no. 3, pp. 293-297, 2015", | |
| "links": null | |
| }, | |
| "BIBREF11": { | |
| "ref_id": "b11", | |
| "title": "Algorithms for non-negative matrix factorization", | |
| "authors": [ | |
| { | |
| "first": "D", | |
| "middle": [ | |
| "D" | |
| ], | |
| "last": "Lee", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [ | |
| "S" | |
| ], | |
| "last": "Seung", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2000, | |
| "venue": "Proceedings of INIP", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "556--562", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "D. D. Lee and H. S. Seung, \"Algorithms for non-negative matrix factorization,\" in Proceedings of INIP, pp.556-562, 2000", | |
| "links": null | |
| }, | |
| "BIBREF12": { | |
| "ref_id": "b12", | |
| "title": "Speech enhancement using a minimum-mean square error short-time spectral amplitude estimator", | |
| "authors": [ | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Ephraim", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Malah", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1984, | |
| "venue": "IEEE Transactions on Audio, Speech, and Language Processing", | |
| "volume": "32", | |
| "issue": "6", | |
| "pages": "1109--1121", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Y. Ephraim and D. Malah, \"Speech enhancement using a minimum-mean square error short-time spectral amplitude estimator,\" IEEE Transactions on Audio, Speech, and Language Processing, 32(6), pp. 1109-1121, 1984.", | |
| "links": null | |
| }, | |
| "BIBREF13": { | |
| "ref_id": "b13", | |
| "title": "Speech enhancement using segmental nonnegative matrix factorization", | |
| "authors": [ | |
| { | |
| "first": "H-T", | |
| "middle": [], | |
| "last": "Fang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J-H", | |
| "middle": [], | |
| "last": "Hung", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "X", | |
| "middle": [], | |
| "last": "Lu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S-S", | |
| "middle": [], | |
| "last": "Wang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Tsao", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2014, | |
| "venue": "Proceedings of the IEEE ICASSP", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "4483--4487", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "H-T. Fang, J-H. Hung, X. Lu, S-S. Wang, Y. Tsao, \"Speech enhancement using segmental nonnegative matrix factorization,\" in Proceedings of the IEEE ICASSP, pp. 4483 -4487, 2014.", | |
| "links": null | |
| }, | |
| "BIBREF14": { | |
| "ref_id": "b14", | |
| "title": "The AURORA experimental framework for the performance evaluations of speech recognition systems under noisy conditions", | |
| "authors": [ | |
| { | |
| "first": "H", | |
| "middle": [ | |
| "G" | |
| ], | |
| "last": "Hirsch", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Pearce", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2000, | |
| "venue": "Proceedings of the ASR", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "H. G. Hirsch and D. Pearce, \"The AURORA experimental framework for the performance evaluations of speech recognition systems under noisy conditions,\" in Proceedings of the ASR, 2000", | |
| "links": null | |
| }, | |
| "BIBREF15": { | |
| "ref_id": "b15", | |
| "title": "Perceptual evaluation of speech quality (PESQ) -a new method for speech quality assessment of telephone networks and codecs", | |
| "authors": [ | |
| { | |
| "first": "A", | |
| "middle": [ | |
| "W" | |
| ], | |
| "last": "Rix", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [ | |
| "G" | |
| ], | |
| "last": "Beerends", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [ | |
| "P" | |
| ], | |
| "last": "Hollier", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [ | |
| "P" | |
| ], | |
| "last": "Hekstra", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2001, | |
| "venue": "Proceedings of ICASSP", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "749--752", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "A. W. Rix, J. G. Beerends, M. P. Hollier and A. P. Hekstra, \"Perceptual evaluation of speech quality (PESQ) -a new method for speech quality assessment of telephone networks and codecs,\" in Proceedings of ICASSP, pp. 749-752, 2001.", | |
| "links": null | |
| }, | |
| "BIBREF16": { | |
| "ref_id": "b16", | |
| "title": "Joint acoustic and modulation frequency", | |
| "authors": [ | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Shihab", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Les", | |
| "middle": [], | |
| "last": "Shamma", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Atlas", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2003, | |
| "venue": "EURASIP Journal on Applied Signal Processing", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Shihab A. Shamma and Les Atlas, \"Joint acoustic and modulation frequency\", EURASIP Journal on Applied Signal Processing, Volume 2003, Jan 2003", | |
| "links": null | |
| }, | |
| "BIBREF17": { | |
| "ref_id": "b17", | |
| "title": "Homomorphic modulation spectra", | |
| "authors": [ | |
| { | |
| "first": "L", | |
| "middle": [], | |
| "last": "Atlas", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Q", | |
| "middle": [], | |
| "last": "Li", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Thompson", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2004, | |
| "venue": "Proceedings of the IEEE ICASSP", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "L. Atlas, Q. Li, and J. Thompson, \"Homomorphic modulation spectra\", in Proceedings of the IEEE ICASSP, 2004.", | |
| "links": null | |
| }, | |
| "BIBREF18": { | |
| "ref_id": "b18", | |
| "title": "FFT-based spectro-temporal analysis and synthesis of sounds", | |
| "authors": [ | |
| { | |
| "first": "C.-C", | |
| "middle": [], | |
| "last": "Hsu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "T.-H", | |
| "middle": [], | |
