| { | |
| "paper_id": "2019", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T14:54:47.773978Z" | |
| }, | |
| "title": "Speech Enhancement for TTS Speech Corpora by using Voice Conversion Technologies", | |
| "authors": [ | |
| { | |
| "first": "Yan-Ting", | |
| "middle": [], | |
| "last": "\u6797\u884d\u5ef7", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "National Taipei University", | |
| "location": {} | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Lin", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "National Taipei University", | |
| "location": {} | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Chen-Yu", | |
| "middle": [], | |
| "last": "\u6c5f\u632f\u5b87", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "National Taipei University", | |
| "location": {} | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Chiang", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "National Taipei University", | |
| "location": {} | |
| }, | |
| "email": "[email protected]" | |
| } | |
| ], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "", | |
| "pdf_parse": { | |
| "paper_id": "2019", | |
| "_pdf_hash": "", | |
| "abstract": [], | |
| "body_text": [ | |
| { | |
| "text": "\u9ad8\u65af\u6df7\u5408\u6a21\u578b GMM\uff0c\u53ef\u4ee5\u5f88\u597d\u7684\u8fd1\u4f3c\u4efb\u610f\u7684\u6a5f\u7387\u5206\u5e03\uff0c\u5728\u9019\u908a\u7528\u4f86\u63cf\u8ff0\u8072\u5b78\u53c3\u6578 \u7684\u6a5f\u7387\u5206\u5e03\uff0c\u7528\u65bc\u8a9e\u8005\u8f49\u63db\u3002\u4ee4\u4f86\u6e90\u7684 MCC \u70ba \u3001\u76ee\u6a19\u70ba \uff0c \u70ba\u97f3\u6846\u6578\uff0c\u70ba\u4e86\u8003\u616e\u97f3 \u6846 \u4e4b\u9593\u7684\u95dc\u806f\u6027\uff0c \u9700\u8981\u4e00\u500b\u6709\u524d\u5f8c\u97f3\u6846\u8cc7\u8a0a\u7684delta = \u2206 = \u22120.5 * \u22121 +0.5 +1 \uff0c\u4ee4 = [ \u2206 ]\u548c = [ \u2206 ]\uff0c\u800c = [ ]\uff0c\u5047\u8a2d \u53ef\u4ee5\u7531 \u9ad8\u65af\u6df7\u5408\u6a21\u578b\u8868\u793a\u6210 ( | ( ) )\uff0c\u03bb (Z) \u70ba\u6a5f\u7387\u53c3\u6578\u3002\u63a5\u4e0b\u4f86\u5c07 ( | ( ) )\u7528\u8c9d\u5f0f\u5b9a\u7406\u5c55\u958b\uff0c ( | , ( ) )\u5c31\u662f\u5c07\u4f86\u6e90\u8f49\u5230\u76ee\u6a19\u7684\u6620\u5c04\u51fd\u6578\u3002 ( | ( ) ) = ( , | ( ) ) = ( | , ( ) ) ( | ( ) ) (3.1) \u5728\u4f30\u8a08\u76ee\u6a19\u0302\u6642\uff0c\u6211\u5011\u6703\u5e0c\u671b likelihood \u6700\u5927\uff0c\u4e5f\u5c31\u662f ( | , ( ) )\u6700\u5927\uff0c\u5beb\u6210 \u0302= ( | , ( ) ) (3.2) \u76ee\u6a19\u51fd\u6578 ( ,\u0302)\u53ef\u5beb\u6210\u5f0f 3.11\uff0c\u70ba\u4e86\u4f7f\u76ee\u6a19\u51fd\u6578\u6700\u5927\uff0c\u5c07 ( ,\u0302)\u5c0dy\u5fae\u5206\u7b97\u51fa\u0302\u6574\u500b\u904e \u7a0b\u7a31\u70ba maximum likelihood parameter generation(MLPG)\u3002 ( ,\u0302) = \u2211 ( | , , ( ) ) (\u0302, | , ( ) ) (3.3)", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "\u8a9e\u6599\uff0c\u5c31\u80fd\u6e1b\u5c11\u8a9e\u6599\u7522\u751f\u7684\u554f\u984c\u4e26\u7e7c\u7e8c\u4f7f\u7528\u9019\u4e9b\u8a9e\u6599\u3002\u672c\u7814\u7a76\u6240\u4f7f\u7528\u7684\u8a9e\u97f3\u8cc7\u6599\u5eab\u662f\u7531 \u4e00\u4f4d\u5c08\u696d\u5973\u6027\u64ad\u97f3\u54e1(\u7e2e\u5beb\u6210\u8a9e\u8005", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "https://drive.google.com/drive/folders/1YsCJNmhw6RFWzfI9DMyiX8ciNBTaAjwu?usp=sharing", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| } | |
| ], | |
| "back_matter": [], | |
| "bib_entries": { | |
| "BIBREF0": { | |
| "ref_id": "b0", | |
| "title": "Continuous Probabilistic Transform for Voice Conversion", | |
| "authors": [ | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Stylianou", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "O", | |
| "middle": [], | |
| "last": "Capp\u00b4e", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "E", | |
| "middle": [], | |
| "last": "Moulines", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1998, | |
| "venue": "IEEE Trans. on Audio, Speech and Language Processing", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Y. Stylianou, O. Capp\u00b4e, and E. Moulines, \"Continuous Probabilistic Transform for Voice Conversion,\" in IEEE Trans. on Audio, Speech and Language Processing, 1998", | |
| "links": null | |
| }, | |
| "BIBREF1": { | |
| "ref_id": "b1", | |
