| { | |
| "paper_id": "2020", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T07:27:21.678769Z" | |
| }, | |
| "title": "A Preliminary Study on Deep Learning-based Chinese Text to Taiwanese Speech Synthesis System", | |
| "authors": [ | |
| { | |
| "first": "\u8a31\u6587\u6f22", | |
| "middle": [ | |
| "\uf02a" | |
| ], | |
| "last": "\u3001\u66fe\u8b49\u878d", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "\uf02a", | |
| "middle": [], | |
| "last": "\u3001\u5ed6\u5143\u752b", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "\uf02a", | |
| "middle": [], | |
| "last": "\u3001\u738b\u6587\u4fca", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "\uf02b", | |
| "middle": [], | |
| "last": "\u3001\u6f58\u632f\u9298 \uf02b", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Wen-Han", | |
| "middle": [], | |
| "last": "Hsu", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Cheng-Jung", | |
| "middle": [], | |
| "last": "Tseng", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Yuan-Fu", | |
| "middle": [], | |
| "last": "Liao", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "yfliao@ntut.edu.tw" | |
| }, | |
| { | |
| "first": "Wern-Jun", | |
| "middle": [], | |
| "last": "Wang", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Chen-Ming", | |
| "middle": [], | |
| "last": "Pan", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "chenming@cht.com.tw" | |
| } | |
| ], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "WaveGlow \u6a21\u578b(Spectrogram to Waveform)\u4f86\u5be6\u73fe\u8a9e\u97f3\u5408\u6210\u3002\u540c\u6642\u6709\u67b6\u8a2d\u7db2\u9801 \u53ef\u4f9b\u4f7f\u7528\u8005\u4e00\u540c\u4f86\u6e2c\u8a66\u6210\u6548\u3002 \u672c\u6587 C2T \u6a5f\u5668\u7ffb\u8b6f\u7684\u5be6\u9a57\u65b9\u9762\u63a1\u53d6\u4e09\u7a2e\u6a21\u5f0f\uff0c\u5305\u62ec(1)\u8f38\u5165\u4e2d\u6587\u5b57\u8a5e\uff0c\u5148\u9032\u884c \u65b7\u8a5e\uff0c\u518d\u8f38\u51fa\u6bcf\u500b\u4e2d\u6587\u8a5e\u7684\u53f0\u8a9e\u53f0\u7f85(T\u00e2i-l\u00f4)\u62fc\u97f3\u3002(2)\u8f38\u5165\u4e2d\u6587\u5b57\u5143\u4e32\uff0c\u76f4 \u63a5\u8f38\u51fa\u53f0\u7f85\u62fc\u97f3\u4e32\u3002(3)\u8f38\u5165\u4e2d\u6587\u5b57\u5143\u4e32\uff0c\u8f38\u51fa\u53f0\u8a9e\u7684\u53f0\u7f85\u62fc\u97f3\u4e32\u8207\u53f0\u8a9e\u8a5e\u7684 \u65b7\u8a5e\u95dc\u4fc2\u3002\u82e5\u4e0d\u8003\u616e\u8072\u8abf\uff0c\u65b9\u6cd5(1)\u7684 syllable error rate(SER)\u70ba 15.66%\u3002\u800c\u65b9 \u6cd5(2)\u7684 SER \u66f4\u53ef\u9054 6.53%\u3002\u9019\u8868\u793a\u6211\u5011\u6240\u7528\u7684 sequence-to-sequence \u6a21\u578b\u78ba\u5be6 \u53ef\u4ee5\u6b63\u78ba\u5730\u5c07\u8f38\u5165\u7684\u4e2d\u6587\u5b57\u5143\u4e32\uff0c\u76f4\u63a5\u8f38\u51fa\u53f0\u7f85\u62fc\u97f3\u4e32\u3002", | |
| "pdf_parse": { | |
| "paper_id": "2020", | |
| "_pdf_hash": "", | |
| "abstract": [ | |
| { | |
| "text": "WaveGlow \u6a21\u578b(Spectrogram to Waveform)\u4f86\u5be6\u73fe\u8a9e\u97f3\u5408\u6210\u3002\u540c\u6642\u6709\u67b6\u8a2d\u7db2\u9801 \u53ef\u4f9b\u4f7f\u7528\u8005\u4e00\u540c\u4f86\u6e2c\u8a66\u6210\u6548\u3002 \u672c\u6587 C2T \u6a5f\u5668\u7ffb\u8b6f\u7684\u5be6\u9a57\u65b9\u9762\u63a1\u53d6\u4e09\u7a2e\u6a21\u5f0f\uff0c\u5305\u62ec(1)\u8f38\u5165\u4e2d\u6587\u5b57\u8a5e\uff0c\u5148\u9032\u884c \u65b7\u8a5e\uff0c\u518d\u8f38\u51fa\u6bcf\u500b\u4e2d\u6587\u8a5e\u7684\u53f0\u8a9e\u53f0\u7f85(T\u00e2i-l\u00f4)\u62fc\u97f3\u3002(2)\u8f38\u5165\u4e2d\u6587\u5b57\u5143\u4e32\uff0c\u76f4 \u63a5\u8f38\u51fa\u53f0\u7f85\u62fc\u97f3\u4e32\u3002(3)\u8f38\u5165\u4e2d\u6587\u5b57\u5143\u4e32\uff0c\u8f38\u51fa\u53f0\u8a9e\u7684\u53f0\u7f85\u62fc\u97f3\u4e32\u8207\u53f0\u8a9e\u8a5e\u7684 \u65b7\u8a5e\u95dc\u4fc2\u3002\u82e5\u4e0d\u8003\u616e\u8072\u8abf\uff0c\u65b9\u6cd5(1)\u7684 syllable error rate(SER)\u70ba 15.66%\u3002\u800c\u65b9 \u6cd5(2)\u7684 SER \u66f4\u53ef\u9054 6.53%\u3002\u9019\u8868\u793a\u6211\u5011\u6240\u7528\u7684 sequence-to-sequence \u6a21\u578b\u78ba\u5be6 \u53ef\u4ee5\u6b63\u78ba\u5730\u5c07\u8f38\u5165\u7684\u4e2d\u6587\u5b57\u5143\u4e32\uff0c\u76f4\u63a5\u8f38\u51fa\u53f0\u7f85\u62fc\u97f3\u4e32\u3002", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Abstract", | |
| "sec_num": null | |
| } | |
| ], | |
| "body_text": [ | |
| { | |