| "last": "Lin", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "T.-S", | |
| "middle": [], | |
| "last": "Chi", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2011, | |
| "venue": "Proceedings of the IEEE ICASSP", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "C.-C. Hsu, T.-H. Lin and T.-S. Chi, \" FFT-based spectro-temporal analysis and synthesis of sounds\", in Proceedings of the IEEE ICASSP, 2011.", | |
| "links": null | |
| } | |
| }, | |
| "ref_entries": { | |
| "FIGREF0": { | |
| "num": null, | |
| "uris": null, | |
| "type_str": "figure", | |
| "text": "\u5f0f\u4e4b\u6b50\u5e7e\u91cc\u5fb7 \u8ddd\u96e2\u5e73\u65b9\uff0c\u56e0\u6b64 \u8207 \u7684\u8fed\u4ee3\u66f4\u65b0\u516c\u5f0f\u70ba (4) \u5f0f\u548c (5) \u5f0f\u3002 (\u4e8c)\u57fa\u65bc\u975e\u8ca0\u77e9\u9663\u5206\u89e3\u4e4b\u8abf\u8b8a\u983b\u8b5c\u5f37\u5316\u6cd5 (NMF-MSE) \u4e4b\u4ecb\u7d39 \u5728\u6b64\u6211\u5011\u5c07\u4ecb\u7d39\u672c\u8ad6\u6587\u6240\u63d0\u7684\u65b0\u65b9\u6cd5-\u57fa\u65bc\u975e\u8ca0\u77e9\u9663\u5206\u89e3\u4e4b\u8abf\u8b8a\u983b\u8b5c\u5f37\u5316\u6cd5 (NMF-based modulation spectrum enhancement, NMF-MSE)\u3002\u4e00\u822c\u800c\u8a00\uff0cNMF \u5e38\u7528\u4ee5\u5f37\u5316 \u8a9e\u97f3\u4e4b\u6642\u983b\u5716 (spectrogram) [15]\uff0c\u4ea6\u5373\u5176\u8072\u5b78\u983b\u8b5c(acoustic spectrum) \u7684\u6642\u5e8f\u5217 (time series)\uff0c\u800c\u672c\u7ae0\u7684\u65b0\u65b9\u6cd5\u88e1\uff0c\u7c21\u55ae\u4f86\u8aaa\uff0c\u5373\u6211\u5011\u5c07\u5404\u8072\u5b78\u983b\u7387\u7684\u983b\u8b5c\u5f37\u5ea6\u6642\u5e8f\u5217\u53d6\u5176\u5085 \u7acb\u8449\u8f49\u63db (Fourier transform)\u3001\u5f97\u5230\u5176\u8abf\u8b8a\u983b\u8b5c (modulation spectrum) [19,20,21]\u5f8c\uff0c\u5728 \u5c0d\u5176\u5f37\u5ea6\u6210\u5206\u4f5c NMF \u7684\u5f37\u5316\u3002 \u7531\u65bc\u5728\u6b64\u65b0\u65b9\u6cd5\u4e2d\uff0c\u6240\u8981\u5f37\u5316\u66f4\u65b0\u7684\u662f\u8a9e\u97f3\u8072\u5b78\u983b\u8b5c\u5f37\u5ea6\u6642\u5e8f\u5217\u4e4b\u7684\u8abf\u8b8a\u983b\u8b5c\u5176\u5f37\u5ea6\u6210 \u5206 (magnitude part)\uff0c\u5728\u6b64\u6211\u5011\u5148\u7c21\u4ecb\u5982\u4f55\u6c42\u5f97\u8abf\u8b8a\u983b\u8b5c\u3002\u5c0d\u4e00\u6bb5\u8a9e\u97f3\u800c\u8a00\uff0c\u6211\u5011\u6cbf\u8457\u6642 \u9593\u8ef8\u628a\u5b83\u5207\u6210\u4e00\u9023\u4e32\u97f3\u6846 (frame)\uff0c\u5f62\u6210\u97f3\u6846\u7684\u6642\u5e8f\u5217\uff0c\u518d\u5c0d\u6bcf\u500b\u97f3\u6846\u5206\u5225\u53d6\u77ed\u6642\u9593\u5085 \u7acb\u8449\u8f49\u63db (short-time Fourier transform, STFT)\uff0c\u5373\u53ef\u4ee5\u5f97\u5230\u6bcf\u500b\u97f3\u6846\u7684\u77ed\u6642\u9593\u8072\u5b78\u983b\u8b5c (short-time acoustic spectrum)\uff0c\u5f0f\u5b50\u5982\u4e0b:" | |
| }, | |
| "TABREF0": { | |
| "content": "<table><tr><td>(2)\u647a\u7a4d\u6027\u96dc\u8a0a (convolutional noise)\uff1a\u4ea6\u7a31\u4f5c\u901a\u9053\u96dc\u8a0a\uff0c\u96dc\u8a0a\u5728\u8a9e\u97f3\u548c\u6642\u9593\u4e0a\u53ef \u4e8c\u3001\u57fa\u65bc\u975e\u8ca0\u77e9\u9663\u5206\u89e3\u4e4b\u8abf\u8b8a\u983b\u8b5c\u5f37\u5316\u6cd5</td></tr><tr><td>\u7c21\u5316\u6210\u647a\u7a4d (convolution) \u7684\u95dc\u4fc2\uff0c\u50cf\u662f\u96fb\u8a71\u901a\u9053\u6548\u61c9\u548c\u9ea5\u514b\u98a8\u901a\u9053\u6548\u61c9\uff0c\u4ea6\u88ab\u7a31\u4e4b\u901a \u9053\u5931\u771f (channel distortion)\uff0c\u5047\u8a2d\u9ea5\u514b\u98a8\u662f\u56fa\u5b9a\u4f4d\u7f6e\uff0c\u5247\u9020\u6210\u7684\u901a\u9053\u6548\u61c9\u8fd1\u4f3c\u70ba\u975e\u6642\u8b8a\uff0c (\u4e00)\u975e\u8ca0\u77e9\u9663\u5206\u89e3\u6cd5\u4e4b\u4ecb\u7d39</td></tr><tr><td>\u4f46\u82e5\u79fb\u52d5\u5feb\u901f (\u5982\u5728\u9ad8\u901f\u884c\u99db\u7684\u4ea4\u901a\u5de5\u5177\u5167) \u7684\u884c\u52d5\u96fb\u8a71\u901a\u8a0a\uff0c\u5247\u901a\u9053\u6548\u61c9\u660e\u986f\u70ba\u6642 NMF \u57fa\u672c\u65b9\u6cd5\u5373\u662f\u5c07\u4e00\u500b\u975e\u8ca0\u77e9\u9663\u5206\u89e3\u6210\u53e6\u5916\u5169\u500b\u975e\u8ca0\u77e9\u9663\uff0c\u524d\u77e9\u9663\u8fd1\u4f3c\u70ba\u5f8c\u5169</td></tr><tr><td>\u8b8a\u3002 \u500b\u77e9\u9663\u7684\u4e58\u7a4d\uff0c\u6240\u8b02\u975e\u8ca0\u77e9\u9663\uff0c\u662f\u6307\u6b64\u77e9\u9663\u5167\u7684\u5143\u7d20 (element) \u90fd\u662f\u5927\u65bc\u6216\u7b49\u65bc\u96f6\u7684\u5be6</td></tr><tr><td>\u8a9e\u97f3\u5f37\u5316\u76ee\u7684\u662f\u8981\u964d\u4f4e\u96dc\u8a0a\u5c0d\u8a9e\u97f3\u7684\u5e72\u64fe\uff0c\u628a\u96dc\u8a0a\u8a9e\u97f3\u4e2d\u7684\u8a9e\u97f3\u5f37\u8abf\u6216\u9084\u539f\uff0c\u9577\u671f \u6578\u3002\u82e5\u4ee5\u6578\u5b78\u63cf\u8ff0\uff0c\u5373\u70ba\u4efb\u4e00\u5c3a\u5bf8\u70ba \u00d7 \u7684\u975e\u8ca0\u77e9\u9663\uff0c\u900f\u904e NMF \u5206\u6790\uff0c\u53ef\u6c42\u5f97\u5169\u500b \u4e00\u3001\u7dd2\u8ad6 \u8fd1\u6578\u5341\u5e74\u4f86\u79d1\u6280\u4e00\u518d\u7684\u7a81\u7834\uff0c\u96fb\u5b50\u7522\u54c1\u7684\u767c\u5c55\u65e5\u65b0\u6708\u7570\uff0c\u5927\u5e45\u63d0\u5347\u4e86\u751f\u6d3b\u7684\u4fbf\u5229\u6027\u4e5f\u7e2e \u77ed\u4e86\u4eba\u8207\u4eba\u4e4b\u9593\u7684\u8ddd\u96e2\u3002\u8a00\u8a9e\uff0c\u6b64\u4e00\u4eba\u985e\u6e9d\u901a\u6700\u91cd\u8981\u7684\u5a92\u4ecb\uff0c\u91cd\u9ede\u5728\u65bc\u4f7f\u7528\u96d9\u65b9\u7686\u660e\u767d \u7684\u8a9e\u8a00\uff0c\u4e14\u5118\u91cf\u6e05\u6670\u3001\u7c21\u6f54\u3001\u660e\u4e86\u5730\u8868\u9054\u81ea\u5df1\u7684\u89c0\u9ede\uff0c\u4ea4\u6d41\u5f7c\u6b64\u610f\u898b\u8207\u60f3\u6cd5\u3002\u800c\u5728\u7121\u6cd5 \u4ee5\u4f86\u6709\u8a31\u591a\u5b78\u8005\u63d0\u51fa\u6b64\u985e\u7684\u6f14\u7b97\u6cd5\uff0c\u4e00\u822c\u4f86\u8aaa\u9019\u4e9b\u8a9e\u97f3\u5f37\u5316\u6cd5\u53ef\u5206\u70ba\u5169\u7a2e: \u76e3\u7763\u5f0f \u975e\u8ca0\u77e9\u9663 \u548c \uff0c\u5982\u4e0b\u5f0f\u8868\u793a: (supervised) \u548c\u975e\u76e3\u7763\u5f0f (unsupervised)\uff0c\u7c21\u55ae\u4f86\u8aaa\uff0c\u5169\u8005\u5dee\u7570\u5728\u65bc\u8a13\u7df4\u8cc7\u6599\u672c\u8eab\u4e4b\u985e\u5225 V\u2245 (1) \u662f\u5426\u5df2\u77e5\u3002 \u76e3\u7763\u5f0f\u7684\u8a9e\u97f3\u5f37\u5316\u6cd5\u901a\u5e38\u540c\u6642\u4f7f\u7528\u8a9e\u97f3\u548c\u96dc\u8a0a\u5169\u65b9\u7684\u6a21\u578b\uff0c\u4e14\u6bcf\u500b\u6a21\u578b\u7684\u53c3\u6578\u7531\u5404 \u5176\u4e2d\uff1a</td></tr><tr><td>\u9762\u5c0d\u9762\u4ee5\u8a9e\u97f3\u6e9d\u901a\u7684\u9650\u5236\u4e0b\uff0c\u4ee5\u5f80\u900f\u904e\u66f8\u4fe1\u7684\u65b9\u5f0f\uff0c\u9060\u65b9\u50b3\u4f86\u7684\u6587\u5b57\u8a0a\u606f\u52d5\u8f12\u9700\u4e00\u9031\u751a \u81ea\u7684\u8a13\u7df4\u6a23\u672c\u6240\u4f30\u6e2c\u7372\u5f97\uff0c\u4f8b\u5982\u97f3\u7d20\u76f8\u95dc\u975e\u8ca0\u77e9\u9663\u5206\u89e3\u6cd5 (phoneme-dependent NMF) [2]\u3001</td></tr><tr><td>\u81f3\u4e00\u500b\u6708\uff0c\u6216\u662f\u9700\u8981\u9ad8\u984d\u7684\u9577\u9014\u6216\u570b\u969b\u96fb\u8a71\u8cbb\u7528\uff0c\u800c\u62dc\u7576\u4eca\u7db2\u8def\u79d1\u6280\u6240\u8cdc\uff0c\u73fe\u5728\u5e7e\u4e4e\u53ef \u57fa\u65bc\u78bc\u7c3f(codebook)\u7684\u5f37\u5316\u6cd5[3] \u3001\u57fa\u65bc\u8c9d\u6c0f\u975e\u8ca0\u77e9\u9663\u5206\u89e3\u7d50\u5408\u96b1\u85cf\u99ac\u53ef\u592b\u6a21\u578b\u6cd5</td></tr><tr><td>\u8aaa\u662f\u96a8\u50b3\u96a8\u5230\uff0c\u9664\u4e86\u6642\u9593\u7684\u7e2e\u6e1b\uff0c\u50b3\u9001\u6a21\u5f0f\u4e5f\u6709\u91cd\u5927\u7684\u6539\u8b8a\uff0c\u7531\u50b3\u7d71\u7684\u6587\u5b57\u50b3\u9001\u63d0\u5347\u70ba (BNMF-HMM)[4]\u3001\u52a0\u5165\u5171\u7a00\u758f\u6027\u4e4b\u647a\u7a4d\u975e\u8ca0\u77e9\u9663\u5206\u89e3\u6cd5 (convolutive nonnegative matrix</td></tr><tr><td>\u8a9e\u97f3\u50b3\u9001\uff0c\u8b93\u4f7f\u7528\u8005\u53ef\u4ee5\u66f4\u5feb\u901f\u7c21\u6377\u4e14\u4f4e\u6210\u672c\u5730\u50b3\u9001\u81ea\u5df1\u7684\u8a0a\u606f\uff0c\u5982\u5fae\u4fe1\u3001LINE \u7b49\u8457 factorization with cosparsity)[5]\u3002\u975e\u76e3\u7763\u5f0f\u7684\u8a9e\u97f3\u5f37\u5316\u6cd5\u4e26\u4e0d\u8981\u6c42\u4e8b\u5148\u6c42\u5f97\u8a9e\u97f3\u7279\u6027\u53ca\u96dc</td></tr><tr><td>\u540d\u7684\u901a\u8a0a\u61c9\u7528\u7a0b\u5f0f (APP)\uff0c\u800c\u9019\u4e9b\u901a\u8a0a APP \u7684\u84ec\u52c3\u767c\u5c55\u8207\u9032\u6b65\uff0c\u66f4\u662f\u8b93\u6d88\u8cbb\u8005\u591a\u4e86\u8a31 \u8a0a\u7a2e\u985e\uff0c\u76ee\u6a19\u662f\u76f4\u63a5\u5f9e\u96dc\u8a0a\u8a9e\u97f3\u4e2d\u4f30\u6e2c\u4e7e\u6de8\u8a9e\u97f3\uff0c\u6b64\u985e\u8457\u540d\u7684\u65b9\u6cd5\u5305\u62ec\u983b\u8b5c\u6d88\u53bb\u6cd5</td></tr><tr><td>\u591a\u5c0d\u8a71\u7684\u9078\u64c7\uff0c\u7531\u539f\u5148\u7684\u6587\u5b57\u8a0a\u606f\u63d0\u5347\u81f3\u8a9e\u97f3\u901a\u8a71\u3001\u751a\u81f3\u9032\u968e\u5230\u8996\u8a0a\u901a\u8a71\uff0c\u4e0d\u7ba1\u8ddd\u96e2\u591a (spectral subtraction , SS)[6]\u3001\u97cb\u7d0d\u6ffe\u6ce2\u5668\u6cd5 (Wiener filter)[7]\u3001\u5361\u723e\u66fc\u6ffe\u6ce2\u5668 (Kalman</td></tr><tr><td>\u9060\uff0c\u6253\u958b\u624b\u6a5f\uff0c\u5c0d\u65b9\u5f77\u5f7f\u5c31\u80fd\u51fa\u73fe\u5728\u4f60\u9762\u524d\u8207\u4f60\u4ea4\u8ac7\uff0c\u5be6\u53ef\u8b02\u7121\u9060\u5f17\u5c46\u3002 filter)[8]\u3001\u57fa\u65bc\u62c9\u666e\u62c9\u65af\u6700\u5c0f\u5747\u65b9\u8aa4\u5dee\u6cd5\u8a9e\u97f3\u5f37\u5316\u6cd5 (Laplacian-based MMSE estimator</td></tr><tr><td>\u7136\u800c\u5728\u5be6\u969b\u9060\u8ddd\u96e2\u901a\u8a0a\u4e2d\uff0c\u8a9e\u97f3\u50b3\u905e\u7684\u904e\u7a0b\u5fc5\u7136\u6703\u53d7\u5230\u901a\u8a0a\u901a\u9053\u8207\u6536\u767c\u97f3\u8a0a\u4e4b\u5468\u906d\u74b0\u5883 for speech enhancement)[9]\u3001\u55ae\u901a\u9053\u9031\u671f\u8a0a\u865f\u4e4b\u5f37\u5316\u6cd5 (Enhancement of single channel</td></tr><tr><td>\u7684\u5e72\u64fe\uff0c\u5c0d\u5f8c\u8005\u800c\u8a00\uff0c\u610f\u5373\u7576\u8aaa\u8a71\u8005\u5728\u4e00\u500b\u5435\u96dc\u7684\u74b0\u5883\u4e2d\u50b3\u8a71\uff0c\u6b64\u8a9e\u97f3\u6703\u593e\u5e36\u8457\u8a31\u591a\u96dc periodic signals in the time-domain)[10]\u3001\u58d3\u7e2e\u8a9e\u97f3\u5f37\u5316 (compressive speech enhancement)</td></tr><tr><td>\u8a0a\u5c0e\u81f4\u63a5\u6536\u8005\u7121\u6cd5\u660e\u78ba\u63a5\u6536\u8cc7\u8a0a\u3002\u56e0\u6b64\u5728\u672c\u8ad6\u6587\u4e2d\uff0c\u6211\u5011\u91dd\u5c0d\u964d\u4f4e\u96dc\u8a0a\u5e72\u64fe\u800c\u8a0e\u8ad6\u7814\u767c [11]\u3001\u57fa\u65bc\u8c9d\u6c0f\u975e\u8ca0\u77e9\u9663\u5206\u89e3\u4e4b\u7dda\u4e0a\u66f4\u65b0\u6cd5 (Online BNMF)[4]\u3001\u5c40\u90e8\u8a5e\u5178\u6df7\u5408 (mixtures</td></tr><tr><td>\u8a9e\u97f3\u5f37\u5316\u7684\u6280\u8853\uff0c\u76ee\u7684\u5c31\u662f\u8981\u964d\u4f4e\u96dc\u8a0a\u5c0d\u8a9e\u97f3\u7684\u5e72\u64fe\uff0c\u53ea\u5f97\u4eba\u8033\u5728\u807d\u89ba\u65b9\u9762\u80fd\u66f4\u52a0\u6e05\u6670 of local dictionaries)[12]\u3002\u76e3\u7763\u5f0f\u7684\u65b9\u6cd5\u56e0\u70ba\u4e8b\u524d\u8cc7\u8a0a\u8f03\u8c50\u5bcc\uff0c\u82e5\u904b\u7528\u5f97\u7576\u901a\u5e38\u6548\u679c\u512a\u65bc</td></tr><tr><td>\u63a5\u6536\u8a9e\u97f3\u3002 \u975e\u76e3\u7763\u5f0f\u7684\u65b9\u6cd5\uff0c\u4f46\u4ecd\u9700\u8996\u65b9\u6cd5\u672c\u8eab\u7684\u8907\u96dc\u5ea6\u8207\u5047\u8a2d\u662f\u5426\u543b\u5408\u4e8b\u5be6\u800c\u5b9a\u3002</td></tr><tr><td>\u7279\u5225\u4e00\u63d0\u7684\u662f\uff0c\u8a9e\u97f3\u8fa8\u8b58\u53ef\u878d\u5165\u5230\u6c7d\u8eca\u529f\u80fd\u4e0a\uff0c\u4f7f\u6c7d\u8eca\u4e0d\u518d\u662f\u4e00\u500b\u55ae\u7d14\u4ea4\u901a\u904b\u8f38\u7684\u6a5f\u68b0 \u672c\u8ad6\u6587\u4e3b\u8981\u662f\u4f7f\u7528\u975e\u8ca0\u77e9\u9663\u5206\u89e3\u6cd5 (nonnegative matrix factorization, NMF)[13] \u4f86</td></tr><tr><td>\u7522\u54c1\u3002\u5fae\u8655\u7406\u5668\u3001\u96fb\u8166\u8f14\u52a9\u7b49\u63a7\u5236\u4e86\u6c7d\u8eca\u7cfb\u7d71\uff0c\u5305\u542b\u5f15\u64ce\u3001\u50b3\u52d5\u3001ABS(Anti-lock Braking System) \u524e\u8eca\u7b49\u539f\u672c\u57fa\u672c\u7684\u6a5f\u68b0\u7cfb\u7d71\uff0c\u800c\u8fd1\u5e74\u4f86\u66f4\u662f\u589e\u52a0\u4e86\u901a\u8a0a\u5a1b\u6a02\u5c0e\u822a\u7cfb\u7d71\u53ca\u5404\u5f0f \u5404\u6a23\u7684\u901a\u8a0a\u8a2d\u5099\uff0c\u4f7f\u6c7d\u8eca\u4e0d\u518d\u662f\u4ee5\u5f80\u55ae\u7d14\u7684\u6a5f\u68b0\u7cfb\u7d71\uff0c\u4f46\u96a8\u8457\u6c7d\u8eca\u529f\u80fd\u8d8a\u4f86\u8d8a\u591a\uff0c\u76f8\u5c0d \u767c\u5c55\u5f37\u5316\u8a9e\u97f3\u6280\u8853\uff0cNMF \u6cd5\u662f\u7531\u8c9d\u723e\u5be6\u9a57\u5ba4 (Bell Laboratory) \u7684 D.D. Lee \u53ca\u9ebb\u7701\u7406\u5de5 \u4ee3\u8868 \u8207 \u7684\u5dee\u8ddd\u6216\u903c\u8fd1\u7a0b\u5ea6\uff0c\u85c9\u7531\u6700\u5c0f\u5316\u6b64\u6210\u672c\u51fd\u6578\u7684\u65b9\u5f0f\u4f7f \u66f4\u6709\u6548\u7684\u8fd1\u4f3c \uff0c \u5b78\u9662 (Massachusetts Institute of Technology, MIT) \u7684 H.S. Seung \u6240\u767c\u5c55\u51fa\u4f86\u7684\u6f14\u7b97\u6cd5\uff0c \u8d77\u521d\u7528\u5728\u5f71\u50cf\u8655\u7406\uff0c\u5f8c\u7e8c\u624d\u6f38\u6f38\u88ab\u4f7f\u7528\u5728\u8a9e\u97f3\u5f37\u5316\u4e0a\u3002\u57fa\u65bc NMF \u5206\u6790\u6cd5\uff0c\u6211\u5011\u63d0\u51fa\u4e86 \u4e0b\u8ff0\u6240\u63d0\u5230\u7684\u662f\u5169\u7a2e\u6bd4\u8f03\u5e38\u7528\u5230\u7684\u6210\u672c\u51fd\u6578:</td></tr><tr><td>\u7684\u6309\u9375\u63a7\u5236\u9215\u4e5f\u8d8a\u4f86\u8d8a\u8907\u96dc\uff0c\u99d5\u99db\u4eba\u9762\u5c0d\u50cf\u662f\u98db\u6a5f\u99d5\u99db\u8259\u7684\u5100\u8868\u6309\u9215\uff0c\u901a\u5e38\u53ea\u6703\u4e82\u4e86\u624b \u66f4\u65b0\u8a9e\u97f3\u8abf\u8b8a\u983b\u8b5c (modulation spectrum) \u7684\u5f37\u5316\u6280\u8853\u3002</td></tr><tr><td>\u8173\u3001\u4e0d\u77e5\u5982\u4f55\u8f15\u9b06\u99d5\u99ad\uff0c\u4f7f\u7528\u9019\u4e9b\u4e94\u82b1\u516b\u9580\u7684\u529f\u80fd\u3002\u70ba\u4e86\u7c21\u5316\u9019\u4e9b\u6309\u9215\uff0c\u8a31\u591a\u8eca\u5ee0\u671d\u5411 \u5c0d\u4e00\u6bb5\u8072\u97f3\u800c\u8a00\uff0c\u6211\u5011\u6cbf\u8457\u6642\u9593\u8ef8\u628a\u5b83\u5207\u6210\u97f3\u6846 (frame) \u7684\u5e8f\u5217\uff0c\u7576\u6bcf\u500b\u97f3\u6846\u900f\u904e</td></tr><tr><td>\u4f7f\u7528\u8a9e\u97f3\u7684\u65b9\u5f0f\uff0c\u76ee\u7684\u5c31\u662f\u8981\u8b93\u99d5\u99db\u300c\u773c\u775b\u4e0d\u96e2\u958b\u8def\u9762\uff0c\u624b\u4e0d\u96e2\u958b\u65b9\u5411\u76e4\u300d\u4e26\u4e14\u53ef\u548c\u8eca \u5085\u7acb\u8449\u8f49\u63db\u5f8c\uff0c\u5f97\u5230\u77ed\u6642\u9593\u8072\u5b78\u983b\u8b5c (short-time acoustic spectrum)\uff0c\u63a5\u8457\u518d\u5c0d\u6bcf\u500b\u8072\u5b78</td></tr><tr><td>\u5b50\u505a\u76f4\u63a5\u7684\u6e9d\u901a\uff0c\u50cf\u662f\u798f\u7279\u548c\u5fae\u8edf\u5408\u4f5c\u958b\u767c\u7684 SYNC[1]\uff0c\u53ef\u4ee5\u85c9\u7531\u8a9e\u97f3\u7684\u547d\u4ee4\u9054\u5230\u63a7\u5236 \u983b\u7387\u9ede\u7684\u6642\u5e8f\u5217\u518d\u53d6\u4e00\u6b21\u5085\u7acb\u8449\u8f49\u63db\uff0c\u5373\u53ef\u5f97\u5230\u5404\u8072\u5b78\u983b\u7387\u9ede\u4e4b\u8abf\u8b8a\u983b\u8b5c\u3002\u800c\u672c\u8ad6\u6587\u6240</td></tr><tr><td>\u6548\u679c\uff0c\u5982:FM \u983b\u7387\u9078\u64c7\u3001\u8abf\u6574\u51b7\u6c23\u6eab\u5ea6\u53ca\u958b\u95dc\u3001\u8072\u63a7\u64a5\u865f\u7b49\uff0c\u4f46\u662f\u6c7d\u8eca\u8a9e\u97f3\u8fa8\u8b58\u6700\u5927\u7684 \u4ee5\u51fa\u7684\u65b9\u6cd5\u662f\u85c9\u7531 NMF \u6cd5\u5c0d\u65bc\u4e0a\u8ff0\u4e4b\u8abf\u8b8a\u983b\u8b5c\u7684\u5f37\u5ea6\u505a\u66f4\u65b0\uff0c\u85c9\u7531\u8a13\u7df4\u8cc7\u6599\u4e2d\u7684\u4e7e\u6de8</td></tr><tr><td>\u96e3\u8655\u4e4b\u4e00\uff0c\u5728\u65bc\u884c\u99db\u4e2d\u4e4b\u8eca\u5916\u96dc\u8a0a\u5e72\u64fe\uff0c\u5982\u98a8\u5207\u8072\u3001\u5f15\u64ce\u8072\u3001\u8f2a\u80ce\u7684\u6efe\u52d5\u96dc\u8a0a\u7b49\uff0c\u9019\u4e9b \u8a9e\u97f3\u53ca\u96dc\u8a0a\uff0c\u5206\u5225\u8a13\u7df4\u5169\u65b9\u4e4b\u8abf\u8b8a\u983b\u8b5c\u5f37\u5ea6\u7684 NMF \u57fa\u5e95 (basis)\uff0c\u518d\u5229\u7528\u5169\u65b9\u7684\u57fa\u5e95\u4f86</td></tr><tr><td>\u8072\u97f3\u6703\u7834\u58de\u539f\u59cb\u8a9e\u97f3\uff0c\u9032\u800c\u964d\u4f4e\u8eca\u5167\u8a9e\u97f3\u8fa8\u8b58\u7cfb\u7d71\u7684\u8fa8\u8b58\u5ea6\uff0c\u4f7f\u6574\u9ad4\u667a\u6167\u578b\u7cfb\u7d71\u505a\u51fa\u932f \u5206\u89e3\u6e2c\u8a66\u8a9e\u97f3\u4e4b\u8abf\u8b8a\u983b\u8b5c\u5f37\u5ea6\uff0c\u5f97\u5230\u8fd1\u4f3c\u4e7e\u6de8\u8a9e\u97f3\u7684\u6210\u5206\u5f8c\uff0c\u914d\u5408\u539f\u59cb\u76f8\u89d2\u3001\u7d93\u7531\u53cd\u5085</td></tr><tr><td>\u8aa4\u7684\u56de\u61c9\uff0c\u53cd\u800c\u9055\u80cc\u7576\u521d\u8a2d\u8a08\u7684\u7406\u5ff5\uff0c\u7531\u6b64\u53ef\u77e5\u8a9e\u97f3\u5f37\u5316\u5728\u6b64\u7684\u91cd\u8981\u6027\u3002\u4f46\u5728\u964d\u4f4e\u96dc\u8a0a \u7acb\u8449\u8f49\u63db (inverse Fourier transform) \u5f97\u5230\u65b0\u7684\u8072\u5b78\u983b\u8b5c\u6642\u5e8f\u5217\uff0c\u5f9e\u800c\u53ef\u5f97\u5f37\u5316\u8a9e\u97f3\u8a0a\u865f\u3002</td></tr><tr><td>\u7684\u904e\u7a0b\u4e2d\uff0c\u904e\u5ea6\u7684\u964d\u566a\u53c8\u6703\u9020\u6210\u539f\u59cb\u8a9e\u97f3\u5931\u771f\uff0c\u6240\u4ee5\u5728\u964d\u4f4e\u96dc\u8a0a\u7684\u540c\u6642\u5fc5\u9808\u6e1b\u5c11\u5c0d\u539f\u59cb \u53e6\u5916\uff0c\u70ba\u4e86\u964d\u4f4e\u904b\u7b97\u8907\u96dc\u5ea6\uff0c\u6211\u5011\u53e6\u5916\u63d0\u51fa\u4e86\u5169\u7a2e\u8b8a\u578b\uff1a\u4e00\u7a2e\u662f\u5c07\u76f8\u9130\u7684\u8072\u5b78\u983b\u7387\u9ede\u4e00</td></tr><tr><td>\u8a9e\u97f3\u7684\u640d\u58de\uff0c\u662f\u8a9e\u97f3\u5f37\u5316\u6280\u8853\u9700\u517c\u9867\u7684\u689d\u4ef6\u3002 \u4f75\u8655\u7406\u3001\u53e6\u4e00\u7a2e\u5247\u662f\u53ea\u8655\u7406\u4f4e\u983b\u7387\u5340\u57df\u7684\u8abf\u8b8a\u983b\u8b5c\uff0c\u6839\u64da\u5be6\u9a57\u7d50\u679c\uff0c\u9019\u4e9b\u65b0\u65b9\u6cd5\u90fd\u53ef\u4ee5</td></tr><tr><td>\u5728 \u73fe \u5be6 \u7684 \u74b0 \u5883 \u4e2d \u8a9e \u97f3 \u8a0a \u865f \u5bb9 \u6613 \u53d7 \u5230 \u5916 \u5728 \u74b0 \u5883 \u7684 \u5f71 \u97ff \uff0c \u964d \u4f4e \u4e86 \u8a9e \u97f3 \u7684 \u53ef \u8b80 \u6027 \u6709\u6548\u63d0\u5347\u96dc\u8a0a\u8a9e\u97f3\u7684\u54c1\u8cea\u3001\u4ea6\u5373\u964d\u4f4e\u96dc\u8a0a\u5e72\u64fe\u7684\u6548\u61c9\u3002</td></tr><tr><td>(intelligibility) \u8207\u54c1\u8cea (quality)\uff0c\u4f9d\u64da\u96dc\u8a0a\u7279\u6027\u662f\u5426\u96a8\u8457\u6642\u9593\u660e\u986f\u8b8a\u5316\u53ef\u5206\u70ba\u5169\u5927\u985e\uff0c \u6700\u5f8c\u503c\u5f97\u4e00\u63d0\u7684\u662f\uff0c\u6211\u5011\u5c07\u6240\u63d0\u7684\u65b0\u65b9\u6cd5\u8207\u4e09\u7a2e\u8457\u540d\u8a9e\u97f3\u5f37\u5316\u6cd5\u505a\u6bd4\u8f03\uff0c\u5206\u5225\u70ba\u983b</td></tr><tr><td>\u82e5\u8b8a\u5316\u6162\u6216\u6c92\u6709\u986f\u8457\u7684\u8b8a\u5316\uff0c\u7a31\u4f5c\u7a69\u614b\u96dc\u8a0a (stationary