| "title": "Voice Conversion Based on Maximum-Likelihood Estimation of Spectral Parameter Trajectory", | |
| "authors": [ | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Toda", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [ | |
| "W" | |
| ], | |
| "last": "Black", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Tokuda", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2007, | |
| "venue": "Speech and Language Processing", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "T. Toda, A. W. Black, and K. Tokuda, \"Voice Conversion Based on Maximum-Likelihood Estimation of Spectral Parameter Trajectory,\" in IEEE Trans. on Audio, Speech and Language Processing, 2007", | |
| "links": null | |
| }, | |
| "BIBREF2": { | |
| "ref_id": "b2", | |
| "title": "Spectral Mapping Using Artificial Neural Networks for Voice Conversion", | |
| "authors": [ | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Desai", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [ | |
| "W" | |
| ], | |
| "last": "Black", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "B", | |
| "middle": [], | |
| "last": "Yegnanarayana", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Prahallad", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2010, | |
| "venue": "Speech and Language Processing", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "S. Desai, A. W. Black, B. Yegnanarayana, and K. Prahallad, \"Spectral Mapping Using Artificial Neural Networks for Voice Conversion,\" in IEEE Trans. on Audio, Speech and Language Processing, 2010", | |
| "links": null | |
| }, | |
| "BIBREF3": { | |
| "ref_id": "b3", | |
| "title": "Voice Conversion Using Deep Neural Networks With Layer-Wise Generative Training", | |
| "authors": [ | |
| { | |
| "first": "L.-H", | |
| "middle": [], | |
| "last": "Chen", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Z.-H", | |
| "middle": [], | |
| "last": "Ling", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "L.-J", | |
| "middle": [], | |
| "last": "Liu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "L.-R", | |
| "middle": [], | |
| "last": "Dai", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2014, | |
| "venue": "IEEE Trans. on Audio, Speech and Language Processing", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "L.-H. Chen, Z.-H. Ling, L.-J. Liu, and L.-R. Dai, \"Voice Conversion Using Deep Neural Networks With Layer-Wise Generative Training,\" in IEEE Trans. on Audio, Speech and Language Processing, 2014", | |
| "links": null | |
| }, | |
| "BIBREF4": { | |
| "ref_id": "b4", | |
| "title": "Voice Conversion Using Deep Bidirectional Long Short-Term Memory Based Recurrent Neural Networks", | |
| "authors": [ | |
| { | |
| "first": "L", | |
| "middle": [], | |
| "last": "Sun", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Kang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Li", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Meng", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2015, | |
| "venue": "ICASSP", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "L. Sun, S. Kang, K. Li and H. Meng, \"Voice Conversion Using Deep Bidirectional Long Short-Term Memory Based Recurrent Neural Networks,\" in ICASSP, Apr. 2015.", | |
| "links": null | |
| }, | |
| "BIBREF5": { | |
| "ref_id": "b5", | |
| "title": "F0 Transformation Techniques for Statistical Voice Conversion with Directwaveform Modification with Spectral Differential", | |
| "authors": [ | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Kobayashi", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Toda", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Nakamura", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2016, | |
| "venue": "IEEE Spoken Language Technology Workshop", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "K. Kobayashi, T. Toda, S. Nakamura, \"F0 Transformation Techniques for Statistical Voice Conversion with Directwaveform Modification with Spectral Differential,\" in IEEE Spoken Language Technology Workshop, Dec. 2016.", | |
| "links": null | |
| }, | |
| "BIBREF6": { | |
| "ref_id": "b6", | |
| "title": "Phonetic Posteriorgrams for Many-to-One Voice Conversion without Parallel Data Training", | |
| "authors": [ | |
| { | |
| "first": "L", | |