| "text": "\u9020\u6210\u53f0\u8a9e\u4f7f\u7528\u4eba\u53e3\u5f0f\u5fae\u7684\u539f\u56e0\u5f88\u591a\uff0c\u6700\u65e9\u53ef\u8ffd\u6eaf\u81f3\u6c11\u570b", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "\u7dd2\u8ad6 (Introduction)", | |
| "sec_num": "1." | |
| }, | |
| { | |
| "text": "\u70ba\u8a13\u7df4\u6b64\u7cfb\u7d71\u4e2d\u7684 C2T \u6a21\u7d44\uff0c\u6211\u5011\u5c07\u5229\u7528\u4e2d\u7814\u9662\u7684 iCorpus \u53f0\u83ef\u5e73\u884c\u8a9e\u6599\u5eab\uff0c\u8207\u5f9e \u7db2 \u8def \u4e0a \u6536 \u96c6 \u7684 \u591a \u672c \u53f0 \u83ef \u5e73 \u884c \u8fad \u5178 ( \u5305 \u542b \u6559 \u80b2 \u90e8 \u95a9 \u5357 \u8a9e \u5e38 \u7528 \u8a5e \u8fad \u5178 ) \uff0c \u4e26 \u63a1 \u7528 sequence-to-sequence \u6df1\u5ea6\u985e\u795e\u7d93\u7db2\u8def\u67b6\u69cb\uff0c\u8b93\u6a21\u578b\u53bb\u5b78\u7fd2\u5982\u4f55\u5c07\u4e2d\u6587\u6587\u5b57\uff0c\u8f49\u63db\u6210\u53f0\u7f85 \u62fc\u97f3\u3002\u4e26\u5229\u7528\u4e2d\u83ef\u96fb\u4fe1\u9304\u88fd\u7684\u55ae\u4e00\u8a9e\u8005\u53f0\u8a9e\u8a9e\u97f3\u5408\u6210\u8a9e\u6599\u5eab\uff0c\u8a13\u7df4 Tacotron2 \u8207 Waveglow\u3002 \u5e0c\u671b\u80fd\u76e1\u53ef\u80fd\u5730\u9054\u5230\u4e2d\u6587\u6587\u5b57\u7ffb\u8b6f\u6210\u53f0\u7f85\u62fc\u97f3\u7684\u6b63\u78ba\u6027\uff0c\u8207\u5408\u6210\u53f0\u8a9e\u8a9e\u97f3\u7684\u9ad8\u5ea6\u81ea\u7136 \u5ea6\u3002 2. \u4e2d \u6587 \u6587 \u5b57 \u8f49 \u53f0 \u8a9e \u8a9e \u97f3 \u5408 \u6210 \u7cfb \u7d71 (", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "\u7dd2\u8ad6 (Introduction)", | |
| "sec_num": "1." | |
| }, | |
| { | |
| "text": "\u6211\u5011\u63d0\u51fa\u7684 C2T \u6a5f\u5668\u7ffb\u8b6f\uff0c\u6700\u597d\u7684 CER \u503c\u70ba 6.53%\u3002\u8868\u793a sequence-to-sequence \u6a21\u578b\u78ba \u5be6\u53ef\u4ee5\u5c07\u8f38\u5165\u7684\u4e2d\u6587\u6587\u672c\uff0c\u7ffb\u8b6f\u6210\u6b63\u78ba\u7684\u53f0\u7f85\u62fc\u97f3\u4e32\u3002\u53f0\u8a9e\u8a9e\u97f3\u5408\u6210\u7684 MOS \u5f97\u5206\u70ba 4.30 \u5206\uff0c\u8868\u793a Tacrtron2+WaveGlow ", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 110, | |
| "end": 128, | |
| "text": "Tacrtron2+WaveGlow", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "\u7d50\u8ad6(Conclusion)", | |
| "sec_num": "5." | |
| }, | |
| { | |
| "text": "\u9b25\u62cd\u5b57\uff0chttps://suisiann.ithuan.tw/", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| } | |
| ], | |
| "back_matter": [ | |
| { | |
| "text": "Ott, M., Edunov, S., Grangier, D., & Auli, M. (2018) . Scaling Neural Machine Translation. In arXiv preprint arxiv: 1806.00187Prenger, R., Valle, R., & Catanzaro, B. (2018) . WaveGlow: A Flow-based Generative Network for Speech Synthesis. In arXiv preprint arxiv: 1811.00002Shen, J., Pang, R., Weiss, R. J., Schuster, M., Jaitly, N., Yang, Z., \u2026Wu, Y. (2018) Inf. Technol., 5(1) , 118-128. \u6a5f\u5668\u7ffb\u8b6f(2020 \u5e74 9 \u6708 17 \u65e5)\u3002In Wikipedia, the free encyclopedia. Retrieved November 10, 2020, from https://zh.wikipedia.org/wiki/ \u673a \u5668 \u7ffb \u8bd1 [Machine Translation (2020, September 17) . In Wikipedia, the free encyclopedia. Retrieved November 10, 2020, from https://zh.wikipedia.org/wiki/\u673a\u5668\u7ffb\u8bd1]", | |
| "cite_spans": [ | |
| { | |
| "start": 9, | |
| "end": 52, | |
| "text": "Edunov, S., Grangier, D., & Auli, M. (2018)", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 139, | |
| "end": 172, | |
| "text": "Valle, R., & Catanzaro, B. (2018)", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 284, | |
| "end": 358, | |
| "text": "Pang, R., Weiss, R. J., Schuster, M., Jaitly, N., Yang, Z., \u2026Wu, Y. (2018)", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 359, | |
| "end": 373, | |
| "text": "Inf. Technol.,", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 374, | |
| "end": 378, | |
| "text": "5(1)", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 522, | |
| "end": 563, | |
| "text": "[Machine Translation (2020, September 17)", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "annex", | |
| "sec_num": null | |
| } | |
| ], | |
| "bib_entries": {}, | |
| "ref_entries": { | |
| "FIGREF0": { | |
| "uris": null, | |