noise)\uff0c\u53cd\u4e4b\u5247\u7a31\u4f5c\u975e\u7a69\u614b\u96dc\u8a0a \u8b5c\u6d88\u53bb\u6cd5 (spectral subtraction, SS)[6]\u3001\u97cb\u7d0d\u6ffe\u6ce2\u5668\u6cd5 (Wiener filter)[7] \u4ee5\u53ca\u6700\u5c0f\u5747\u65b9\u8aa4</td></tr><tr><td>(non-stationary noise)\u3002\u800c\u6839\u64da\u8207\u8a9e\u97f3\u548c\u96dc\u8a0a\u4f86\u6e90\u7684\u7d44\u5408\u95dc\u4fc2\uff0c\u96dc\u8a0a\u4f86\u6e90\u5206\u70ba\u5169\u985e\uff1a \u5dee\u77ed\u6642\u9593\u983b\u8b5c\u632f\u5e45\u4f30\u6e2c\u6cd5 (minimum mean-squared error short-time spectral amplitude</td></tr><tr><td>(1)\u52a0\u6210\u6027\u96dc\u8a0a (additive noise)\uff1a\u4ea6\u7a31\u4f5c\u80cc\u666f\u96dc\u8a0a\uff0c\u96dc\u8a0a\u5728\u8a9e\u97f3\u548c\u6642\u9593\u4e0a\u6210\u7dda\u6027\u76f8 estimation, MMSE-STSA) [14]\uff0c\u521d\u6b65\u5be6\u9a57\u7d50\u679c\u53ef\u770b\u51fa\uff0c\u6211\u5011\u6240\u63d0\u7684\u65b0\u65b9\u6cd5\u5728\u67d0\u4e9b\u96dc\u8a0a\u74b0</td></tr><tr><td>\u52a0 (linear addition) \u7684\u95dc\u4fc2\uff0c\u6b64\u985e\u96dc\u8a0a\u53ef\u9032\u4e00\u6b65\u5206\u70ba\u975e\u6642\u8b8a\u52a0\u6210\u6027\u96dc\u8a0a\u8207\u6642\u8b8a\u52a0\u6210\u6027\u96dc \u5883\u4e0b\u80fd\u6bd4\u9019\u4e09\u7a2e\u65b9\u6cd5\u5f97\u5230\u66f4\u4f73\u7684\u5f37\u5316\u6548\u679c\u3002</td></tr><tr><td>\u8a0a\uff0c\u524d\u8005\u50cf\u662f\u8eca\u8072\u3001\u6a5f\u5668\u767c\u51fa\u7684\u8072\u97f3\u7b49\uff0c\u5f8c\u8005\u5247\u5305\u542b\u5e02\u96c6\u96dc\u8a0a (babble) \u8207\u8b66\u5831\u96dc\u8a0a</td></tr><tr><td>(siren noise) \u7b49\u3002</td></tr></table>", | |
| "num": null, | |
| "type_str": "table", | |
| "html": null, | |
| "text": "\u95dc\u9375\u8a5e\uff1a\u975e\u8ca0\u77e9\u9663\u5206\u89e3\u6cd5\u3001\u8a9e\u97f3\u5f37\u5316\u3001\u6642\u983b\u5716\u3002Keywords: non-negative matrix factorization, speech enhancement, modulation spectrum, spectrogram." | |
| }, | |
| "TABREF2": { | |
| "content": "<table><tr><td>\u7387\u7bc4\u570d\u70ba 0 \u81f3 16 Hz\uff0c\u4e14\u4ee5 4 Hz \u7684\u6210\u5206\u70ba\u6700\u91cd\u8981\u3002\u6211\u5011\u4f7f\u7528\u9019\u6a23\u7684\u89c0\u5ff5\uff0c\u4f86\u7c21\u5316 4.1 \u7bc0 \u56db\u3001\u5be6\u9a57\u6578\u64da\u8207\u8a0e\u8ad6 Baseline 1.8266 2.1700 2.5367 2.8015 3.1490 \u65b0\u7684\u6642\u983b\u5716\uff0c\u9032\u4e00\u6b65\u5f97\u5230\u5f37\u5316\u5f8c\u7684\u8a9e\u97f3\u8a0a\u865f\u3002\u540c\u6642\uff0c\u6211\u5011\u5c0d\u4e0a\u8ff0\u65b0\u65b9\u6cd5\u9032\u4e00\u6b65\u63d0\u51fa\u4e86\u5169</td></tr><tr><td>\u4e2d\u6240\u63d0\u7684 NMF-MSE \u6cd5\uff0c\u4ea6\u5373\u6211\u5011\u53ea\u5c0d\u65bc\u8abf\u8b8a\u983b\u8b5c\u7684\u4f4e\u983b\u5e36\u6210\u5206\u52a0\u4ee5\u5f37\u5316\uff0c\u800c\u5c07\u5269\u9918\u7684 \u9ad8\u983b\u5e36\u6210\u5206\u4fdd\u7559\u4e0d\u52d5\u3002\u5716\u4e8c\u70ba\u793a\u610f\u5716\uff0c\u6211\u5011\u5728\u6b64\u5b9a\u7fa9\u4e86\u4e00\u500b\u53c3\u6578\uff0c\u7a31\u505a\u4f4e\u983b\u5e36\u76f8\u5c0d\u5168\u983b \u5e36\u7684\u983b\u5bec\u6bd4\u4f8b (Low-to-full ratio)\uff0c\u7c21\u5beb\u70ba \uff0c\u4f8b\u5982\u7576 = 0.25\uff0c\u5247\u4ee3\u8868\u53ea\u66f4\u65b0\u5168\u8abf \u672c\u7bc0\u5c07\u7531\u56db\u90e8\u5206\u6240\u7d44\u6210\u3002 (\u4e00) \u539f\u59cb NMF-MSE \u4e4b\u5be6\u9a57\u7d50\u679c NMF-MSE = 1 2.2215 2.3892 2.7011 2.9321 3.2008 = 0.75 2.2176 2.3815 2.6931 2.9387 3.1991 = 0.5 2.1632 2.3476 2.6819 2.9267 3.1967 \u53c3\u6578\u8a2d\u5b9a\u70ba 3\uff0c\u76f8\u7576\u65bc\u8072\u5b78\u983b\u7387\u89e3\u6790\u5ea6\u964d\u70ba\u539f\u59cb\u76841 3 \u7a2e\u65b9\u5f0f\u4f86\u964d\u4f4e\u6f14\u7b97\u8907\u96dc\u5ea6\uff0c\u5206\u5225\u70ba\u4e0d\u540c\u6bd4\u4f8b\u4e4b\u4f4e\u8abf\u8b8a\u983b\u5e36\u4e4b NMF-MSE \u66f4\u65b0\u53ca\u4e0d\u540c\u8072 \u2044 \uff0c\u5728\u6b64\u8f03\u4f4e\u7684\u904b\u7b97\u8907\u96dc\u5ea6\u689d\u4ef6 \u5b78\u983b\u7387\u9ede\u6578\u4e4b\u5168\u8abf\u8b8a\u983b\u5e36\u4e4b NMF-MSE \u66f4\u65b0\uff0c\u5f9e PESQ \u70ba\u6307\u6a19\u7684\u8a9e\u97f3\u54c1\u8cea\u8a55\u4f30\u4e0a\uff0c\u4e0a\u8ff0 \u4e0b\uff0c\u4ecd\u53ef\u5f97\u5230\u986f\u8457\u7684\u8a9e\u97f3\u5f37\u5316\u6548\u679c\u3002 \u65b9\u6cd5\u7686\u80fd\u6709\u6548\u63d0\u5347\u96dc\u8a0a\u8a9e\u97f3\u7684\u6e05\u6670\u5ea6\u3002</td></tr><tr><td>\u8b8a\u983b\u5e36\u524d 25%\u7684\u4f4e\u983b\u5e36\u6210\u5206\u3002\u4f7f\u7528\u4f4e\u65bc 1 \u7684 \u6cd5\u7684\u6f14\u7b97\u8907\u96dc\u5ea6\uff0c\u56e0\u70ba\u9700\u5f37\u5316\u7684\u8abf\u8b8a\u983b\u8b5c\u5f37\u5ea6\u5411\u91cf\u7684\u5c3a\u5bf8\u8b8a\u5c0f\u4e86\u3002\u4f46 \u5e36\u4f86\u7684\u597d\u8655\u7576\u7136\u662f\u53ef\u4ee5\u964d\u4f4e NMF-MSE \u7576\u7136\u4e0d\u80fd\u7121\u9650 \u5236\u5730\u8b8a\u5c0f\uff0c\u5426\u5247\u8655\u7406\u904e\u7a84\u7684\u4f4e\u8abf\u8b8a\u983b\u5e36\u5c07\u7121\u52a9\u65bc\u964d\u4f4e\u96dc\u8a0a\u7684\u6210\u5206\u3002 \u5728\u9019\u88e1\uff0c\u6211\u5011\u9a57\u8b49\u6240\u63d0\u51fa\u4e4b\u539f\u59cb NMF-MSE \u6cd5\u5c0d\u65bc\u53d7\u5230\u96dc\u8a0a\u5e72\u64fe\u4e4b\u8a9e\u97f3\u7684\u5f37\u5316\u6548\u679c\u3002\u7279 = 0.25 2.0703 2.2486 2.6280 2.8880 \u7279\u5225\u4e00\u63d0\u7684\u662f\uff0cNMF-MSE \u6cd5\u662f\u91dd\u5c0d\u8a9e\u97f3\u6642\u983b\u5716\u7684\u8abf\u8b8a\u983b\u8b5c\u52a0\u4ee5\u66f4\u65b0\uff0c\u5c31\u6211\u5011\u6240\u77e5\uff0c 3.1885 \u8868\u56db\uff1a\u8b66\u5831\u96dc\u8a0a\u74b0\u5883\u4e0b\uff0c\u57fa\u790e\u5be6\u9a57\u3001NMF-MSE (\u5176 \u8a2d\u70ba 3\u3001 \u8a2d\u70ba 1)\u8207\u4e09\u7a2e\u8a9e \u76ee\u524d\u57fa\u65bc NMF \u7684\u8a9e\u97f3\u5f37\u5316\u6cd5\u5927\u591a\u662f\u91dd\u5c0d\u6642\u983b\u5716\u672c\u8eab\u4f5c\u5f37\u5316\uff0c\u751a\u5c11\u6709\u9032\u4e00\u6b65\u8655\u7406\u8abf\u8b8a\u983b \u5225\u8aaa\u660e\u7684\u662f\uff0c\u6b64\u6642\u539f\u59cb NMF-MSE \u6cd5\u63a1\u7528\u4e86\u55ae\u4e00\u8072\u5b78\u983b\u7387(\u5373 \u53c3\u6578\u8a2d\u70ba 1)\u4e26\u57f7\u884c\u65bc \u5168\u8abf\u8b8a\u983b\u5e36(\u5373 \u53c3\u6578\u8a2d\u70ba 1)\u3002 \u8868\u4e00\u5217\u51fa\u4e86\u53d7\u96dc\u8a0a\u5e72\u64fe\u4e4b\u8a9e\u97f3\u5176\u5f37\u5316\u5f8c\u7684 PESQ \u503c\u3002 \u97f3\u5f37\u5316\u6cd5(SS, WF, MMSE-STSA)\u6240\u5f97\u7684 PESQ \u503c \u8b5c\uff0c\u56e0\u6b64\u6211\u5011\u7684\u65b0\u65b9\u6cd5\u76f8\u7576\u65bc\u62d3\u5c55\u4e86 NMF \u6cd5\u5728\u8a9e\u97f3\u5f37\u5316\u4e0a\u7684\u61c9\u7528\u3002\u5728\u672a\u4f86\u5c55\u671b\u4e2d\uff0c\u6211 (\u4e09)\u4e0d\u540c\u8072\u5b78\u983b\u7387\u9ede\u6578\u4e4b\u5168\u8abf\u8b8a\u983b\u5e36\u4e4b NMF-MSE \u4e4b\u5be6\u9a57\u7d50\u679c SNR 0 dB 5 dB 10 dB 15 dB 20 dB \u5011\u5e0c\u671b\u53ef\u4ee5\u5c07 NMF-MSE \u6cd5\u7d50\u5408\u5176\u4ed6\u7a2e\u8a9e\u97f3\u5f37\u5316\u6cd5\u9054\u5230\u6291\u5236\u96dc\u8a0a\u6216\u63d0\u5347\u8a9e\u97f3\u54c1\u8cea\uff0c\u5982</td></tr><tr><td>\uf06c \u964d\u4f4e\u983b\u7387\u89e3\u6790\u5ea6\u4e4bNMF-MSE \u5728\u524d\u4e00\u7bc0\u6240\u8ff0\u7684 NMF-MSE \u6cd5\u4e2d\uff0c\u6211\u5011\u662f\u91dd\u5c0d\u500b\u5225\u8072\u5b78\u983b\u7387\u9ede\u7684\u8abf\u8b8a\u983b\u8b5c\u52a0\u4ee5\u5f37 \u5316\uff0c\u4f46\u7531\u65bc\u76f8\u9130\u8072\u5b78\u983b\u7387\u7684\u7279\u6027\u901a\u5e38\u76f8\u4f3c\uff0c\u5982\u679c\u6211\u5011\u5c07\u76f8\u9130\u8072\u5b78\u983b\u7387\u8996\u70ba\u4e00\u9ad4\u3001\u5176\u8abf\u8b8a \u983b\u8b5c\u5f37\u5ea6\u5171\u540c\u7528\u4ee5 NMF \u7684\u57fa\u5e95\u6c42\u53d6\u7684\u8a13\u7df4\u4e0a\uff0c\u6b64\u6642\u56e0\u70ba\u8a13\u7df4\u8cc7\u6599\u91cf\u7684\u589e\u52a0\uff0c\u57fa\u5e95\u77e9\u9663 \u672c\u8eab\u61c9\u8a72\u66f4\u5177\u4ee3\u8868\u6027\u3001\u6709\u52a9\u65bc\u6e2c\u8a66\u968e\u6bb5\u7684\u8abf\u8b8a\u983b\u8b5c\u5f37\u5316\u4e0a\uff0c\u5716\u4e09\u70ba\u793a\u610f\u5716\uff0c\u6211\u5011\u5728\u6b64\u5b9a \u5716\u4e8c\uff1a\u53c3\u6578 \u4e4b\u793a\u610f\u5716 \u5f9e\u8868\u4e00\u660e\u986f\u770b\u51fa\uff0c\u96dc\u8a0a\u5c0d\u65bc\u8a9e\u97f3\u54c1\u8cea\u7686\u6709\u660e\u986f\u7684\u5f71\u97ff\uff0c\u539f\u59cb NMF-MSE \u6cd5\u5728\u96dc\u8a0a\u5e72\u64fe\u7684 \u5728\u672c\u7bc0\uff0c\u6211\u5011\u5c07\u5448\u73fe\u4e0d\u540c\u8072\u5b78\u983b\u7387\u9ede\u6578\u4e4b\u5168\u8abf\u8b8a\u983b\u5e36\u4e4b NMF-MSE \u66f4\u65b0\u5f8c\u4e4b PESQ \u503c\uff0c Baseline 1.8266 2.1700 2.5367 2.8015 3.1490 \u672c\u8ad6\u6587\u4e2d\u7528\u4ee5\u6bd4\u8f03\u7684\u7684\u983b\u8b5c\u6d88\u53bb\u6cd5 (spectral subtraction, SS)\u3001\u97cb\u7d0d\u6ffe\u6ce2\u6cd5 (Wiener \u4e94\u7a2e\u8a0a\u96dc\u6bd4 (SNR) \u7a0b\u5ea6\u4e4b\u74b0\u5883\u4e0b\uff0c\u90fd\u80fd\u660e\u986f\u63d0\u5347 PESQ \u503c\uff0c\u521d\u6b65\u9a57\u8b49\u4e86\u6b64\u65b0\u65b9\u6cd5\u78ba\u5be6 \u5982\u524d\u7ae0\u6240\u8ff0\uff0c\u53c3\u6578 \u4ee3\u8868\u4e86\u540c\u6642\u88ab NMF-MSE \u66f4\u65b0\u4e4b\u76f8\u9130\u8072\u5b78\u983b\u7387\u7684\u9ede\u6578\u3002\u5728\u672c\u7bc0 NMF-MSE 2.2453 2.3969 2.7062 2.9356 3.2015 filtering, WF) \u8207\u5e73\u5747\u6700\u5c0f\u5316\u8aa4\u5dee\u77ed\u6642\u983b\u8b5c\u632f\u5e45\u4f30\u6e2c\u6cd5 (minimum mean-square error \u5728\u6291\u5236\u96dc\u8a0a\u6548\u61c9\u4e0a\u6709\u6240\u5e6b\u52a9\u3002\u4f8b\u5982\u8207\u57fa\u790e\u5be6\u9a57\u76f8\u6bd4\uff0cNMF-MSE \u6cd5\u4f7f PESQ \u503c\u5e73\u5747\u63d0\u5347 \u4e86 10%\uff0c\u5176\u4e2d\u53c8\u4ee5 SNR \u70ba 0dB \u6642\u6700\u70ba\u660e\u986f\uff0c\u63d0\u5347\u4e86 21%\u3002 \u8868\u4e00\uff1a\u57fa\u790e\u5be6\u9a57\u53ca\u539f\u59cb NMF-MSE \u6cd5\u6240\u5f97\u7684 PESQ \u503c \u5be6\u9a57\u4e2d\u6b64\u53c3\u6578 \u5206\u5225\u88ab\u8a2d\u5b9a\u70ba 1\u30012\u30013 \u8207 4\uff0c\u96a8\u8457 \u503c\u7684\u589e\u52a0\uff0c\u4ee3\u8868\u4e86 NMF-MSE \u6240 \u8655\u7406\u4e4b\u8072\u5b78\u983b\u7387\u89e3\u6790\u5ea6\u76f8\u5c0d\u8b8a\u5c0f\u3002\u6b64\u6642\uff0cNMF-MSE \u6cd5\u4e2d\u7684 SS 1.9320 2.2362 2.7176 2.9483 3.2350 short-time spectral amplitude estimation, MMSE-STSA)\u3002\u6b64\u5916\uff0c\u6211\u5011\u5c07\u6e2c\u8a66\u64f4\u5c55 NMF-MSE \u53c3\u6578\u503c\u56fa\u5b9a\u70ba 1\uff0c\u4ee3\u8868 WF 1.6569 2.2042 2.6856 2.9632 3.3036 \u6cd5\u7684\u904b\u7528\uff0c\u5305\u62ec\u7528\u5728\u591a\u8a9e\u8005\u8207\u591a\u985e\u578b\u96dc\u8a0a\u74b0\u5883\u7684\u8a9e\u97f3\u5f37\u5316\u4e0a\uff0c\u53e6\u5916\u4e5f\u53ef\u4ee5\u66f4\u9032\u4e00\u6b65\u5728\u5176 \u8655\u7406\u7684\u5c0d\u8c61\u662f\u5168\u983b\u5e36\u4e4b\u8abf\u8b8a\u983b\u8b5c\u3002 MMSE-STSA 2.0098 2.4355 2.7472 3.0262 3.2873 \u4ed6\u8cc7\u6599\u5eab\u4e0a\u8655\u7406 (\u5982\u4e2d\u6587\u6578\u5b57\u8a9e\u97f3\u6216\u662f\u66f4\u591a\u5b57\u5f59\u7684\u8cc7\u6599\u5eab) \uff0c\u4ee5\u63a2\u8a0e\u5176\u5be6\u969b\u5c64\u9762\u61c9\u7528\u7684</td></tr><tr><td>\u7fa9\u4e00\u500b\u53c3\u6578 \u4e09\u3001\u5be6\u9a57\u74b0\u5883\u8a2d\u5b9a \u96dc\u8a0a\u70ba siren\uff0c\u56fa\u5b9a = 1\u8207 \u8868\u4e09\u5217\u51fa\u4e86\u53d7\u96dc\u8a0a\u5e72\u64fe\u4e4b\u8a9e\u97f3\u7d93\u7531 NMF-MSE \u5f37\u5316\u5f8c\u7684 PESQ \u503c\u3002\u5f9e\u9019\u8868\uff0c\u6211\u5011\u770b\u5230\u5728 = 1\u4e4b NMF-MSE \u6cd5\u4e4b PESQ \u7d50\u679c \u50f9\u503c\u3002 (block)\uff0c\u4ee3\u8868\u4e86\u4e00\u4f75\u8655\u7406\u4e4b\u76f8\u9130\u8072\u5b78\u983b\u7387\u7684\u9ede\u6578\u3002\u5728\u539f\u59cb\u7684 MSE-NMF \u6a21\u5f0f\u4e2d\uff0c = 1\uff0c\u6211\u5011\u5c07\u5728\u4e4b\u5f8c\u7684\u5be6\u9a57\u88e1\uff0c\u5c07\u6b64\u503c\u8b8a\u5316\u70ba 2, 3 \u8207 4\u3002\u5f88\u660e\u986f\u7684\uff0c\u7576 \u8d8a \u5927\uff0c\u4ee3\u8868\u8d8a\u591a\u76f8\u9130\u8072\u5b78\u983b\u7387\u88ab\u8996\u70ba\u4e00\u9ad4\u4f86\u52a0\u4ee5\u8655\u7406\uff0c\u610f\u5373\u983b\u7387\u89e3\u6790\u5ea6\u8b8a\u4f4e\uff0c\u56e0\u6b64\uff0c\u96d6\u7136 \u589e\u52a0 \u53ef\u4f7f NMF \u8a13\u7df4\u8cc7\u6599\u8b8a\u591a\uff0c\u4f46\u53ef\u80fd\u56e0\u70ba\u72a7\u7272\u983b\u7387\u89e3\u6790\u5ea6\u800c\u964d\u4f4e\u6574\u9ad4\u5f37\u5316\u7684\u6548\u679c\u3002 (\u4e00)\u6240\u4f7f\u7528\u7684\u8a9e\u97f3\u8cc7\u6599\uff1a \u672c\u8ad6\u6587\u7528\u4ee5\u8a55\u4f30\u5f37\u5316\u65b9\u6cd5\u4e4b\u5be6\u9a57\u6240\u4f7f\u7528\u7684\u8a9e\u97f3\u8cc7\u6599\uff0c\u662f\u53d6\u81ea\u6b50\u6d32 \u96fb \u4fe1 \u6a19 \u6e96 \u5354 \u6703 (European-Telecommunications Standards Institute, ETSI) \u6240\u767c\u884c\u7684 