| "middle": [], | |
| "last": "Sun", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Li", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Wang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Kang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Meng", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2016, | |
| "venue": "ICME", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "L. Sun, K. Li, H. Wang, S. Kang and H. Meng, \"Phonetic Posteriorgrams for Many-to-One Voice Conversion without Parallel Data Training,\" in ICME, July. 2016.", | |
| "links": null | |
| }, | |
| "BIBREF7": { | |
| "ref_id": "b7", | |
| "title": "WORLD: A Vocoder-Based High-Quality Speech Synthesis System for Real-Time Applications", | |
| "authors": [ | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Morise", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "F", | |
| "middle": [], | |
| "last": "Yokomori", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Ozawa", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2016, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "M. MORISE, F. YOKOMORI, K. OZAWA, \"WORLD: A Vocoder-Based High-Quality Speech Synthesis System for Real-Time Applications,\" in IEICE, 2016", | |
| "links": null | |
| } | |
| }, | |
| "ref_entries": { | |
| "TABREF0": { | |
| "content": "<table><tr><td>\u8a9e\u8005\u8f49\u63db\u4f5c\u6cd5\u6d41\u7a0b\u5982\u5716\u4e00\uff0c\u8f38\u5165\u8a0a\u865f\u7d93\u7531\u5206\u6790\u6210\u8072\u5b78\u53c3\u6578\uff0c\u5728\u8a13\u7df4\u8cc7\u6599\u7684\u90e8\u5206\uff0c WORLD [8]\u8072\u78bc\u5668\u662f\u8fd1\u5e74\u88ab\u8a8d\u70ba\u76ee\u524d\u97f3\u8cea\u6700\u597d\u7684\u8072\u78bc\u5668\uff0c\u5b83\u5c07\u8a0a\u865f\u5206\u6210\u97f3\u8abf(pitch)\u3001\u983b</td></tr><tr><td>\u8a13\u7df4\u7684\u6a21\u578b\u90fd\u662f\u7528\u4f86\u8f49\u63db\u8072\u5b78\u53c3\u6578\uff0c\u7d71\u8a08\u5f0f\u8f49\u63db(statistical conversion)\u6240\u7528\u7684\u6a21\u578b\u6709 \u8b5c\u5305\u7d61\u548c\u975e\u9031\u671f\u6027(aperiodicity)\u4e09\u500b\u90e8\u5206\uff0cWORLD \u5c07\u975e\u9031\u671f\u6027\u5f9e\u983b\u8b5c\u5305\u7d61\u5206\u51fa\u4f86\u4e4b\u5f8c\uff0c</td></tr><tr><td>GMM \u7684\u6a5f\u7387\u6a21\u578b\u4e5f\u6709\u795e\u7d93\u7db2\u8def(neuron network, NN)\u7684\u6a5f\u5668\u5b78\u7fd2\u6a21\u578b\uff0c\u90fd\u662f\u5728\u5c07\u8f38\u5165 \u7684\u53c3\u6578\u8f49\u63db\u6210\u63a5\u8fd1\u76ee\u6a19\u7684\u53c3\u6578\uff0c\u5c07\u8072\u5b78\u53c3\u6578\u8f49\u63db\u5b8c\u4e4b\u5f8c\uff0c\u9032\u5165\u5408\u6210\u5668\u5408\u6210\u51fa\u8072\u97f3\uff0c\u9019 \u505a\u983b\u8b5c\u5305\u7d61\u7684\u8f49\u63db\u6703\u66f4\u597d\uff0c\u6574\u9ad4\u97f3\u8cea\u6703\u8f03\u7a69\u5b9a\u3002</td></tr><tr><td>\u5c31\u662f\u8a9e\u8005\u8f49\u63db\u7684\u904b\u4f5c\u6d41\u7a0b\u3002 (\u4e8c)\u6620\u5c04\u51fd\u6578</td></tr><tr><td>\u4e8c\u3001\u8a9e\u97f3\u8cc7\u6599\u5eab</td></tr><tr><td>1\u3001Gaussian Mixture Model (GMM)</td></tr><tr><td>\u53e6\u5916\uff0c\u672c\u7814\u7a76\u4e5f\u5617\u8a66\u8f49\u63db\u53e6\u4e00\u500b\u8a9e\u8005 Tao \u6240\u9304\u88fd\u7684\u4e2d\u82f1\u593e\u96dc\u7684\u8a9e\u6599\uff0c\u8a72\u8a9e\u6599\u6709\u4e00\u6a23\u7684\u97f3 (i). Treebank-Tao-SR-Corpus\uff1a\u662f\u7531\u4e00\u4f4d\u5c08\u696d\u5973\u6027\u64ad\u97f3\u54e1(\u7e2e\u5beb\u6210\u8a9e\u8005 Tao)\u8b80\u7a3f\u9304\u88fd\u4e4b 4</td></tr><tr><td>\u8cea\u554f\u984c\uff0c\u5617\u8a66\u80fd\u5426\u4e5f\u8f49\u63db\u975e\u540c\u4e00\u8a9e\u8a00\u7684\u60c5\u5f62\u3002 \u7a2e\u8a9e\u901f\u5e73\u884c\u8a9e\u6599\u5eab\uff0c\u7e3d\u8a08 1,478 \u500b\u97f3\u6a94\uff0c\u5171\u6709 203,746 \u500b\u97f3\u7bc0\uff0c\u8a9e\u901f\u5206\u70ba\uff1a\u5feb\u3001\u6b63\u5e38\u3001</td></tr><tr><td>\u4e2d\u3001\u4ee5\u53ca\u6162\uff0c\u5e73\u5747\u97f3\u7bc0\u9577\u5ea6\u5206\u5225\u70ba 0.181 \u79d2\u30010.198 \u79d2\u30010.244 \u79d2\u53ca 0.264 \u79d2\uff0c\u97f3\u6a94\u5747</td></tr><tr><td>(\u4e8c)\u8a9e\u8005\u8f49\u63db\u76f8\u95dc\u7814\u7a76\u6587\u737b\u63a2\u8a0e \u70ba 20kHz \u7684\u53d6\u6a23\u7387\u53ca 16-bit \u4e4b PCM \u683c\u5f0f\uff0c\u4e3b\u8981\u5167\u5bb9\u5927\u591a\u6458\u9304\u81ea\u65b0\u805e\u3001\u7db2\u8def\u6587\u7ae0\u3002</td></tr><tr><td>\u8a9e\u8005\u8f49\u63db\u53ef\u4ee5\u5206\u70ba\u5e73\u884c\u8a9e\u6599\u7684\u8f49\u63db\u548c\u975e\u5e73\u884c\u8a9e\u6599\u7684\u8f49\u63db\u5169\u5927\u5340\u584a\u8a0e\u8ad6\uff0c\u5e73\u884c\u8a9e\u6599\u7684 (ii). English-Tao-CEMix-Spell-Corpus\uff1a\u662f\u7531\u8a9e\u8005 Tao \u6240\u9304\u88fd\u800c\u6210\u7684\u4e2d\u82f1\u593e\u96dc\u8a9e\u6599\uff0c\u4ee5\u4e2d</td></tr><tr><td>\u8f49\u63db\u6700\u5ee3\u70ba\u4eba\u77e5\u7684\u5c31\u662f\u7528 Gaussion mixture model(GMM)[1]\u4f86\u505a\u8f49\u63db\uff0c\u5c07\u4f86\u6e90\u97f3\u6846\u548c\u76ee \u6587\u70ba\u4e3b\u9ad4\u4e26\u7a7f\u63d2\u82f1\u6587\u5b57\u6bcd\u65bc\u4e2d\u6587\u8a9e\u53e5\u4e2d\uff0c\u5171 539 \u500b\u8a9e\u53e5\uff0c\u7e3d\u97f3\u7bc0\u6578\u70ba 13,540 \u500b\u97f3\u7bc0\uff0c</td></tr><tr><td>\u5305\u542b 11,688 \u500b\u4e2d\u6587\u97f3\u7bc0\u8207 1,872 \u500b\u82f1\u6587\u5b57\u6bcd\u3002\u97f3\u6a94\u70ba\u53d6\u6a23\u983b\u7387 20,000 \u8d6b\u8332(Hertz)\u53ca 16</td></tr><tr><td>\u4f4d\u5143\u6578\u4e4b PCM \u683c\u5f0f\uff0c\u5e73\u5747\u8a9e\u901f\u70ba\u4e00\u79d2 3.5 \u500b\u97f3\u7bc0\u3002</td></tr><tr><td>\u6cd5\uff0c\u7528\u6a5f\u7387\u4f86\u89e3\u9019\u4e00\u554f\u984c\u5230\u9019\u908a\u7b97\u662f\u6709\u4e00\u500b\u4e0d\u932f\u7684\u7d50\u679c\u3002</td></tr><tr><td>\u518d\u4f86\u5c31\u662f\u8fd1\u5e74\u6d41\u884c\u7684\u6a5f\u5668\u5b78\u7fd2\uff0c\u5f9e[3]\u958b\u59cb\u4e86\u7528\u985e\u795e\u7d93\u7db2\u8def(artificial neural networks, (iii). English-Tao-CEMix-Word-Corpus\uff1a\u662f\u7531\u8a9e\u8005 Tao \u6240\u9304\u88fd\u800c\u6210\u7684\u4e2d\u82f1\u593e\u96dc\u8a9e\u6599\uff0c\u4ee5\u4e2d</td></tr><tr><td>ANN)\u4f86\u8f49\u63db\u4f86\u6e90\u8a9e\u8005\u5230\u76ee\u6a19\u8a9e\u8005\uff0c\u5728[4]\u4e2d\u5be6\u9a57\u4e86\u5f88\u591a DNN \u7684\u8b8a\u7a2e\uff0c\u9084\u6709\u53ef\u4ee5\u7528\u4e00\u6bb5 \u6587\u70ba\u4e3b\u9ad4\u4e26\u7a7f\u63d2\u82f1\u6587\u8a5e(word)\u65bc\u4e2d\u6587\u8a9e\u53e5\u4e2d\uff0c\u5171 843 \u500b\u8a9e\u53e5\uff0c\u7e3d\u97f3\u7bc0\u6578\u70ba 18,103 \u500b\u97f3</td></tr><tr><td>\u8a0a\u865f\u4f5c\u70ba\u8f38\u5165\u7684 DBLSTM [5]\u3002\u9664\u4e86\u7528\u4e0a\u8ff0\u7684\u65b9\u5f0f\u505a\u8f49\u63db\uff0c\u9084\u6709\u7528 spectral differential[6] \u7bc0\uff0c\u5305\u542b 15,885 \u500b\u4e2d\u6587\u97f3\u7bc0\u8207 2,218 \u500b\u82f1\u6587\u97f3\u7bc0\u3002\u97f3\u6a94\u70ba\u53d6\u6a23\u983b\u7387 20,000 \u8d6b\u8332(Hertz)</td></tr><tr><td>\u7684\u4f5c\u6cd5\uff0c\u53bb\u5b78\u7fd2\u4f86\u6e90\u8a9e\u8005\u548c\u76ee\u6a19\u8a9e\u8005\u7684\u5dee\u7570\uff1b\u5f8c\u4f86\u51fa\u73fe\u4e86\u7528\u8a9e\u97f3\u8fa8\u8b58\u52a0\u8a9e\u97f3\u5408\u6210\u7684\u4f5c\u6cd5 \u53ca 16 \u4f4d\u5143\u6578\u4e4b PCM \u683c\u5f0f\uff0c\u5e73\u5747\u8a9e\u901f\u70ba\u4e00\u79d2 4.85 \u500b\u97f3\u7bc0\u3002</td></tr><tr><td>[7]\uff0c\u5148\u7528\u8a9e\u97f3\u8fa8\u8b58\u8fa8\u5225\u4f86\u6e90\u8a9e\u8005\u7684\u5167\u5bb9\uff0c\u518d\u8a13\u7df4\u76ee\u6a19\u8a9e\u8005\u7684\u8072\u5b78\u6a21\u578b\uff0c\u7528\u8072\u5b78\u6a21\u578b\u5c07\u5167 \u7531\u65bc\u672c\u5be6\u9a57\u6240\u9700\u5e73\u884c\u7684\u8a9e\u6599\uff0c\u5c07(i)\u8a9e\u6599\u5eab\u4e2d\u7684 4 \u500b\u8a9e\u901f\u7d93\u6587\u672c\u6bd4\u5c0d\u4e4b\u5f8c\uff0c\u6211\u5011\u7528\u5176</td></tr><tr><td>\u5bb9\u5408\u6210\u8072\u97f3\uff0c\u9019\u7a2e\u4f5c\u6cd5\u8b93\u975e\u5e73\u884c\u8a9e\u6599\u7684\u8a9e\u8005\u8f49\u63db\u6709\u66f4\u9032\u4e00\u6b65\u7684\u7a81\u7834\u3002 \u4e2d 1048 \u500b\u97f3\u6a94\u4f5c\u70ba\u5be6\u9a57\u7684\u8a9e\u6599\u4e14\u53d6\u6a23\u7387\u70ba 16kHz\u3002</td></tr><tr><td>(\u4e09)\u8a9e\u8005\u8f49\u63db\u67b6\u69cb \u4e09\u3001\u8a9e\u8005\u8f49\u63db\u6280\u8853\u7c21\u4ecb</td></tr><tr><td>(\u4e00)\u8072\u78bc\u5668</td></tr><tr><td>\u8072\u78bc\u5668\u5c31\u662f\u5c07\u8072\u97f3\u8a0a\u865f\u5206\u89e3\u6210\u6709\u610f\u7fa9\u7684\u53c3\u6578\uff0c\u50cf Mel-cepstral coefficient(MCC)\u5c31\u662f</td></tr><tr><td>\u5c07\u8072\u97f3\u8a0a\u865f\u7684\u983b\u8b5c\u53d6\u5c0d\u6578\u5f8c\u6309\u7167\u4eba\u8033\u5c0d\u983b\u7387\u7684\u807d\u89ba\u654f\u92b3\u505a\u7e2e\u653e\u518d\u505a\u5085\u7acb\u8449\u9006\u8f49\u63db\u5f97\u5230</td></tr><tr><td>\u7684\u4fc2\u6578\u3002\u9019\u500b\u4f9d\u64da\u4eba\u8033\u7279\u6027\u5c0d\u983b\u7387\u8ef8\u7e2e\u653e\u7684\u5c31\u662f\u6885\u723e\u523b\u5ea6(Mel-scale)\uff0c\u6bcf\u4e00\u500b\u523b\u5ea6\u90fd\u662f</td></tr><tr><td>\u4e00\u7dad\u7684 MCC\u3002Mel-log spectral approximation filter (MLSA filter)\u5c31\u662f\u5c07 MCC \u8f49\u56de\u983b\u8b5c\u5305</td></tr><tr><td>\u5716 \u4e00\uff1a\u8a9e\u8005\u8f49\u63db\u67b6\u69cb\u5716 \u7d61(spectral envelope)\uff0c\u5c31\u53ef\u4ee5\u628a\u983b\u8b5c\u5305\u7d61\u548c\u6fc0\u767c\u8a0a\u865f\u5377\u7a4d\u4f86\u5f97\u5230\u8072\u97f3\u8a0a\u865f\u3002\u53e6\u5916\uff0c</td></tr></table>", | |
| "text": "\u8b80\u7a3f\u9304\u88fd\u4e4b 4 \u7a2e\u8a9e\u901f\u5e73\u884c\u8a9e\u6599\u5eab\uff0c\u7e3d\u8a08 1,478 \u500b \u97f3\u6a94\uff0c\u8a9e\u901f\u5206\u70ba\uff1a\u5feb\u3001\u6b63\u5e38\u3001\u4e2d\u3001\u4ee5\u53ca\u6162\uff0c\u800c\u6709\u96dc\u8a0a\u7684\u8a9e\u6599\u70ba\u8a9e\u901f\u6b63\u5e38\uff0c\u672c\u7814\u7a76\u8a66\u8457\u7528 \u8a9e\u8005\u8f49\u63db\u7684\u65b9\u6cd5\uff0c\u5229\u7528\u540c\u4e00\u8a9e\u8005\u7684\u7279\u6027\u4f7f\u5f97\u8a9e\u8005\u8f49\u63db\u8b8a\u6210\u97f3\u8cea\u8f49\u63db\uff0c\u5c07 \u8a9e\u901f\u6b63\u5e38\u7684\u97f3\u8cea \u8f49\u6210\u5176\u4ed6\u8a9e\u901f\u7684\u97f3\u8cea\u3002\u73fe\u6709\u7684\u8f49\u63db\u6280\u8853\u5f9e\u9ad8\u65af\u6df7\u5408\u6a21\u578b(Gaussion mixture model,GMM) \u5230\u6df1\u5ea6\u985e\u795e\u7d93\u7db2\u8def(deep neuron network,DNN)\u7b49\u7b49\uff0c\u80fd\u591a\u65b9\u5617\u8a66\u627e\u51fa\u6700\u597d\u7684\u8f49\u63db\u6a21\u578b\u3002 \u6a19\u97f3\u6846\u7528\u52d5\u614b\u6642\u9593\u626d\u66f2(dynamic time warping,DTW)\u505a\u5c0d\u9f4a\u5f8c\uff0c\u4f86\u8a13\u7df4\u9ad8\u65af\u6df7\u5408\u6a21\u578b\u63cf \u8ff0\u4e0d\u540c\u7684\u767c\u97f3\uff0c\u9084\u6709[2]GMM \u52a0\u4e0a maximum likelihood parameter generation(MLPG)\u7684\u4f5c", | |
| "html": null, | |
| "type_str": "table", | |
| "num": null | |
| }, | |
| "TABREF1": { | |
| "content": "<table><tr><td>DBLSTM \u7684\u8868\u73fe\u6bd4\u8d77 GMM \u548c DNN \u90fd\u4f86\u7684\u597d\uff1b(2)\u85c9\u7531\u6587\u672c\u7684\u767c\u97f3\u6a19\u8a18\u548c\u767c\u97f3\u4f4d\u7f6e\u505a</td></tr><tr><td>\u70ba\u8f38\u5165\u4f86\u8f14\u52a9\u8a13\u7df4\u6a21\u578b\uff0c\u78ba\u5be6\u80fd\u8b93\u7d50\u679c\u66f4\u8fd1\u4f3c\u65bc\u76ee\u6a19\u8072\u97f3\uff1b(3) \u975e\u9031\u671f\u6027\u7684\u8cc7\u6599\u5c0d\u9f4a\u61c9</td></tr><tr><td>\u8ddf\u96a8\u983b\u8b5c\u5305\u7d61\uff0c\u4e14\u8f49\u63db\u975e\u9031\u671f\u6027\u5c0d\u65bc\u97f3\u8cea\u6709\u4e00\u5b9a\u7684\u5e6b\u52a9\uff1b(4)\u4ee5\u672c\u8ad6\u6587\u6240\u4f7f\u7528\u7684 4 \u7a2e\u8a9e</td></tr><tr><td>\u7684\u7d50\u679c 86.6%\u7684\u7968\u6578\u89ba\u5f97 Diff \u7684\u8072\u97f3\u6bd4\u8f03\u597d\u807d\uff0c\u5728\u6c92\u6709\u7d93\u904e\u8072\u78bc\u5668\u7684\u60c5\u6cc1\u4e0b\u97f3\u8cea\u807d\u8d77\u4f86 \u901f\u8a9e\u6599\u5eab\u800c\u8a00\uff0c\u5c07\u8a9e\u901f\u6b63\u5e38\u8f49\u5230\u8a9e\u901f\u5feb\u6216\u662f\u8a9e\u901f\u4e2d\u662f\u8f03\u70ba\u6070\u7576\u7684\uff0c\u97f3\u8cea\u4e5f\u6bd4\u8f03\u597d\uff1b(5)\u8de8</td></tr><tr><td>\u5c31\u50cf\u662f\u901a\u904e\u6ffe\u6ce2\u5668\u3002 \u8a9e\u8a00\u7684\u8f49\u63db\u4e2d\uff0c\u767c\u97f3\u76f8\u8fd1\u7684\u6703\u5171\u7528\u76f8\u540c\u7684\u8f49\u63db\u5c0d\u3002\u672c\u7814\u7a76\u5617\u8a66\u7528\u8a9e\u8005\u8f49\u63db\u6280\u8853\u5728\u540c\u4e00\u8a9e</td></tr><tr><td>(\u4e09)\u8f49\u5230\u4e0d\u540c\u8a9e\u901f\u4e4b\u97f3\u8cea\u6bd4\u8f03 \u8005\u7684\u8a9e\u6599\u5eab\u9032\u884c\u97f3\u8cea\u4fee\u5fa9\uff0c\u7531\u4e3b\u89c0\u6e2c\u8a66\u7d50\u679c\u53ef\u5f97\u77e5\uff0c\u5728 DBLSTM\u3001DBLSTM 172 \u548c\u52a0</td></tr><tr><td>\u210e \u20d7 = ( \u210e \u20d7 \u20d7 + \u210e \u20d7 \u20d7 \u210e \u20d7 \u20d7 \u210e \u20d7 \u22121 + \u210e \u20d7 \u20d7 ) \u6211\u5011\u7528 DBLSTM172 \u4f86\u5be6\u9a57\u8a9e\u901f\u6b63\u5e38\u8f49\u5230\u54ea\u7a2e\u8a9e\u901f\u7684\u97f3\u8cea\u6bd4\u8f03\u597d\uff0c\u5728 MCD \u7684\u6bd4\u8f03 (3.4) \u5165 ap \u8f49\u63db\u7684\u5be6\u9a57\u4e2d\u7686\u6709\u6539\u5584\u97f3\u8cea\u548c\u6e1b\u5c11\u96dc\u8a0a\u3002</td></tr><tr><td>\u611f\u6e2c\u5668\u6240\u7d44\u6210\uff0c\u6bcf\u500b\u611f\u6e2c\u5668\u7684\u8f38\u51fa\u662f\u7531\u8f38\u5165\u5411\u91cf\u8207\u6b0a\u91cd\u5411\u91cf\u7684\u5167\u7a4d\u5f8c\uff0c\u7d93\u904e\u4e00\u500b\u975e\u7dda \u6027\u51fd\u6578\u6240\u5f97\u7684\u7d14\u91cf\uff0c\u5176\u67b6\u69cb\u662f\u4e00\u5c64\u8f38\u5165\u5c64\u3001\u4e00\u5c64\u96b1\u85cf\u5c64\u548c\u4e00\u5c64\u8f38\u51fa\u5c64\uff0c\u800c DNN \u5c31\u662f \u591a\u5c64\u7684\u96b1\u85cf\u5c64\u3002\u672c\u7814\u7a76\u7528\u7684\u975e\u7dda\u6027\u7684\u51fd\u6578\u662f sigmoid\uff0c\u8a13\u7df4\u6574\u500b DNN \u7684\u6e96\u5247\u7528\u7684\u662f minimum mean squared error (MMSE)\u3002\u5728\u8a9e\u8005\u8f49\u63db\u4e2d\uff0cDNN \u662f\u97f3\u6846\u5c0d\u97f3\u6846\u7684\u8f49\u63db\uff0c\u8ddf GMM \u4e00\u6a23\u9700\u8981\u6709 delta \u4f5c\u70ba\u8f14\u52a9\u8f38\u5165\u4f86\u8003\u616e\u524d\u5f8c\u97f3\u6846\u7684\u5f71\u97ff\uff0c\u8f49\u63db\u7d50\u679c\u624d\u6703\u6bd4\u8f03\u597d\u3002 3\u3001Deep Bidirectional Long Short-Term Memory (DBLSTM) \u210e \u20d6\u20d7 = ( \u210e \u20d6\u20d7 \u20d7 + \u210e \u20d6\u20d7 \u20d7 \u210e \u20d6\u20d7 \u20d7 \u210e \u20d6\u20d7 +1 + \u210e \u20d6\u20d7 \u20d7 ) (3.5) = \u210e \u20d7 \u20d7 \u210e \u20d7 + \u210e \u20d6\u20d7 \u20d7 \u210e \u20d6\u20d7 + (3.6) 4\u3001Spectral Differential [ \u2206 ]\uff0c\u5c07\u76ee\u6a19\u51fd\u6578\u5beb\u6210\u4e0b\u5217\uff1a \u0302= ( | , ( ) ) (3.7) 2\u3001\u589e\u52a0\u8f38\u5165\u53c3\u6578\u4e4b\u6bd4\u8f03 \u5728\u6620\u5c04\u51fd\u6578\u6bd4\u8f03\u7684\u5be6\u9a57\u4e2d\uff0cDBLSTM \u4f86\u8f49\u63db\u5f97\u5230\u8f03\u597d\u7684\u7d50\u679c\uff0c\u4f46\u662f\u78b0\u5230\u4e86\u5728\u767c\u97f3 \u908a\u754c\u6703\u6709\u96dc\u97f3\u7684\u554f\u984c\uff0c\u70ba\u4e86\u6539\u5584\u9019\u4e00\u554f\u984c\u5728\u6a21\u578b\u8a13\u7df4\u7684\u8f38\u5165\u7aef\u52a0\u5165\u4e86\u8a9e\u8a00\u53c3\u6578(language \u4f86\u5f97\u5230\u5dee\u7570\u7684\u8cc7\u8a0a\uff0c\u672c\u5be6\u9a57\u4ee5 sprocket \u7684\u4f5c\u6cd5\u5be6\u8e10\uff0c\u4e26\u548c\u539f GMM \u8f49\u63db\u7684\u8072\u97f3\u505a\u6bd4\u8f03\u3002 (\u4e8c)\u5be6\u9a57\u7d50\u679c 1 \u6211\u5011\u5c0d\u5916\u90e8\u97f3\u6a94\u5171 25 \u53e5\u4f86\u8f49\u5230\u8a9e\u901f\u4e2d\uff0c\u6bcf\u4e00\u500b\u5be6\u9a57\u6709\u5ba2\u89c0\u8a55\u91cf\u6216\u4e3b\u89c0\u8a55\u91cf\u3002\u5ba2\u89c0 \u8a55\u91cf\u4e2d\uff0c\u6211\u5011\u4e26\u672a\u6709\u6b63\u78ba\u4e14\u9577\u5ea6\u4e00\u81f4\u7684\u97f3\u6a94\u53ef\u4ee5\u4f86\u7b97 stoi \u548c pesq\uff0c\u6240\u4ee5\u6211\u5011\u63d0\u51fa\u7528 source\u3001 = 10 10 \u221a2 \u2211 ( [ ] \u2212 [ ]) 2 =1 (4.1) \u5716 \u4e09\uff1aMCD \u6bd4\u8f03\u793a\u610f\u5716 \u5982\u5716\u4e09\u5c07 source\u3001target \u548c transformed \u7684 MCD \u7b97\u51fa\uff0c\u53ea\u8981\u80fd\u7b26\u5408 > > \uff0c\u5c31\u80fd\u4fdd\u8b49 \u9019\u500b\u8f49\u63db\u7684\u7d50\u679c\u6bd4\u539f\u4f86\u7684\u597d\uff0c\u5982\u679c\u53ea\u6bd4\u8f03 > \uff0c\u90a3\u9ebc\u4ee5\u76ee\u6a19\u70ba\u5713\u5fc3\uff0c\u534a\u5f91\u70ba c \u7684\u5713\u90fd \u6703\u7b26\u5408\u6b64\u689d\u4ef6\u800c\u7121\u6cd5\u78ba\u5b9a\u5713\u4e0a\u7684\u9ede\u4e4b\u9593\u7684\u512a\u52a3\uff1b\u53e6\u5916\u6c92\u6709\u6bd4\u8f03 > \u7684\u8a71\uff0c\u4e5f\u4e0d\u80fd\u78ba\u5b9a \u8f49\u63db\u5f8c\u7684\u7d50\u679c\u662f\u5426\u6709\u6bd4\u8f03\u9760\u8fd1\u76ee\u6a19\u3002\u5728\u505a\u5176\u4ed6\u8b8a\u56e0\u6bd4\u8f03\u7684\u6642\u5019\uff0c\u5728\u7b26\u5408 > > \u60c5\u5f62\u4e0b \u6bd4\u8f03 \u4e5f\u53ef\u4ee5\u770b\u51fa\u54ea\u500b\u8f49\u63db\u7684\u6548\u679c\u8f03\u512a\uff1b\u5728\u4e3b\u89c0\u6e2c\u8a66\uff0c\u6211\u5011\u627e\u4e86 6 \u4f4d\u53d7\u6e2c\u8005\uff0c\u6bcf\u4e00\u500b\u5be6 \u8868 \u4e00\uff1a\u6620\u5c04\u51fd\u6578\u4e4b MCD \u6bd4\u8f03\u548c\u807d\u89ba\u6e2c\u8a66 a b c \u807d\u89ba\u6e2c\u8a66 GMM 7.942 12.815 13.061 26.6% DNN 7.942 6.644 5.298 73.3% 30% DBLSTM 7.942 7.433 5.433 70% 2\u3001\u589e\u52a0\u8f38\u5165\u53c3\u6578\u4e4b\u6bd4\u8f03\u7d50\u679c \u5728\u78ba\u8a8d DBLSTM \u662f\u8f03\u597d\u7684\u67b6\u69cb\u4e4b\u5f8c\uff0c\u6211\u5011\u958b\u59cb\u589e\u52a0\u8f38\u5165\u53c3\u6578\u7684\u5be6\u9a57\uff0c\u70ba\u4e86\u4f7f\u5f97\u8a13 \u7df4\u8f49\u63db\u7684\u7d50\u679c\u66f4\u63a5\u8fd1\u76ee\u6a19\uff0c\u52a0\u5165\u4e86\u8a9e\u8a00\u53c3\u6578-DBLSTM 172 \u548c\u984d\u5916\u518d\u52a0\u5165 uv \u8cc7\u8a0a\u7684-DBLSTM 176\uff0c\u7531\u8868\u4e8c\u53ef\u4ee5\u770b\u51fa\u8ddd\u96e2 c \u7684\u503c\u6709\u6e1b\u5c11\u4e86\u4e00\u4e9b\uff1b\u807d\u89ba\u6e2c\u8a66\u4e2d\u662f 46.6%\u6bd4 53.3%\uff0c \u807d\u611f\u4e0a\u6709\u9ede\u63a5\u8fd1\uff0c\u5fc5\u9808\u7528\u8033\u6a5f\u807d\u624d\u80fd\u807d\u51fa\u767c\u97f3\u908a\u754c\u96dc\u97f3\u7684\u5dee\u7570\u3002\u53e6\u5916\u589e\u52a0 voice/unvoice \u53c3\u6578\u7684\u5be6\u9a57\uff0c\u7531\u65bc\u8ddd\u96e2 b \u548c\u8ddd\u96e2 c \u8ddf DBLSTM 172 \u662f\u4e00\u6a23\u7684\u6240\u4ee5\u4e0d\u505a\u807d\u529b\u6e2c\u8a66\u3002 \u8868 \u4e8c\uff1a\u589e\u52a0\u8f38\u5165\u53c3\u6578\u5be6\u9a57\u4e4b\u6bd4\u8f03 MCD \u548c\u807d\u89ba\u6e2c\u8a66 a b c \u807d\u89ba\u6e2c\u8a66 DBLSTM 7.942 7.433 5.433 46.6% DBLSTM 172 7.942 6.818 4.732 53.3% DBLSTM 176 7.942 6.818 4.732 \u8a9e\u901f\u7684\u7d50\u679c\u6bd4\u8f03\u50cf\u8a72\u8a9e\u901f\uff0c\u7531\u8868\u4e09\u53ef\u4ee5\u770b\u51fa\u8a9e\u901f\u6162\u8ddd\u96e2\u8a9e\u901f\u6b63\u5e38\u662f\u6700\u9060\u7684\uff0c\u8a9e\u901f\u5feb\u548c\u8a9e \u901f\u4e2d\u8ddd\u96e2\u8a9e\u901f\u6b63\u5e38\u7684\u9577\u5ea6\u5dee\u4e0d\u591a\u4e14\u8a9e\u901f\u5feb\u7684\u8ddd\u96e2 c \u6bd4\u8a9e\u901f\u4e2d\u7684\u7565\u5c0f\uff0c\u53ef\u4ee5\u5f97\u77e5\u8a9e\u901f\u5feb\u548c \u8a9e\u901f\u4e2d\u5404\u81ea\u7684\u8f49\u63db\u6548\u679c\u5dee\u4e0d\u591a\u597d\uff1b\u800c\u5728\u807d\u89ba\u6e2c\u8a66\u4e2d 42.8%\u7684\u8a9e\u901f\u4e2d\u5c0d\u4e0a 50%\u8a9e\u901f\u5feb\uff0c\u5728 \u807d\u611f\u4e0a\u53ef\u80fd\u8a9e\u901f\u5feb\u6703\u597d\u4e00\u4e9b\uff0c\u7d50\u8ad6\u662f\u8f49\u5230\u8a9e\u901f\u4e2d\u6216\u662f\u8a9e\u901f\u5feb\u90fd\u662f\u4e0d\u932f\u7684\u3002 \u8868 \u4e09\uff1a\u8f49\u5230\u4e0d\u540c\u8a9e\u901f\u5be6\u9a57\u4e4b MCD \u6bd4\u8f03 a b c \u807d\u89ba\u6e2c\u8a66 \u8a9e\u901f\u4e2d 7.942 6.818 4.732 42.8% \u8a9e\u901f\u5feb 7.976 6.530 4.710 50% \u8a9e\u901f\u6162 8.632 6.706 5.085 7.1% (\u56db)\u4e2d\u6587\u8f49\u63db\u6a21\u578b\u5c0d\u4e2d\u82f1\u593e\u96dc\u8a9e\u6599\u4e4b\u5617\u8a66 \u5728\u78ba\u8a8d\u4e2d\u6587\u5c0d\u4e2d\u6587\u7684\u8f49\u63db\u4e0a\u97f3\u8cea\u78ba\u5be6\u6709\u8b8a\u597d\u4e4b\u5f8c\uff0c\u6211\u5011\u5c0d CE_word \u548c CE_spell \u5206 \u5225\u505a\u8a9e\u8005\u8f49\u63db\uff0c\u8f49\u63db\u6a21\u578b\u662f DBLSTM\uff0c\u8f38\u5165\u662f 40 \u7dad\u3002\u7531\u65bc\u662f\u8f49\u6210\u4e2d\u82f1\u593e\u96dc\u7684\u53e5\u5b50\uff0c\u6c92 \u6709\u5167\u5bb9\u4e00\u81f4\u4e14\u97f3\u8cea\u8f03\u597d\u7684\u97f3\u6a94\u53ef\u4f9b\u8a08\u7b97 MCD\uff0c\u53ea\u80fd\u505a\u807d\u89ba\u6e2c\u8a66\u3002CE_spell \u6295\u7968\u7d50\u679c\u986f \u793a\u8f49\u5f97\u4e26\u4e0d\u597d\uff0c\u807d\u611f\u4e0a\u56de\u97f3\u7684\u90e8\u5206\u5c11\u4e86\u5f88\u591a\u4f46\u6709\u9ede\u7834\u55d3\uff0c\u6574\u9ad4\u611f\u89ba\u9084\u662f\u6c92\u6709\u539f\u672c\u597d\uff1b CE_word \u6295\u7968\u7d50\u679c\u5404 50%\uff0c\u6709\u6539\u5584\u5230\u67d0\u4e9b\u4f4e\u983b\u96dc\u97f3\u3002\u96d6\u7136\u6574\u9ad4\u5be6\u9a57\u6548\u679c\u6c92\u6709\u7279\u5225\u986f\u8457\uff0c \u4f46\u6211\u5011\u767c\u73fe\u5373\u4f7f\u539f\u672c\u4e2d\u6587\u8f49\u63db\u6a21\u578b\u6c92\u770b\u904e\u82f1\u6587\u7684\u767c\u97f3\uff0c\u5728\u8f49\u63db\u7684\u904e\u7a0b\u9084\u662f\u6703\u8f49\u6210\u985e\u4f3c\u7684 \u767c\u97f3\uff0c\u4f7f\u5f97\u97f3\u6a94\u4e2d\u82f1\u6587\u7684\u90e8\u5206\u9084\u662f\u80fd\u7684\u807d\u51fa\u4f86\u3002 \u4e94\u3001\u7d50\u8ad6 \u672c\u8ad6\u6587\u63d0\u51fa\u7528\u8a9e\u8005\u8f49\u63db\u7684\u6280\u8853\u7528\u65bc\u4fee\u5fa9\u8a9e\u97f3\u8cc7\u6599\u5eab\uff0c\u85c9 \u7531\u540c\u4e00\u4f4d\u8a9e\u8005\u7684\u7279\u6027\u4f7f\u5f97\u8a9e \u9a57\u5f9e 25 \u53e5\u88e1\u96a8\u6a5f\u62bd 5 \u4e2d\uff0cDBLSTM \u662f\u6211\u5011\u89ba\u5f97\u8868\u73fe\u8f03\u597d\u7684\u3002 \u4e2d\uff0c\u7531\u65bc\u5404\u8a9e\u901f\u548c\u8a9e\u901f\u6b63\u5e38\u8ddd\u96e2\u4e26\u4e0d\u4e00\u81f4\u4e14\u5206\u5e03\u4e0d\u540c\uff0c\u6240\u4ee5\u53ea\u80fd\u6bd4\u8f03\u8a9e\u901f\u6b63\u5e38\u8f49\u5230\u54ea\u7a2e \u53c3\u8003\u6587\u737b</td></tr><tr><td>\u8005\u8f49\u63db\u8b8a\u5f97\u50cf\u662f\u97f3\u8cea\u8f49\u63db\uff0c\u7d93\u904e\u7814\u7a76\u548c\u5206\u6790\u5f97\u51fa\u4ee5\u4e0b\u7d50\u8ad6\uff1a(1)\u5728\u8f49\u63db\u6a21\u578b\u7684\u67b6\u69cb\u4e0a\uff0c</td></tr></table>", | |