| "text": "supported partially by \u4e2d\u83ef\u96fb\u4fe1 under the project \"\u57fa\u65bc\u6df1\u5ea6\u5b78\u7fd2\u4e4b\u53f0\u8a9e\u8a9e\u97f3\u5408\u6210 \u300e\u6587\u5b57-\u8072\u5b78\u53c3\u6578\u6a21\u578b\u300f \"\uff0c Taiwan's Ministry of Education under the project \"\u6559\u80b2\u90e8\u95a9\u5357\u8a9e\u8a9e\u97f3\u8a9e\u6599\u5eab\u5efa\u7f6e\u8a08\u5283\" and partially by Ministry of Science and Technology under the contract number 107-2221-E-027-102, 107-2911-I-027-501, 107-3011-F-027-003, 108-2221-E-027-067 and 109-2221-E-027-108. \u53c3\u8003\u6587\u737b (References) Kuo, W.-C., Wang, Y.-R., & Chen, S.-H. (2004). A MODEL-BASED TONE LABELING METHOD FOR MIN-NAN/TAIWANESE SPEECH. In Proceddings of IEEE International Conference on Acoustics, Speech, and Signal Processing 2004, 505-508. doi: 10.1109/ICASSP.2004.1326033 Lin, C.-J. & Chen, H.-H. (1999). A Mandarin to Taiwanese Min Nan Machine Translation System with Speech Synthesis of Taiwanese Min Nan. Int. J. Comput. Linguist. Chinese Lang. Process., 4(1), 59-84.", | |
| "num": null, | |
| "type_str": "figure" | |
| }, | |
| "TABREF0": { | |
| "num": null, | |
| "text": "", | |
| "type_str": "table", | |
| "content": "<table><tr><td>\u57fa\u65bc\u6df1\u5ea6\u5b78\u7fd2\u4e4b\u4e2d\u6587\u6587\u5b57\u8f49\u53f0\u8a9e\u8a9e\u97f3\u5408\u6210\u7cfb\u7d71\u521d\u6b65\u63a2\u8a0e</td><td>71 \u8a31\u6587\u6f22 \u7b49</td></tr><tr><td colspan=\"2\">\u6d41\u5229\u3002\u6b64\u5916\uff0c\u53f0\u8a9e\u8a9e\u8a00\u7684\u6f14\u9032\u4e5f\u6162\u6162\u5730\u8207\u751f\u6d3b\u812b\u7bc0\uff0c\u5e38\u6709\u4e00\u4e9b\u65b0\u7684\u6642\u3001\u4e8b\u3001\u7269\uff0c\u4f8b\u5982\"\u78d0 \u4fc2\u3002\u4f8b\u5982\u7d66 encoder \u7aef\u7684 RNN \u8f38\u5165\u4e00\u500b\u4f86\u6e90\u8a9e\u8a00\u7684\u53e5\u5b50\u5f8c\uff0c\u5148\u5229\u7528 RNN \u5206\u6790\u4f86\u6e90\u6587\u5b57</td></tr><tr><td colspan=\"2\">\u77f3\u8266\u3001\u8a0e\u62cd\u3001\u6ed1\u9f20\"\u7b49\u7b49\uff0c\u90fd\u4e0d\u77e5\u9053\u5982\u4f55\u7528\u53f0\u8a9e\u4f86\u8aaa\uff0c\u9020\u6210\u5927\u5bb6\u5728\u7528\u53f0\u8a9e\u8b1b\u8a71\u6642\uff0c\u53ea\u597d \u7684\u8a9e\u610f\uff0c\u7de8\u78bc\u6210\u4e00\u500b\u80fd\u4ee3\u8868\u539f\u8a9e\u53e5\u7684\u8a9e\u610f\u5411\u91cf\u5e8f\u5217\u3002\u518d\u8b93 decoder \u7aef\u7684 RNN\uff0c\u4ee5\u76ee\u6a19\u8a9e</td></tr><tr><td colspan=\"2\">\u5e38\u5e38\u593e\u96dc\u570b\u8a9e\u3002 \u8a00\u7684\u8a9e\u8a00\u6a21\u578b\u77e5\u8b58\uff0c\u91cd\u65b0\u89e3\u8b6f\u8a72\u8a9e\u610f\uff0c\u8f38\u51fa\u5408\u4e4e\u76ee\u6a19\u8a9e\u8a00\u67b6\u69cb\u7684\u8a9e\u53e5(Lee, 2019)\u3002\u9019\u6a23</td></tr><tr><td colspan=\"2\">\u91dd\u5c0d\u53f0\u8a9e\u73fe\u968e\u6bb5\u7684\u56f0\u5883\uff0c\u82e5\u80fd\u505a\u51fa\u4e00\u5957\u4e2d\u6587\u6587\u5b57\u8f49\u53f0\u8a9e\u8a9e\u97f3\u5408\u6210\u7684\u6a5f\u5668\u7ffb\u8b6f\u7cfb\u7d71\uff0c \u5c31\u53ef\u4ee5\u8b93\u7ffb\u8b6f\u7d50\u679c\u540c\u6642\u7b26\u5408\u8a5e\u5f59\u3001\u6587\u6cd5\u8207\u8a9e\u610f\u3002</td></tr><tr><td colspan=\"2\">\u8b93\u4f7f\u7528\u8005\u8f38\u5165\u4e2d\u6587\u6587\u5b57\u5f8c\uff0c\u80fd\u81ea\u52d5\u5408\u6210\u53f0\u8a9e\u8a9e\u97f3\uff0c\u5c31\u53ef\u4ee5\u6559\u5927\u5bb6\u5982\u4f55\u8b1b\u53f0\u8a9e\u3002\u8b93\u4f7f\u7528\u8005 \u53e6\u4e00\u65b9\u9762\uff0c\u76ee\u524d\u7684\u4e3b\u6d41\u8a9e\u97f3\u5408\u6210\uff0c\u4e5f\u5e7e\u4e4e\u90fd\u662f\u57fa\u65bc\u985e\u795e\u7d93\u7db2\u8def\u6280\u8853\uff0c\u5c24\u5176\u4ee5 Google</td></tr><tr><td colspan=\"2\">\u5c0d\u53f0\u8a9e\u63d0\u8d77\u8208\u8da3\uff0c\u4e26\u52a0\u5f37\u53f0\u8a9e\u5728\u65e5\u5e38\u751f\u6d3b\u4e2d\u7684\u61c9\u7528\uff0c\u9032\u800c\u6d3b\u5316\u53f0\u8a9e\u3002\u5c24\u5176\u82e5\u80fd\u540c\u6642\u986f\u793a \u63d0\u51fa\u7684 Tacotron2+WaveNet Vocoder \u8f03\u70ba\u51fa\u540d\u3002Tacotron2 \u53ef\u76f4\u63a5\u4ee5\u985e\u795e\u7d93\u7db2\u8def\uff0c\u9032\u884c\u6587</td></tr><tr><td colspan=\"2\">\u6559\u80b2\u90e8\u5b98\u65b9\u63a8\u85a6\u7684\u53f0\u7f85\u62fc\u97f3\u66f8\u5beb\u7cfb\u7d71\uff0c\u5c31\u80fd\u8b93\u5b78\u751f\u5c0d\u53f0\u7f85\u62fc\u97f3\u6709\u521d\u6b65\u7684\u8a8d\u8b58\u53ca\u77ad\u89e3\uff0c\u9032 \u8108\u8a0a\u606f\u8655\u7406\uff0c\u5efa\u7acb\u4e00\u300c\u6587\u5b57\u300d\u8f49\u300cMel-Spectrogram\u300d\u7684 end-to-end \u67b6\u69cb\u3002WaveNet Vocoder</td></tr><tr><td colspan=\"2\">\u800c\u80fd\u76f4\u63a5\u66f8\u5beb\u53f0\u8a9e\u3002\u5efa\u7acb\u8d77\u4e00\u5957\u4e2d\u6587\u6587\u5b57\u8f49\u53f0\u8a9e\u8a9e\u97f3\u5408\u6210\u7684\u6a5f\u5668\u7ffb\u8b6f\u7cfb\u7d71\uff0c\u901a\u5e38\u9700\u8981\u4e09 \u63a5\u8457\u5c07\u300cMel-Spectrogram\u300d\u8f49\u6210\u300cSpeech Waveform\u300d\u3002\u6b64 Vocoder \u51fa\u73fe\u4ee5\u5f8c\uff0c\u8a9e\u97f3\u5408\u6210</td></tr><tr><td colspan=\"2\">\u500b\u6a21\u7d44\uff0c\u5305\u62ec(1)\u5c07\u4e2d\u6587\u6587\u5b57\u8f49\u6210\u4ee5\u53f0\u7f85(T\u00e2i-l\u00f4)\u62fc\u97f3\u8868\u793a\u7684\u53f0\u8a9e\u8b1b\u6cd5\uff0c(2)\u5c07\u53f0\u7f85\u62fc\u97f3 \u7684\u97f3\u8cea\u5c31\u5e7e\u4e4e\u63a5\u8fd1\u4eba\u8072\u3002Tacotron2+WaveNet Vocoder \u5169\u8005\u7684\u7d44\u5408\u57fa\u672c\u4e0a\u5c31\u662f\u76ee\u524d\u7684</td></tr><tr><td colspan=\"2\">\u8f49\u70ba\u53f0\u8a9e\u5408\u6210\u8a9e\u97f3\u53c3\u6578\uff0c\u6700\u5f8c(3)\u5c07\u5408\u6210\u8a9e\u97f3\u53c3\u6578\u8f49\u6210\u5be6\u969b\u53f0\u8a9e\u5408\u6210\u97f3\u6a94\u3002\u5176\u4e2d\uff0c\u5176\u4e2d\u4ee5 State-of-the-Art \u8a9e\u97f3\u5408\u6210\u6280\u8853\u3002\u4f46\u6b64\u8655\u7684 WaveNet Vocoder\uff0c\u662f\u4e00\u500b\u4ee5 sample \u70ba\u55ae\u4f4d\u505a</td></tr><tr><td colspan=\"2\">\u5c07\u4e2d\u6587\u6587\u5b57\u8f49\u6210\u53f0\u7f85\u62fc\u97f3\u7684\u6a5f\u5668\u7ffb\u8b6f\u6a21\u7d44\u6700\u70ba\u91cd\u8981\uff0c\u56e0\u70ba\u82e5\u7ffb\u8b6f\u7684\u6b63\u78ba\u7387\u4e0d\u9ad8\uff0c\u5408\u6210\u7aef \u8a08\u7b97\u7684\u5e8f\u5217\u5f0f\u905e\u8ff4\u7db2\u8def\u67b6\u69cb\uff0csample \u9700\u8981\u4e00\u500b\u63a5\u8457\u4e00\u500b\u7167\u524d\u5f8c\u9806\u5e8f\u7522\u751f\u3002\u9664\u8a08\u7b97\u91cf\u76f8\u7576</td></tr><tr><td colspan=\"2\">\u6709\u518d\u597d\u7684\u97f3\u8272\u54c1\u8cea\u548c\u5408\u6210\u901f\u5ea6\u90fd\u662f\u5f92\u52de\u3002 \u5927\u5916\uff0c\u4e5f\u4e0d\u6613\u5e73\u884c\u5316\uff0c\u5c0e\u81f4\u8a9e\u97f3\u751f\u6210\u901f\u5ea6\u975e\u5e38\u6162\uff0c\u5e7e\u4e4e\u7121\u6cd5\u7528\u4e00\u822c\u7684 GPU \u8a2d\u5099\u9054\u5230</td></tr><tr><td colspan=\"2\">\u8f03\u65e9\u671f\u7684\u6a5f\u5668\u7ffb\u8b6f\u65b9\u6cd5\uff0c\u6709\u57fa\u65bc\u898f\u5247\u7684\u5b57\u5c0d\u5b57\u6a5f\u5668\u7ffb\u8b6f(RBMT)\uff0c\u57fa\u65bc\u7bc4\u4f8b\u7684\u53e5\u5c0d\u53e5 real-time \u7684\u6548\u80fd\u8981\u6c42\u3002</td></tr><tr><td colspan=\"2\">\u6a5f\u5668\u7ffb\u8b6f(EBMT)\uff0c\u4ee5\u53ca\u7d71\u8a08\u6a5f\u5668\u7ffb\u8b6f(SMT) (\"\u6a5f\u5668\u7ffb\u8b6f\uff0c\" 2020)\u3002\u8a5e\u5c0d\u8a5e\u7684\u898f\u5247\u6cd5\u5373\u70ba \u56e0\u6b64\uff0c\u76ee\u524d\u8a9e\u97f3\u5408\u6210\u7814\u7a76\u4e3b\u8981\u662f\u8981\u89e3\u6c7a\u5408\u6210\u901f\u5ea6\u554f\u984c\u3002\u4f8b\u5982 Wave-RNN \u8207 WaveGlow\u3002</td></tr><tr><td colspan=\"2\">\u5c07\u4e00\u500b\u4e2d\u6587\u8a5e\uff0c\u4f9d\u64da\u898f\u5247\u8207\u53f0\u83ef\u5e73\u884c\u8fad\u5178\uff0c\u5c0d\u7167\u5230\u4e00\u500b\u53f0\u8a9e\u62fc\u97f3\u7684\u7ffb\u8b6f\u6cd5\uff0c\u9069\u7528\u65bc\u53ea\u6ce8 \u5176\u4e2d\uff0cNVIDIA \u63d0\u51fa\u7684 WaveGlow \u8ddf WaveNet \u76f8\u6bd4\uff0c\u53ef\u4ee5\u907f\u958b\u905e\u8ff4\u7db2\u8def\u67b6\u69cb\u8a08\u7b97\u91cf\u5927\uff0c</td></tr><tr><td colspan=\"2\">\u91cd\u55ae\u8a5e\u7684\u975e\u5b8c\u6574\u53e5\u5b50\u4e4b\u7ffb\u8b6f\uff0c\u4f46\u7ffb\u8b6f\u51fa\u4f86\u7684\u53f0\u8a9e\u6587\u6cd5\u53ef\u80fd\u4e0d\u6b63\u78ba\u3002\u53e5\u5c0d\u53e5\u7684\u7bc4\u4f8b\u6cd5\u70ba\u4e00 \u4e14\u4e0d\u6613\u5e73\u884c\u5316\u7684\u554f\u984c\uff0c\u5408\u6210\u6240\u9700\u6642\u9593\u6bd4\u5927\u5e45\u6e1b\u5c11\uff0c\u7d04\u70ba 1:400\uff0c\u82e5\u662f\u5408\u6210\u7d04 10 \u79d2\u4ee5\u4e0b\u7684</td></tr><tr><td colspan=\"2\">\u6574\u4e32\u4e2d\u6587\u53e5\u5b50\u5c0d\u7167\u5230\u4e00\u6574\u4e32\u53f0\u8a9e\u7684\u53f0\u7f85\u62fc\u97f3\uff0c\u53ef\u9069\u7528\u65bc\u53e5\u5b50\u7684\u7ffb\u8b6f\uff0c\u4e5f\u8f03\u80fd\u8003\u616e\u6587\u6cd5\u5dee \u8a9e\u97f3\uff0c\u5927\u5e45\u6e1b\u5c11\u7684\u5408\u6210\u6642\u9593\u5df2\u7d93\u5e7e\u4e4e\u63a5\u8fd1\u9ad4\u611f\u7684\u5373\u6642\u5408\u6210\uff0c\u4e14\u5176\u516c\u958b\u7684\u5e73\u5747\u610f\u898b\u5f97\u5206</td></tr><tr><td colspan=\"2\">\u7570\u3002\u4f46\u6b64\u6cd5\u5e38\u9700\u4f9d\u8cf4\u8a9e\u6599\u5eab\u4e2d\u53f0\u83ef\u5e73\u884c\u53e5\u5b50\u7684\u591a\u6a23\u5316\u548c\u6578\u91cf\uff0c\u5982\u8981\u7ffb\u8b6f\u5f9e\u672a\u51fa\u73fe\u65bc\u8a9e\u6599 (MOS)\u6e2c\u8a66\u4e5f\u8868\u660e\uff0cWaveGlow \u7684\u97f3\u8cea\u4e5f\u4e0d\u905c\u65bc