AURORA 2.0 \u8a9e\u97f3\u8cc7 SNR 0 dB 5 dB 10 dB 15 dB \u8f03\u4f4e\u8072\u5b78\u983b\u7387\u89e3\u6790\u5ea6\u7684\u8a2d\u5b9a(\u5373 \u5927\u65bc 1)\u6642\uff0cNMF-MSE \u5f37\u5316\u7684\u6548\u679c\u901a\u5e38\u6bd4\u539f\u8a2d\u5b9a(\u539f 20 dB Baseline (\u672a\u8655\u7406) 1.8266 2.1700 2.5367 2.8015 3.1490 NMF-MSE 2.2215 2.3892 2.7011 2.9321 \u78ba\u7684\u57fa\u5e95\u77e9\u9663\uff0c\u9032\u800c\u4f7f NMF-MSE \u7684\u6548\u679c\u66f4\u660e\u986f\u3002\u6b64\u5916\uff0c\u63a1\u7528\u5927\u65bc 1 \u7684 \u503c(\u5373\u8f03\u4f4e 2.8 3.2008 \u983b\u7387\u89e3\u6790\u5ea6\uff0c\u5373 \u7b49\u65bc 1)\u4f86\u7684\u597d\uff0c\u6b64\u5927\u81f4\u547c\u61c9\u4e86\u6211\u5011\u524d\u9762\u6240\u63d0\uff0c\u76f8\u9130\u8072\u5b78\u983b\u7387\u4e4b\u8abf\u8b8a \u983b\u8b5c\u7684\u7279\u6027\u76f8\u8fd1\uff0c\u56e0\u6b64\u4e00\u4f75\u8655\u7406\u662f\u53ef\u884c\u7684\uff0c\u4e14\u56e0\u70ba\u589e\u52a0\u6a23\u672c\u6578\uff0c\u4f7f NMF \u6cd5\u80fd\u5f97\u5230\u8f03\u7cbe PESQ\u5e73\u5747\u503c \u53c3\u8003\u6587\u737b</td></tr><tr><td>\u6599\u5eab[16]\uff0c\u767c\u97f3\u8a9e\u8005\u70ba\u7f8e\u570b\u6210\u5e74\u7537\u5973\uff0c\u5167\u5bb9\u70ba\u4e00\u7cfb\u5217\u9023\u7e8c\u7684\u82f1\u6587\u6578\u5b57\u5b57\u53e5\uff0c\u6b64\u8a9e\u6599\u5eab\u8d77 \u521d\u662f\u7528\u5728\u96dc\u8a0a\u74b0\u5883\u4e4b\u8a9e\u97f3\u8fa8\u8b58\u8a55\u4f30\u4e0a\uff0c\u8a13\u7df4\u8072\u5b78\u6a21\u578b\u7684\u74b0\u5883\u6709\u5169\u7a2e:\u4e7e\u6de8\u8a13\u7df4 (clean-2.7 \u679c\uff0c\u6b64\u70ba\u4e00\u986f\u8457\u7684\u512a\u9ede\u3002 (\u4e8c)\u4e0d\u540c\u6bd4\u4f8b\u4e4b\u4f4e\u8abf\u8b8a\u983b\u5e36\u4e4b NMF-MSE \u4e4b\u5be6\u9a57\u7d50\u679c \u7684\u983b\u7387\u89e3\u6790\u5ea6)\u53ef\u4f7f NMF-MSE \u904b\u7b97\u8907\u96dc\u5ea6\u964d\u4f4e\u3001\u537b\u4ecd\u53ef\u4ee5\u5e36\u4f86\u76f8\u4f3c\u751a\u81f3\u66f4\u597d\u7684\u5f37\u5316\u6548 2.75</td></tr><tr><td>condition training) \u74b0\u5883\uff1a\u4f7f\u7528\u7684\u8a13\u7df4\u8a9e\u6599\u662f\u4e0d\u5305\u542b\u96dc\u8a0a\u7684\u4e7e\u6de8\u8a9e\u97f3 (clean speech);\u53e6\u4e00 \u7a2e\u5247\u662f\u8907\u5408\u8a13\u7df4 (multi-condition training) \u74b0\u5883\uff1a\u8a13\u7df4\u8a9e\u6599\u662f\u4e0d\u540c\u7a0b\u5ea6\u4e4b\u8a0a\u96dc\u6bd4 (Signal-to-noise ratio, SNR) \u7684\u96dc\u8a0a\u8a9e\u97f3\u3002\u7531\u65bc\u6211\u5011\u5728\u672c\u8ad6\u6587\u4e2d\uff0c\u91cd\u8996\u7684\u662f\u8a9e\u97f3\u5f37\u5316\u6280\u8853\u7684\u767c\u5c55 \u8207\u8a55\u4f30\uff0c\u4e26\u672a\u8a13\u7df4\u8072\u5b78\u6a21\u578b\u7528\u4ee5\u8a9e\u97f3\u8fa8\u8b58\uff0c\u56e0\u6b64\u8207\u4e0a\u8ff0\u5169\u7a2e\u8a13\u7df4\u74b0\u5883\u7121\u95dc\u3002 \u5728\u672c\u7bc0\uff0c\u6211\u5011\u5c07\u5448\u73fe\u4e0d\u540c\u6bd4\u4f8b\u4e4b\u4f4e\u8abf\u8b8a\u983b\u5e36\u4e4b NMF-MSE \u66f4\u65b0\u5f8c\u4e4b PESQ \u503c\uff0c\u5982\u524d\u6240 \u8ff0\uff0c\u53c3\u6578 to-full ratio)\u3002\u5728\u672c\u7bc0\u5be6\u9a57\u4e2d\u6b64\u53c3\u6578 \u5206\u5225\u88ab\u8a2d\u5b9a\u70ba 1\u30010.75\u30010.50 \u8207 0.25\uff0c\u96a8\u8457 \u4ee3\u8868\u6bcf\u500b\u55ae\u4e00\u8072\u5b78\u983b\u7387\u4e4b\u8abf\u8b8a\u983b\u8b5c\u88ab\u7368\u7acb\u5f37\u5316\u3002 SNR 0 dB 5 dB 10 dB 15 dB 2.5 20 dB \u7684\u905e\u6e1b\uff0c\u6240\u66f4\u65b0\u8655\u7406\u7684\u983b\u5bec\u4e5f\u8ddf\u8457\u8b8a\u5c0f\u3002\u540c\u6642\uff0cNMF-MSE \u6cd5\u4e2d\u7684 \u53c3\u6578\u503c\u56fa\u5b9a\u70ba 1\uff0c \u96dc\u8a0a\u70ba siren\uff0c\u56fa\u5b9a 2.55 = 1\u3001\u8b8a\u5316 \u503c\u4e4b NMF-MSE \u6cd5\u4e4b PESQ \u7d50\u679c \u503c \u8868\u4e09\uff1a\u8b66\u5831\u96dc\u8a0a\u74b0\u5883\u4e0b\uff0c\u57fa\u790e\u5be6\u9a57\u53ca\u4e0d\u540c \u503c\u4e4b NMF-MSE \u6cd5\u6240\u5f97\u7684 PESQ \u503c 2.6 \u4ee3\u8868\u4e86\u88ab NMF-MSE \u66f4\u65b0\u4e4b\u4f4e\u983b\u5e36\u983b\u5bec\u76f8\u5c0d\u65bc\u5168\u983b\u5e36\u983b\u5bec\u4e4b\u6bd4\u4f8b (low-2.65</td></tr><tr><td>(\u5176\u4e2d\uff0c\"\u2215\"\u8207\"\u00d7\" \u5206\u5225\u70ba\u77e9\u9663\u55ae\u4e00\u5143\u7d20\u9ede\u7684\u9664\u6cd5\u8207\u4e58\u6cd5\u904b\u7b97) \uff0c\u6b64\u5f37\u5ea6\u914d\u5408 \u7684\u539f\u59cb\u76f8 \u4f4d\u6210\u5206\uff0c\u53ef\u5f97\u5230\u66f4\u65b0\u904e\u5f8c\u7684\u8abf\u8b8a\u983b\u8b5c\uff0c\u518d\u7d93\u7531\u53cd\u5085\u7acb\u8449\u8f49\u63db (Inverse Fourier transform) \u5373\u53ef\u7372\u5f97\u5f37\u5316\u7684 (\u55ae\u8072\u5b78\u983b\u7387) \u8072\u5b78\u983b\u8b5c\u5f37\u5ea6\u6642\u5e8f\u5217\uff0c\u6700\u5f8c\uff0c\u5f37\u5316\u5f8c\u4e4b\u6240\u6709\u8072\u5b78\u983b \u7387\u4e4b\u8072\u5b78\u983b\u8b5c\u5f37\u5ea6\u914d\u5408\u539f\u59cb\u76f8\u4f4d\uff0c\u5f97\u5230\u65b0\u7684\u6642\u983b\u5716\u5f8c\uff0c\u5404\u97f3\u6846\u7d93\u7531\u53cd\u77ed\u6642\u9593\u5085\u7acb\u8449\u8f49 \u63db (Inverse STFT) \u5c31\u53ef\u5f97\u5230\u5f37\u5316\u5f8c\u7684\u8a9e\u97f3\u8a0a\u865f\u3002 (\u4e09)NMF-MSE\u6cd5\u7684\u5169\u7a2e\u8b8a\u5f62 \u5716\u4e00\uff1a\u53c3\u6578 \u4e4b\u793a\u610f\u5716 \u5728\u8a55\u4f30\u6211\u5011\u6240\u63d0\u51fa\u7684\u57fa\u65bc NMF \u4e4b\u8abf\u8b8a\u8072\u5b78\u983b\u8b5c\u5f37\u5316\u6cd5\u4e2d\uff0c\u6211\u5011\u4f7f\u7528\u4e86\u4f86\u81ea Aurora-dB\u300115 dB\u300110 dB\u30015 dB\u30010 dB \u4e94\u7a2e\u3002\u7d14\u8b66\u5831\u5668\u96dc\u8a0a\u5247\u7528\u4ee5\u8a13\u7df4\u4e7e\u6de8\u8a9e\u97f3 NMF \u57fa\u5e95\u77e9\u9663 \uff0c\u6211\u5011\u5c07\u5169\u57fa\u5e95 \u77e9\u9663 ( \u8207 ) \u7684\u884c\u6578\u56fa\u5b9a\u70ba 20\u3002 (\u4e8c)\u6240\u4f7f\u7528\u7684\u8a9e\u97f3\u54c1\u8cea\u4f30\u6e2c\u65b9\u6cd5 \u5225\u3002\u4e3b\u89c0\u985e\u5225\u7684\u8a9e\u97f3\u54c1\u8cea\u8a55\u4f30\u5176\u4e2d\u8457\u540d\u7684\u65b9\u6cd5\u70ba\u5e73\u5747\u610f\u898b\u5f97\u5206 (mean opinion score, Evaluation of Speech Quality (PESQ)\u6cd5[18]\u3002\u5728\u672c\u8ad6\u6587\u4e2d\uff0c\u5c07\u63a1\u7528\u5ba2\u89c0\u985e\u5225\u7684 PESQ \u6cd5\u4f86 SNR \u5f88\u4f4e(\u5982 0 dB \u6216 5 dB \u6642) \uff0c\u8655\u7406\u5168\u983b\u5e36\u7684 NMF-MSE \u4ecd\u662f\u8f03\u9069\u7576\u7684\u9078\u64c7\u3002 \u4e4b PESQ \u5e73\u5747\u503c\u8f49\u6210\u5716\u4e00\u4e4b\u9577\u689d\u5716\u4ee5\u65b9\u4fbf\u6bd4\u8f03\u3002 \u91cf\u5316\u6578\u64da\u7684\u65b9\u5f0f\u4f86\u8868\u793a\u8a9e\u97f3\u8a0a\u865f\u53d7\u5230\u96dc\u8a0a\u640d\u58de\u7684\u7a0b\u5ea6\u9ad8\u4f4e\uff0c\u8457\u540d\u7684\u65b9\u6cd5\u5982\uff1aPerceptual 3. \u5728\u9ad8 SNR \u6642\uff0c\u8655\u7406\u4f4e\u81f31 4 \u2044 \u7684\u8abf\u8b8a\u983b\u8b5c\u5bec\u5ea6\u4e5f\u80fd\u9054\u5230\u8fd1\u4f3c\u8655\u7406\u5168\u983b\u5e36\u7684\u6548\u679c\uff0c\u4f46\u7576 (MMSE-STSA)\uff0c\u5be6\u9a57\u7d50\u679c(PESQ \u503c)\u5982\u8868\u56db\u3002\u540c\u6642\uff0c\u6211\u5011\u4e5f\u5c07\u8868\u56db\u5404\u65b9\u6cd5\u8de8\u4e0d\u540c SNR MOS)\u6cd5[17]\u3002\u5ba2\u89c0\u985e\u5225\u7684\u8a9e\u97f3\u54c1\u8cea\u8a55\u4f30\u6cd5\uff0c\u662f\u900f\u904e\u96fb\u8166\u6f14\u7b97\u6cd5\u5206\u6790\u8a9e\u97f3\u8a0a\u865f\u6216\u983b\u8b5c\u3001\u7528 \u5225\u70ba\u983b\u8b5c\u6d88\u53bb\u6cd5(SS)\u3001\u97cb\u7d0d\u6ffe\u6ce2\u5668\u6cd5(WF)\u4ee5\u53ca\u6700\u5c0f\u5747\u65b9\u8aa4\u5dee\u4e4b\u77ed\u6642\u983b\u8b5c\u632f\u5e45\u4f30\u6e2c\u6cd5 \u70ba\u53ea\u8655\u7406 50%\u7684\u8abf\u8b8a\u983b\u8b5c) \u3002 \u5728\u672c\u7bc0\uff0c\u6211\u5011\u5c07\u5448\u73fe\u524d\u4e00\u7bc0\u6700\u4f73\u8a2d\u5b9a\u7684 NMF-MSE \u6cd5\u8207\u4e09\u7a2e\u8457\u540d\u8a9e\u97f3\u5f37\u5316\u6cd5\u505a\u6bd4\u8f03\uff0c\u5206 \u614b\u4e0b\uff0cPSEQ \u5927\u7d04\u53ea\u4e0b\u964d\u4e86 0.01 \u81f3 0.02\uff0c\u4f46\u904b\u7b97\u8907\u96dc\u5ea6\u537b\u53ef\u4ee5\u56e0\u6b64\u4e0b\u964d\u4e86\u4e00\u534a(\u56e0 \u76ee\u524d\u5df2\u6709\u7684\u8a9e\u97f3\u54c1\u8cea\u4f30\u6e2c\u65b9\u5f0f\u5206\u70ba\u5169\u985e\uff0c\u4e3b\u89c0 (subjective) \u985e\u5225\u53ca\u5ba2\u89c0(objective) \u985e Baseline 1.8266 2.1700 2.5367 2.8015 3.1490 2.45 \u8868\u4e8c\u5217\u51fa\u4e86\u53d7\u96dc\u8a0a\u5e72\u64fe\u4e4b\u8a9e\u97f3\u7d93\u7531 NMF-MSE \u5f37\u5316\u5f8c\u7684 PESQ \u503c\u3002\u5f9e\u9019\u8868\uff0c\u6211\u5011\u6709\u4ee5\u4e0b = 1 2.2215 2.3892 2.7011 2.9321 3.2008 2.4 \u7684\u89c0\u5bdf\uff1a 1. \u7576\u8655\u7406\u7684\u8abf\u8b8a\u983b\u5e36\u5bec\u8d8a\u5c0f(\u5373 \u8d8a\u5c0f)\u6642\uff0cNMF-MSE \u6240\u5c0d\u61c9\u7684 PSEQ \u63d0\u5347\u7a0b\u5ea6 \u4e00\u822c\u6703\u8d8a\u4f4e\uff0c\u552f\u4e00\u4f8b\u5916\u662f 15 dB \u7684\u8a0a\u96dc\u6bd4\u6642\uff0c = 0.75\u7684\u8a2d\u5b9a\u6bd4 = 1\u7684\u8a2d\u5b9a \u5f97\u5230\u66f4\u9ad8\u7684 PESQ \u503c\u3002 2. \u96d6\u7136\u964d\u4f4e \u76f8\u5c0d\u5e36\u4f86\u8f03\u4f4e\u7684\u8a9e\u97f3\u5f37\u5316\u6548\u679c\uff0c\u4f46\u5176\u5be6 PESQ \u8b8a\u5dee\u7684\u6548\u61c9\u4e26\u4e0d\u986f\u8457\uff0c \u7279\u5225\u662f\u7576\u8a0a\u96dc\u6bd4\u8f03\u9ad8\u7684\u60c5\u5f62\u4e0b\uff0c\u4f8b\u5982\u7576\u628a \u5f9e 1 \u964d\u70ba 0.5 \u6642\uff0cSNR \u70ba 10 dB \u7684\u72c0 NMF-MSE = 2 2.2445 2.3981 2.7020 2.9372 3.1991 = 3 2.2453 2.3969 2.7062 2.9356 3.2015 = 4 2.2322 2.3882 2.6960 2.9319 3.1870 (\u56db)NMF-MSE \u6cd5\u8207\u4e09\u7a2e\u8a9e\u97f3\u5f37\u5316\u6cd5\u4e4b\u6548\u80fd\u6bd4\u8f03 \u5716\u4e00\u3001 \u57fa\u790e\u5be6\u9a57\u3001NMF-MSE(\u5176 \u8a2d\u70ba 3\u3001 \u8a2d\u70ba 1)\u8207\u4e09\u7a2e\u8a9e\u97f3\u5f37\u5316\u6cd5(SS, WF, 2.35 Baseline NMF-MSE SS WF MMSE-STSA</td></tr><tr><td>\uf06c \u4f4e\u8abf\u8b8a\u983b\u5e36\u4e4bNMF-MSE \u5728\u8af8\u591a\u6587\u737b\u4e2d\uff0c\u7686\u63d0\u5230\u8a9e\u97f3\u4e3b\u8981\u7684\u8cc7\u8a0a\u662f\u96c6\u4e2d\u5728\u4f4e\u983b\u7387\u7684\u8abf\u8b8a\u983b\u8b5c\u4e2d\uff0c\u4f8b\u5982\u7576\u97f3\u6846 \u8a55\u4f30\u8a9e\u97f3\u54c1\u8cea\uff0cPESQ \u6240\u5f97\u7684\u5206\u6578\u7bc4\u570d\u70ba 1.0~4.5 \u4e4b\u9593\uff0c\u4e3b\u8981\u5c0d\u61c9\u5230\u4e3b\u89c0\u8a55\u5206\u7684 MOS \u6cd5\uff0c \u8868\u4e8c\uff1a\u8b66\u5831\u96dc\u8a0a\u74b0\u5883\u4e0b\uff0c\u57fa\u790e\u5be6\u9a57\u53ca\u4e0d\u540c \u503c\u4e4b NMF-MSE \u6cd5\u6240\u5f97\u7684 PESQ \u503c \u8d8a\u9ad8\u5206\u4ee3\u8868\u8a9e\u97f3\u54c1\u8cea\u8d8a\u4f73\u3002\u4e00\u822c\u4f86\u8aaa\u5206\u6578\u8d85\u904e 4 \u5206\u4ee3\u8868\u8457\u807d\u8005\u89ba\u5f97\u6eff\u610f\u6216\u975e\u5e38\u6eff\u610f\u3002 \u96dc\u8a0a\u70ba siren\uff0c\u56fa\u5b9a = 1\u3001\u8b8a\u5316 \u503c\u4e4b NMF-MSE \u6cd5\u4e4b PESQ \u7d50\u679c \u53d6\u6a23\u7387\u70ba 100 Hz \u6642\uff0c\u96d6\u7136\u6574\u9ad4\u8abf\u8b8a\u983b\u8b5c\u7684\u983b\u7387\u7bc4\u570d\u70ba 0 \u5230 50 Hz\uff0c\u5c0d\u8a9e\u97f3\u8fa8\u8b58\u4e4b\u95dc\u9375\u983b SNR 0 dB 5 dB 10 dB 15 dB 20 dB</td></tr></table>", | |
| "num": null, | |
| "type_str": "table", | |
| "html": null, | |
| "text": "\u76f8\u6bd4\uff0c\u63d0\u5347\u4e86 0.4187\u3002\u800c\u7576\u96dc\u8a0a\u7a0b\u5ea6\u5f88\u4f4e(SNR \u70ba 20dB \u6642) \uff0c WF \u6cd5\u6240\u5c0d\u61c9\u7684 PESQ \u70ba\u6700\u4f73\uff0c\u800c NMF-MSE \u548c\u57fa\u790e\u5be6\u9a57\u76f8\u6bd4\u4ecd\u9ad8\u4e86 0.0525\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\u6b64\u6642 NMF-MSE \u7684 MMSE-STSA)\u5404\u65b9\u6cd5\u6240\u5f97\u4e4b\u8de8\u4e0d\u540c SNR \u7684 PESQ \u5e73\u5747\u503c \u4e94\u3001\u7d50\u8ad6 \u5728\u672c\u8ad6\u6587\u4e2d\uff0c\u6211\u5011\u63d0\u51fa\u4e86\u57fa\u65bc\u975e\u8ca0\u77e9\u9663\u5206\u89e3\u6cd5 (NMF) \u5728\u8a9e\u97f3\u5f37\u5316\u4e0a\u7684\u61c9\u7528\u6280\u8853\uff0c \u4f7f\u7528 NMF \u6cd5\u5c0d\u8a9e\u97f3\u6642\u983b\u5716\u4e4b\u8abf\u8b8a\u983b\u8b5c\u7684\u5f37\u5ea6\u505a\u66f4\u65b0\uff0c\u7c21\u7a31\u70ba NMF-MSE (NMF-based modulation spectrum enhancement)\u3002\u5728 NMF-MSE \u6cd5\u4e2d\uff0c\u85c9\u7531\u5206\u958b\u8a13\u7df4\u8a9e\u53e5\u4e2d\u4e7e\u6de8\u8a9e\u97f3 \u53ca\u96dc\u8a0a\u7684\u8abf\u8b8a\u983b\u8b5c\u5f37\u5ea6\u7684 NMF \u57fa\u5e95 (basis)\uff0c\u5c07\u6240\u5f97\u57fa\u5e95\u7528\u4ee5\u5206\u89e3\u6e2c\u8a66\u8a9e\u97f3\u6642\u983b\u5716\u7684\u8abf \u8b8a\u983b\u8b5c\u5f37\u5ea6\uff0c\u518d\u5c07\u5f37\u5316\u5f8c\u7684\u8abf\u8b8a\u983b\u8b5c\u900f\u904e\u53cd\u5085\u7acb\u8449\u8f49\u63db (inverse Fourier transform) \u5f97\u5230" | |
| } | |
| } | |
| } | |
| } |