| "text": "Term Memory (LSTM)\u3002\u672c\u7814\u7a76\u4f7f\u7528\u7684\u662f Deep Bidirectional LSTM (DBLSTM)\uff0cDBLSTM \u662f\u5c07\u591a\u5c64 BLSTM \u758a\u5728\u4e00\u8d77\u4f86\u52a0\u5f37\u5b78\u7fd2\u6548\u679c\u3002 BLSTM \u96b1\u85cf\u5c64\u5206\u70ba\u6642\u5e8f\u5411\u524d\u548c\u6642\u5e8f\u5411\u5f8c\u5169\u7a2e\uff0c\u9019\u6a23\u7684\u67b6\u69cb\u53ef\u4ee5\u4f7f\u4e0b\u4e00\u6642\u523b\u7684\u8f38\u51fa\u8003\u616e \u524d\u5f8c\u6587\u7684\u95dc\u4fc2\uff0c\u5c31\u4e0d\u9700\u8981 delta \u4f5c\u70ba\u8f14\u52a9\u8f38\u5165\uff0c\u210e \u20d7 \u70ba\u6642\u5e8f\u5411\u524d\uff0c\u8fed\u4ee3\u8a08\u7b97t = 1 \u5230 t = T\uff1b \u210e \u20d6\u20d7 \u70ba\u6642\u5e8f\u5411\u5f8c\uff0c\u8fed\u4ee3\u8a08\u7b97t = T \u5230 t = 1\uff0c ( * )\u8868\u793a\u6574\u500b LSTM \u55ae\u5143\u904e\u7a0b\uff0cBLSTM \u55ae\u5c64 \u904b\u7b97\u5982\u5f0f(3.4~3.6) \uff0c\u6574\u500b DBLSTM \u53ef\u4ee5\u7528 back propagation though time(BPTT)\u7b97\u51fa\u6b0a\u91cd\u3002 \u8f38\u5165 40 \u7dad\uff0c\u8f38\u51fa 40 \u7dad\uff1b\u53e6\u4e00\u7a2e\u662f\u52a0\u5165\u8a9e\u8a00\u53c3\u6578\u7684\u8f38\u5165 172 \u7dad\uff0c\u8f38\u51fa 40 \u7dad\uff0c\u5169\u7a2e\u7684\u8a13 \u7df4\u9664\u4e86\u8f38\u5165\u7dad\u5ea6\u4e0d\u540c\uff0c\u5176\u67b6\u69cb\u90fd\u540c DBLSTM \u7684\u5be6\u9a57\u3002 4\u3001Spectral Differential \u8207 GMM \u8f49\u63db\u4e4b\u6bd4\u8f03 Diff \u7684\u505a\u6cd5\u548c GMM \u8a9e\u8005\u8f49\u63db\u5728\u8cc7\u6599\u5c0d\u9f4a\u4e4b\u524d\u90fd\u4e00\u81f4\uff0c\u5c0d\u9f4a\u8cc7\u8a0a\u9084\u662f\u4f9d\u64da MCC \u5c0d \u9f4a\u5f8c\u7684\u7d22\u5f15\u503c\uff0c\u518d\u8a08\u7b97 MCC differential \u7684\u90e8\u5206 = \u2212 \uff0c\u8a13\u7df4\u51fa\u4e00\u500b \u8f49\u5230 \u7684\u8f49\u63db\u6a21 \u578b\uff0c\u5c07\u8f49\u51fa \u2032 \u7684\u4f5c\u70ba MLSA filter \u7684\u53c3\u6578\uff0c\u8f38\u5165\u70ba\u4f86\u6e90\u97f3\u6a94\uff0c\u5f97\u51fa\u8072\u97f3\u3002\u4f46 sprocket \u4e2d\uff0c \u4e26\u672a\u8a13\u7df4\u4e00\u500b \u8f49\u5230 \u7684\u8f49\u63db\u6a21\u578b\uff0c\u800c\u662f\u7528 \u8f49\u5230 y \u7684\u8f49\u63db\u6a21\u578b\uff0c\u8f49\u51fa \u2032 \u4e4b\u5f8c\uff0c\u4ee4 \u2032 = \u2032 \u2212 \u53e5\u807d\u6e2c\uff0c\u8b93\u6e2c\u8a66\u8005\u6295\u7968\u54ea\u500b\u97f3\u8cea\u6bd4\u8f03\u597d\uff0c\u518d\u6bd4\u8f03\u7968\u6578\u7b97\u6bd4\u4f8b\u3002 1\u3001\u6620\u5c04\u51fd\u6578\u4e4b\u6bd4\u8f03\u7d50\u679c \u8868\u4e00\u662f GMM\u3001DNN \u548c DBLSTM \u7684 MCD \u6bd4\u8f03\u548c\u7968\u6578\u6bd4\u8f03\u3002\u5728\u5ba2\u89c0\u8a55\u91cf\u4e2d\uff0cGMM \u4e26\u672a\u7b26\u5408 > > \uff0c\u4e14 \u548c \u9084\u6bd4 \u9577\uff0c\u5728\u807d\u611f\u4e0a\u6709\u9ede\u50cf\u8a9e\u901f\u4e2d\u4f46\u53c8\u6709\u9ede\u4e0d\u50cf\u7684\u6a23\u5b50\u3002DNN \u8f49\u63db\u7684\u7d50\u679c\u7b26\u5408 > > \uff0c\u5728\u8ddd\u96e2 c \u7684\u5dee\u7570\u4e0a DNN \u7684\u8868\u73fe\u6bd4\u8f03\u597d\uff1b\u5728\u807d\u89ba\u6e2c\u8a66\u7684\u6295\u7968 \u4e2d\uff0c73.3%\u8a8d\u70ba DNN \u7684\u97f3\u8cea\u6bd4 GMM \u597d\u3002DBLSTM \u7684\u8f49\u63db\u7d50\u679c\u4e5f\u7b26\u5408 > > \u3002\u5728 DNN \u8ddf DBLSTM \u7684\u6bd4\u8f03\u4e2d\uff0c\u8ddd\u96e2 c \u7684\u503c\u662f DNN \u7565\u5c0f\uff0c\u4f46\u5728\u807d\u89ba\u6e2c\u8a66\u7684\u6295\u7968\u4e2d\uff0c\u50c5\u6709 30%\u8a8d\u70ba DNN \u97f3\u8cea\u8f03\u597d\uff0c70%\u8a8d\u70ba DBLSTM \u97f3\u8cea\u6bd4 DNN \u597d\uff0c\u5728\u8f49\u63db\u6a21\u578b\u7684\u5be6\u9a57\u6bd4\u8f03 3\u3001\u52a0\u5165\u975e\u9031\u671f\u6027\u8f49\u63db\u4e4b\u6bd4\u8f03\u7d50\u679c \u5728\u52a0\u5165 ap \u8f49\u63db\u7684\u5be6\u9a57\u4e2d\uff0cMCC \u7684\u8f49\u63db\u662f\u7528 DBLSTM 172\uff0c\u6240\u4ee5 MCD \u7684\u6bd4\u8f03\u6703\u4e00 \u81f4\uff0c\u6211\u5011\u7528 ap \u52a0\u8a9e\u8a00\u53c3\u6578\u7684\u5be6\u9a57\u4f86\u505a\u807d\u89ba\u6e2c\u8a66\uff0c\u807d\u89ba\u6e2c\u8a66\u7684\u7d50\u679c 53.3%\u7684\u7968\u6578\u89ba\u5f97\u6709 \u8f49 ap \u7684\u7d50\u679c\u807d\u8d77\u4f86\u6bd4\u8f03\u597d\u4e00\u9ede\uff0c\u807d\u611f\u4e0a\u6709\u8f49 ap \u7684\u7d50\u679c\u6bd4\u8f03\u6c92\u6709\u55e1\u55e1\u7684\u8072\u97f3\u3002 4\u3001Spectral Differential \u8207 GMM \u8f49\u63db\u4e4b\u6bd4\u8f03\u7d50\u679c \u56e0\u70ba GMM Diff \u6c92\u6709\u7d93\u904e\u8072\u78bc\u5668\uff0c\u6240\u4ee5\u4e0d\u505a MCD \u6bd4\u8f03\uff0c\u53ea\u505a\u807d\u89ba\u6e2c\u8a66\u3002\u807d\u89ba\u6e2c\u8a66", | |
| "html": null, | |
| "type_str": "table", | |
| "num": null | |
| } | |
| } | |
| } | |
| } |