WaveNet\u3002</td></tr><tr><td colspan=\"2\">\u5eab\u4e2d\u7684\u53e5\u5b50\uff0c\u901a\u5e38\u6703\u8f03\u70ba\u56f0\u96e3\u3002\u7d71\u8a08\u6cd5\u76ee\u524d\u70ba\u975e\u9650\u5b9a\u9818\u57df\u6a5f\u5668\u7ffb\u8b6f\u4e2d\u6027\u80fd\u8f03\u4f73\u7684\u4e00\u7a2e\u65b9 \u56e0\u6b64\uff0c\u57fa\u65bc\u4ee5\u4e0a\u8a0e\u8ad6\uff0c\u6211\u5011\u5c07\u4f7f\u7528 sequence-to-sequence + Tacotron2 + WaveGlow \u7b49</td></tr><tr><td colspan=\"2\">\u6cd5\uff0c\u901a\u904e\u5c0d\u5927\u91cf\u7684\u53f0\u83ef\u5e73\u884c\u8a9e\u6599\u9032\u884c\u7d71\u8a08\u5206\u6790\uff0c\u69cb\u5efa\u7d71\u8a08\u7ffb\u8b6f\u6a21\u578b\u4e26\u9032\u884c\u7ffb\u8b6f\uff0c\u5df2\u7d93\u53ef \u6a21\u578b\u4f86\u5be6\u73fe\u9ad8\u54c1\u8cea\u4e14\u5373\u6642\u4e4b\u53f0\u8a9e\u8a9e\u97f3\u5408\u6210\u3002\u5176\u67b6\u69cb\u70ba\u4f7f\u7528\u8005\u8f38\u5165\u7684\u4e2d\u6587\u6587\u672c\u900f\u904e C2T</td></tr><tr><td colspan=\"2\">\u4ee5\u878d\u5408\u6587\u53e5\u4e2d\u8a9e\u6cd5\u7b49\u4fe1\u606f\u9032\u4e00\u6b65\u63d0\u9ad8\u7ffb\u8b6f\u7684\u7cbe\u78ba\u6027\u3002 (Ott, Edunov, Grangier & Auli, 2018)\u8f49\u70ba\u53f0\u7f85\u62fc\u97f3\uff0c\u518d\u900f\u904e Tacotron2 (Shen et al., 2018)</td></tr><tr><td colspan=\"2\">\u4f8b\u5982\uff0c\u4ea4\u5927\u9673\u4fe1\u5b8f(Kuo, Wang & Chen, 2004) (\u8d99\u826f\u57fa\uff0c2012)\uff0c\u4e2d\u8208\u4f59\u660e\u8208(\u6f58\u80fd\u714c\u3001 \u5c07\u53f0\u7f85\u62fc\u97f3\u8f49\u70ba\u983b\u8b5c\uff0c\u6700\u5f8c\u900f\u904e WaveGlow (Prenger, Valle & Catanzaro, 2018)\u5c07\u983b\u8b5c\u5408\u6210</td></tr><tr><td colspan=\"2\">\u4f59\u660e\u8208\u3001\u8a31\u66f8\u8c6a\uff0c2011)\u8207\u53f0\u5927\u9673\u4fe1\u5e0c(Lin & Chen, 1999)\u7b49\u8001\u5e2b\u8207\u90fd\u66fe\u9032\u884c\u904e\u4e2d\u6587\u6587\u5b57\u8f49 \u51fa\u53f0\u8a9e\u8a9e\u97f3\uff0c\u5982\u5716 1 \u6240\u793a\u3002</td></tr><tr><td colspan=\"2\">\u6210\u53f0\u7f85\u62fc\u97f3\u6a5f\u5668\u7ffb\u8b6f\u7684\u76f8\u95dc\u7814\u7a76\u3002\u5176\u4e2d\uff0c\u9673\u4fe1\u5e0c\u8001\u5e2b\u66fe\u5728 1999 \u5e74\uff0c\u5c31\u767c\u5c55\u51fa\u4e00\u500b\u57fa\u65bc\u8fad</td></tr><tr><td colspan=\"2\">\u5178\u7ffb\u8b6f\uff0c\u5177\u6709\u8a9e\u97f3\u5408\u6210\u529f\u80fd\u7684 Mandarin to Taiwanese Min Nan Machine Translation</td></tr><tr><td colspan=\"2\">System(\u76ee\u524d\u5df2\u7d42\u6b62\u7dad\u8b77)\u3002\u800c\u4e14\u610f\u50b3\u79d1\u6280\u4e5f\u63a1\u7528\u7d71\u8a08\u65b9\u6cd5\u8a13\u7df4\u51fa\u4e00\u5957\u7db2\u9801\u7248\u672c\u7684\u4e2d\u6587\u8f49</td></tr><tr><td colspan=\"2\">\u53f0\u8a9e\u6a5f\u5668\u7ffb\u8b6f 1 \u3002\u4e0d\u904e\uff0c\u6b64\u7a2e\u6a5f\u5668\u7ffb\u8b6f\u6a21\u7d44\uff0c\u9084\u9700\u8981\u5148\u6709\u4e00\u500b\u83ef\u6587\u65b7\u8a5e\u8207 POS \u5256\u6790\u5668(\u81ea</td></tr><tr><td colspan=\"2\">\u7136\u8a9e\u8a00\u5256\u6790\u5668\uff0cNLP parser)\uff0c\u624d\u80fd\u9806\u5229\u9032\u884c\u5f8c\u7e8c\u7684\u6a5f\u5668\u7ffb\u8b6f\u7a0b\u5e8f\u3002\u4f46 NLP parser \u672c\u8eab\u5c31</td></tr><tr><td colspan=\"2\">\u5df2\u7d93\u662f\u4e00\u500b\u96e3\u89e3\u7684\u554f\u984c\uff0c\u800c\u4e14\u901a\u5e38\u6703\u6709\u5927\u7d04 5%\u7684\u5206\u6790\u932f\u8aa4(\u5305\u62ec\u65b7\u8a5e\u8207 POS \u6a19\u8a18)\u3002\u82e5\u9084</td></tr><tr><td colspan=\"2\">\u662f\u4f7f\u7528\u6b64\u50b3\u7d71\u5169\u968e\u6bb5\u67b6\u69cb\uff0c\u5c31\u6703\u8b93\u524d\u7d1a\u7522\u751f\u7684\u932f\u8aa4\uff0c\u9023\u5e36\u5c0e\u81f4\u5f8c\u9762\u7684\u7ffb\u8b6f\u8207\u8a9e\u97f3\u5408\u6210\u932f</td></tr><tr><td>\u8aa4\uff0c\u800c\u4e14\u5f8c\u7d1a\u53ea\u80fd\u63a5\u53d7\uff0c\u7121\u6cd5\u518d\u52a0\u4ee5\u633d\u6551\u3002</td><td/></tr><tr><td colspan=\"2\">\u800c\u8fd1\u5e74\u4f86\u4e3b\u6d41\u7684\u6a5f\u5668\u7ffb\u8b6f\u65b9\u6cd5\u70ba\u985e\u795e\u7d93\u7db2\u8def\u6a5f\u5668\u7ffb\u8b6f(NMT)\uff0c\u9867\u540d\u601d\u7fa9\u4f7f\u7528\u985e\u795e\u7d93</td></tr><tr><td colspan=\"2\">\u7db2\u8def(Neural Network)\u4f86\u505a\u6a5f\u5668\u7ffb\u8b6f\uff0c\u5176\u901a\u5e38\u662f\u57fa\u65bc sequence-to-sequence \u6a21\u578b\uff0c\u4f7f\u7528</td></tr><tr><td colspan=\"2\">34 \u5e74\uff0c\u570b\u6c11\u653f\u5e9c\u63a5\u7ba1\u81fa\u7063\u5f8c\u6975\u529b encoder-decoder \u67b6\u69cb\u4f86\u5b78\u7fd2\u8f38\u5165\u4f86\u6e90\u8a9e\u8a00\u8207\u8f38\u51fa\u76ee\u6a19\u8a9e\u8a00\u9593\u7684\u5c0d\u61c9\u95dc\u4fc2\u3002NMT \u5c24\u5176\u5e38\u4f7f</td></tr><tr><td colspan=\"2\">\u63a8\u884c\u7684\u300c\u570b\u8a9e\u904b\u52d5\u300d\uff0c\u4f7f\u5f97\u7576\u6642\u7684\u5b78\u6821\u7981\u6b62\u5404\u7701\u65b9\u8a00\u53ca\u539f\u4f4f\u6c11\u8a9e\uff0c\u4e26\u56b4\u683c\u63a8\u884c\u300c\u570b\u8a9e\u6559 \u7528 CNN \u6216\u662f RNN\uff0c\u4f86\u5b78\u7fd2\u81ea\u7136\u8a9e\u8a00\u9019\u7a2e\u5177\u6709\u6642\u9593\u9806\u5e8f\u7684\u5e8f\u5217\u6578\u64da(Sequence Data)\u7684\u95dc</td></tr><tr><td colspan=\"2\">\u80b2\u300d(\u674e\u739f\u9038\u3001\u674e\u7950\u8431\u3001\u5468\u695a\uff0c2017)\u3002\u6642\u81f3\u4eca\u65e5\uff0c\u4eba\u6c11\u751f\u6d3b\u4e2d\u5927\u591a\u8aaa\u570b\u8a9e\u70ba\u4e3b\uff0c\u5c0e\u81f4\u73fe \u5716 1</td></tr><tr><td colspan=\"2\">\u4ee3\u4eba\u719f\u6089\u53f0\u8a9e\u7684\u4eba\u6578\u8d8a\u4f86\u8d8a\u5c11\uff0c\u5c24\u5176\u662f\u5e74\u8f15\u4eba\uff0c\u5927\u90e8\u5206\u61c2\u7684\u53f0\u8a9e\u8a5e\u5f59\u4e0d\u591a\uff0c\u8b1b\u5f97\u4e5f\u4e0d\u751a</td></tr></table>", | |
| "html": null | |
| }, | |
| "TABREF1": { | |
| "num": null, | |
| "text": "\u6b21\u3002 \u76f8\u6bd4\u4e4b\u4e0b\uff0cCNN \u6539\u5584\u4e86\u4ee5\u4e0a\u554f\u984c\uff0c\u9664\u4e86\u80fd\u5920\u4e26\u884c\u8655\u7406\u6578\u64da\uff0cPosition Embedding \u6642 \u8f38\u5165\u9664\u4e86\u8a5e\u5411\u91cf\u9084\u52a0\u5165\u4f4d\u7f6e\u5411\u91cf\uff0c\u4e14 CNN \u70ba\u5c64\u7d1a\u7d50\u69cb\uff0c\u53ef\u9867\u616e\u5230\u6574\u6bb5\u6587\u53e5\u7684\u6bcf\u4e00\u500b\u55ae \u8a5e\uff0c\u8f03\u5e95\u5c64 CNN \u6355\u6349\u9593\u9694\u8f03\u8fd1\u7684\u8a5e\u4e4b\u9593\u7684\u4f9d\u8cf4\u95dc\u4fc2\uff0c\u8f03\u9ad8\u5c64 CNN \u5247\u6355\u6349\u9593\u9694\u8f03\u9060\u7684\u8a5e \u4e4b\u9593\u7684\u4f9d\u8cf4\u95dc\u4fc2\u3002CNN \u5728 encoder \u7aef\uff0c\u4ee5 GLU \u4f5c\u70ba\u975e\u7dda\u6027\u55ae\u5143\uff0c\u8f38\u5165\u8207\u8f38\u51fa\u76f8\u52a0\u5f8c\uff0c\u624d \u8f38\u5165\u5230\u4e0b\u4e00\u5c64\u7db2\u7d61\u4e2d\uff0c\u5728 decoder \u7aef\uff0c\u6709 multi-hop attention \u6a5f\u5236\uff0cencoder \u7aef\u7684\u8f38\u51fa\u9032\u884c \u52a0\u6b0a\u6642\uff0c\u6703\u8003\u616e\u539f\u59cb\u7684\u8f38\u5165\u5411\u91cf\uff0c\u5728\u6bcf\u4e00\u500b\u5377\u7a4d\u5c64\u90fd\u6703\u9032\u884c attention \u7684\u64cd\u4f5c\uff0c\u4f7f\u5f97\u6a21\u578b \u5728\u5f97\u5230\u4e0b\u4e00\u500b attention \u6642\uff0c\u80fd\u5920\u8003\u616e\u5230\u4e4b\u524d\u7684\u5df2\u7d93 attention \u904e\u7684\u8a5e\u3002\u5f9e IBM Research \u767c \u8868\u7684\u7814\u7a76\u8ad6\u6587\"Comparative Study of CNN and RNN for Natural Language Processing\" (Yin, Kann, Yu & Sch\u00fctze, 2017)\uff0c\u4e5f\u53ef\u4ee5\u770b\u51fa\u5728\u8655\u7406\u53e5\u5b50\u914d\u5c0d\u7684\u4efb\u52d9\u4e0a\uff0c\u6bd4\u8d77 RNN \u7684 GRU\uff0c LSTM \u7b49\u6a21\u578b\uff0cCNN \u64c1\u6709\u4e00\u5b9a\u7684\u512a\u52e2\u3002 \u8981\u8a13\u7df4\u4e00\u500b C2T \u6a21\u578b\uff0c\u5fc5\u9808\u5148\u6e96\u5099\u597d\u53f0\u83ef\u5e73\u884c\u8a9e\u6599\u4ee5\u53ca\u53f0\u83ef\u5e73\u884c\u8fad\u5178\u3002\u672c\u6a21\u578b\u4f7f\u7528\u7684\u8a9e \u6599\uff0c\u70ba\u4e2d\u7814\u9662\u8cc7\u8a0a\u6240\u9673\u5b5f\u5f70\u8001\u5e2b\u8a08\u756b\u5167\u7684 iCorpus\uff0c\u6b64\u8a9e\u6599\u5eab\u6536\u96c6 3266 \u7bc7\u65b0\u805e\uff0c\u5171 83544 \u53e5\u3002\u7b97\u6a19\u9ede\u7b26\u865f\uff0c\u53f0\u8a9e 504037 \u8a5e\u30011030671 \u5b57\uff0c\u83ef\u8a9e 501202 \u8a5e\u30011028218 \u5b57\u3002\u4ee5\u4e0b\u70ba iCorpus \u7684\u90e8\u4efd\u6587\u7ae0\u5167\u5bb9\uff0c\u5982\u5716 2 \u6240\u793a\u3002 \u4e2d\u6587\u8f49\u53f0\u7f85\u62fc\u97f3\u70ba\u4e00\u7a2e\u6a5f\u5668\u7ffb\u8b6f\uff0c\u505a\u6cd5\u70ba\u5c07\u4e2d\u6587\u6587\u5b57\u5e8f\u5217\u8f49\u53f0\u7f85\u62fc\u97f3\u5e8f\u5217\uff0c\u5229\u7528\u57fa\u65bc sequence-to-sequence \u6df1\u5ea6\u985e\u795e\u7d93\u7db2\u8def\u67b6\u69cb\uff0c\u5b78\u7fd2\u5982\u4f55\u9032\u884c\u8f49\u63db\u3002\u6b64 C2T \u63a1\u7528\u7db2\u8def\u4e0a\u958b\u6e90 \u7684 fairseq \u67b6\u69cb(Ott et al., 2018)\u9032\u884c\u8a13\u7df4\uff0c\u5176\u5305\u62ec\u4e00 encoder \u524d\u7aef\u8207\u4e00 decoder \u5f8c\u7aef\u3002\u524d\u7aef encoder \u8ca0\u8cac\u63a5\u6536\u8f38\u5165\u4e2d\u6587\u6587\u5b57\u5e8f\u5217\uff0c\u5206\u6790\u5176\u8a9e\u610f\u4e26\u64f7\u53d6\u51fa\u6587\u8108\u8cc7\u8a0a\u5411\u91cf\u3002\u5f8c\u7aef decoder \u5728\u6587\u8108\u8cc7\u8a0a\u5411\u91cf\u4e4b\u9593\u52a0\u5165 attention \u4e4b\u6a5f\u5236\u8207 Convolutional Neural Network \u4e4b\u8a13\u7df4\u6a21\u578b\u4e0b \u6bcf\u500b encoder \u6b0a\u91cd\uff0c\u5229\u7528\u4e00\u4e2d\u6587\u5c0d\u61c9\u53f0\u8a9e\u62fc\u97f3\u5e73\u884c\u8a9e\u6599\u5eab(iCorpus)\uff0c\u518d\u52a0\u4e0a\u53f0\u83ef\u5e73\u884c\u8fad\u5178 \u9032\u884c\u8a13\u7df4\uff0c\u4ee5\u6b64\u5f97\u5230\u6700\u4f73\u7684\u8f49\u8b6f\u53f0\u7f85\u62fc\u97f3\u5e8f\u5217\uff0c\u5982\u5716 4 \u6240\u793a\u3002 \u5716", | |
| "type_str": "table", | |
| "content": "<table><tr><td>74 82</td><td>\u57fa\u65bc\u6df1\u5ea6\u5b78\u7fd2\u4e4b\u4e2d\u6587\u6587\u5b57\u8f49\u53f0\u8a9e\u8a9e\u97f3\u5408\u6210\u7cfb\u7d71\u521d\u6b65\u63a2\u8a0e \u57fa\u65bc\u6df1\u5ea6\u5b78\u7fd2\u4e4b\u4e2d\u6587\u6587\u5b57\u8f49\u53f0\u8a9e\u8a9e\u97f3\u5408\u6210\u7cfb\u7d71\u521d\u6b65\u63a2\u8a0e \u57fa\u65bc\u6df1\u5ea6\u5b78\u7fd2\u4e4b\u4e2d\u6587\u6587\u5b57\u8f49\u53f0\u8a9e\u8a9e\u97f3\u5408\u6210\u7cfb\u7d71\u521d\u6b65\u63a2\u8a0e</td><td>\u8a31\u6587\u6f22 \u7b49 75 \u8a31\u6587\u6f22 \u7b49 77 \u8a31\u6587\u6f22 \u7b49 \u8a31\u6587\u6f22 \u7b49 81 \u8a31\u6587\u6f22 \u7b49</td></tr><tr><td colspan=\"3\">\u5716 2. iCorpus \u83ef\u53f0\u5e73\u884c\u8a9e\u6599\u5eab [Figure 2. iCorpus Chinese-TLPA parallel corpus] \u8fad\u5178\u65b9\u9762\uff0c\u5247\u662f\u900f\u904e\"ChhoeTaigi \u627e\u53f0\u8a9e \"\u7db2\u7ad9\u4e4b\u53f0\u8a9e\u5b57\u8a5e\u8cc7\u6599\u5eab\uff0c\u8490\u96c6\u5230\u7684 9 \u672c\u4e0d\u540c\u53f0\u8a9e\u8fad\u5178\u5408\u4f75\u6210\u7684\u53f0\u83ef\u5e73\u884c\u8fad\u5178\uff0c\u5404\u8fad\u5178\u7d71\u8a08\u7684\u83ef\u8a9e\u8a5e\u6578\uff0c\u5982\u5716 3 \u6240\u793a\u3002 \u5716 3\u56e0\u5404\u500b\u8fad\u5178\u7531\u4e0d\u540c\u4f5c\u8005\u6240\u64b0\u5beb\uff0c\u683c\u5f0f\u4e26\u975e\u4e00\u81f4\uff0c\u70ba\u80fd\u5728\u5408\u6210\u7cfb\u7d71\u4e2d\u4f7f\u7528\uff0c\u9700\u8981\u7d93\u904e \u591a\u6b21\u6821\u6b63\uff0c\u6821\u6b63\u76ee\u7684\u4e3b\u8981\u662f\u5c07\u53f0\u8a9e\u8a5e\u7ffb\u8b6f\u6210\u83ef\u8a9e\u8a5e\uff0c\u83ef\u8a9e\u8a5e\u5373\u70ba\u5e73\u5e38\u751f\u6d3b\u4e2d\u53e3\u8a9e\u7684\u6163\u7528 \u6587\u5b57\uff0c\u6aa2\u67e5\u83ef\u8a9e\u8a5e\u7684\u610f\u601d\u8207\u683c\u5f0f\u662f\u5426\u6b63\u78ba\uff0c\u4e26\u4e14\u8207\u4e4b\u5c0d\u61c9\u7684\u53f0\u7f85\u62fc\u97f3\u662f\u5426\u70ba\u4e00\u5c0d\u4e00\u3002\u53f0 \u97f3\u9577\u7b49\uff0c\u7136\u5f8c\u628a\u8aaa\u8a71\u7684\u8072\u8abf\uff0c\u8a9e\u6c23\uff0c\u505c\u9813\u65b9\u5f0f\uff0c\u767c\u97f3\u9577\u77ed\u8f49\u63db\u6210\u97fb\u5f8b\u53c3\u6578(\u6731\u5b5d\u570b\uff0c2005)\u3002 \u9078\u64c7\u51fa\u5408\u9069\u7684\u8072\u5b78\u53c3\u6578\uff0c\u7136\u5f8c\u6839\u64da\u5728\u97fb\u5f8b\u6a21\u578b\u4e2d\u5f97\u5230\u7684\u97fb\u5f8b\u53c3\u6578\uff0c\u900f\u904e\u8a9e\u97f3\u5408\u6210\u6f14\u7b97\u6cd5 \u6578 88881\uff0c\u53f0\u8a9e\u8a5e\u689d\u6578 153132\u3002 \u6578\u63a5\u8457\u9001\u5165\u97fb\u5f8b\u7522\u751f\u5668\u4f86\u7522\u751f\u6587\u672c\u88e1\u6bcf\u500b\u97f3\u7bc0\u7684\u5c0d\u61c9\u97fb\u5f8b\u8a0a\u606f\uff0c\u5305\u542b\u57fa\u983b\u8ecc\u8de1\uff0c\u97f3\u91cf\uff0c \u591a\u6b21\u51fd\u5f0f\u8f49\u63db\uff0c\u9010\u6b65\u8f49\u63db\u6210\u771f\u5be6\u8a9e\u97f3\u6ce2\u5f62\u8a0a\u865f x\u3002\u6700\u5f8c\u6839\u64da\u9700\u8981\u767c\u51fa\u7684\u8072\u97f3\u5f9e\u8cc7\u6599\u5eab\u4e2d \u53f0\u8a9e\u767c\u97f3\u3002\u56e0\u6b64\u610f\u50b3\u79d1\u6280\u6b64\u6a5f\u5668\u7ffb\u8b6f\u7684\u300c\u4e2d\u6587\u300d\u8f49\u53f0\u8a9e\u7ffb\u8b6f\uff0c\u9084\u662f\u6709\u6240\u727d\u5f37\u3002 \u7f85\u62fc\u97f3\u4e5f\u9700\u6aa2\u67e5\uff0c\u5254\u9664\u591a\u9918\u4e4b\u610f\u601d\u6216\u7b26\u865f\u3002\u6700\u7d42\u53ef\u4f7f\u7528\u7684\u53f0\u83ef\u8a5e\u689d\u6578 225965\uff0c\u83ef\u8a9e\u8a5e\u689d \u5408\u4f5c\u7522\u88fd\uff0c\u70ba\u4e00\u6709\u5927\u5b78\u6559\u80b2\u7a0b\u5ea6\u4e4b 34 \u6b72\u7537\u6027\u9304\u88fd\uff0c\u53f0\u8a9e\u8154\u8abf\u504f\u6f33\u5dde\u8154\uff0c\u97f3\u6a94\u7b46\u6578 9625 \u7b46\uff0c\u9577\u5ea6\u7d04 10.4 \u5c0f\u6642\u3002Tacotron2 \u70ba\u4e00 end-to-end \u65b9\u5f0f\u505a\u8a13\u7df4\u8207\u63a8\u8ad6\u4e4b\u6a21\u578b\uff0c\u4f7f\u7528\u67b6\u69cb\u70ba encoder-decoder + Location Sensitive Attention\u3002\u505a\u6cd5\u70ba\u5c07\u7ffb\u8b6f\u5b8c\u6210\u7684\u53f0\u7f85\u62fc\u97f3\u8f38\u5165\u5f8c\uff0c\u985e \u795e\u7d93\u7db2\u8def\u9032\u884c\u6587\u672c\u5206\u6790\uff0c\u628a\u8a9e\u6cd5\u8207\u8a9e\u610f\u8f49\u6210\u8a9e\u8a00\u7279\u5fb5\u53c3\u6578\uff0c\u8b93\u7cfb\u7d71\u77e5\u9053\u6587\u672c\u4e2d\u54ea\u4e9b\u662f\u8a5e\uff0c \u54ea\u4e9b\u662f\u53e5\u5b50\uff0c\u767c\u4ec0\u9ebc\u97f3\uff0c\u600e\u9ebc\u767c\u97f3\uff0c\u767c\u97f3\u6642\u5230\u54ea\u61c9\u8a72\u505c\u9813\uff0c\u505c\u9813\u591a\u9577\u7b49\u7b49\u3002\u8a9e\u8a00\u7279\u5fb5\u53c3 \u6700\u5f8c\u8f38\u51fa\u6885\u723e\u983b\u8b5c\u5716\uff0c\u518d\u7d93\u7531\u5c0d\u9f4a\u9054\u6210\u53f0\u7f85\u62fc\u97f3\u8207\u983b\u8b5c\u4e00\u5c0d\u4e00\u7684\u5c0d\u61c9\uff0c\u5982\u5716 5 \u6240\u793a\u3002 \u5716 5. \u6b64\u8655\u6a21\u578b\u8a13\u7df4\u7528\u4e4b\u8a9e\u6599\u5eab\u4f86\u6e90\u540c Tacotron2\uff0cWaveGlow \u662f\u4e00\u7a2e flow-based generative networks\uff0c\u7d50\u5408 Glow \u548c WaveNet \u7684\u539f\u7406\uff0c\u900f\u904e\u8f38\u5165\u97f3\u6a94\u8207\u5176\u751f\u6210\u4e4b\u983b\u8b5c\uff0c\u50c5\u4f7f\u7528\u55ae\u500b \u7db2\u8def\u8207\u55ae\u500b\u640d\u5931\u51fd\u5f0f\u9032\u884c\u8a13\u7df4\uff0c\u751f\u6210\u4e00\u9ad8\u65af\u5206\u5e03 z\uff0c\u5408\u6210\u6642\u53ea\u9700\u900f\u904e z \u8207\u983b\u8b5c\u5c31\u53ef\u5373\u6642 \u5408\u6210\u9ad8\u54c1\u8cea\u8a9e\u97f3\uff0c\u5982\u5716 6 \u6240\u793a\u3002 \u5716 6. \u8a13\u7df4\u6642\u4f9d\u64da\u57fa\u65bc\u6a5f\u7387\u4e4b cost function \u5c0e\u5f15\uff0c\u591a\u6b21\u5229\u7528\u51fd\u5f0f\u8f49\u63db\uff0c\u9010\u6b65\u5b78\u7fd2\u5982\u4f55\u5c07\u771f \u5be6\u8a9e\u97f3\u6ce2\u5f62\u8a0a\u865f x \u6295\u5c04\u5230\u4e00\u5177\u9ad8\u65af\u5206\u4f48\u4e4b\u96b1\u85cf\u8b8a\u6578 z \u7684\u7a7a\u9593\u3002\u4e26\u5728\u8a13\u7df4\u6642\u9650\u5236 mapping \u51fd\u5f0f\u70ba\u53ef\u9006\u51fd\u5f0f\uff0cWaveGlow \u5728\u751f\u6210\u6ce2\u578b\u5716\u6642\u5373\u53ef\u4f9d\u64da\u96b1\u85cf\u8b8a\u6578 z \u7a7a\u9593\u53d6\u6a23\u7684\u7d50\u679c\uff0c\u7d93 \u7522\u751f\u8a9e\u97f3(\u6731\u5b5d\u570b\uff0c2005)\u3002\u5982\u5716 7 \u6240\u793a\u3002 \u5716 7. \u6b64\u57fa\u65bc\u6df1\u5ea6\u5b78\u7fd2\u4e4b\u4e2d\u6587\u8f49\u53f0\u8a9e\u8a9e\u97f3\u5408\u6210\u7cfb\u7d71\uff0c\u5df2\u67b6\u8a2d\u7db2\u9801\u7248\u672c\u4ee5\u4f9b\u4f7f\u7528\uff0c\u9023\u7d50\u7db2\u5740\u70ba http://140.115.54.90:31810/\u3002\u4f7f\u7528\u8005\u8f38\u5165\u4e2d\u6587\u6587\u5b57\uff0c\u6309\u4e0b\u5408\u6210\u6309\u9215\u5c31\u80fd\u64a5\u653e\u5c0d\u61c9\u7684\u53f0\u8a9e\u8a9e \u97f3\uff0c\u4e26\u80fd\u4e00\u4f75\u986f\u793a\u51fa\u7ffb\u8b6f\u904e\u5f8c\u7684\u53f0\u7f85\u62fc\u97f3\u4f9b\u4f7f\u7528\u8005\u67e5\u8a62\uff0c\u4e14\u8a2d\u8a08\u4e86\u53ef\u8f38\u5165\u53f0\u7f85\u62fc\u97f3\u7684\u6b04 \u4f4d\u8b93\u64c1\u6709\u76f8\u95dc\u53f0\u7f85\u77e5\u8b58\u7684\u4f7f\u7528\u8005\u53ef\u4ee5\u9375\u5165\u4e0d\u540c\u7684\u767c\u97f3\u65b9\u5f0f\u4e26\u5408\u6210\u8a9e\u97f3\u3002\u5716 8 \u5c55\u793a\u70ba\u672c\u6587 \u4e4b\u4f7f\u7528\u8005\u4ecb\u9762\u3002\u7db2\u9801\u4e4b\u7d05\u8272\u5206\u9694\u7dda\u4e0a\u65b9\u4f7f\u7528\u672c\u6587\u4e4b C2T \u6a5f\u5668\u7ffb\u8b6f\uff0c\u4e0b\u65b9\u70ba\u4f5c\u70ba\u6bd4\u8f03\u7528\u4e4b \u610f\u50b3\u79d1\u6280\u7d71\u8a08\u6cd5\u6a5f\u5668\u7ffb\u8b6f\uff0c\u5408\u6210\u7aef\u4e00\u5f8b\u4f7f\u7528\u672c\u6587\u4e4b Tacotron2+WaveGlow\u3002 \u5176\u4e2d\uff0c\u521d\u6b65\u6e2c\u8a66\u7d71\u8a08\u5f0f\u6a5f\u5668\u7ffb\u8b6f\u5f8c\u53ef\u4ee5\u767c\u73fe\uff0c\u7d71\u8a08\u5f0f\u7ffb\u8b6f\u7684\u7d50\u679c\u8f03\u70ba\u63a5\u8fd1\u4e2d\u6587\u6587\u5b57 \u672c\u8eab\u5ff5\u6cd5\u7684\u97f3\u8b6f\uff0c\u4ea6\u5373\u5982\u679c\u8981\u7ffb\u8b6f\u51fa\u4e2d\u6587\u8f49\u53f0\u8a9e\u5f8c\u6b63\u78ba\u7684\u53f0\u7f85\u62fc\u97f3\uff0c\u61c9\u8f38\u5165\u53f0\u6587\u70ba\u4f73\uff0c \u6240\u8b02\u53f0\u6587\u5373\u70ba\u4e2d\u6587\u8f49\u70ba\u95a9\u5357\u8a9e\u7684\u53e6\u4e00\u7a2e\u66f8\u5beb\u8868\u793a\u5f62\u5f0f\u3002\u5982\u4e2d\u6587\u7684\u300c\u73fe\u5728\u662f\u665a\u4e0a\u516b\u9ede\u300d\uff0c \u7d71\u8a08\u5f0f\u6a5f\u5668\u7ffb\u8b6f\u7d50\u679c\u70ba\u300chian7 tsai7 si7 mng2 siong7 peh4 tiam2\u300d \uff0c\u4e0d\u662f\u6b63\u78ba\u7684\u53f0\u8a9e\u767c\u97f3\uff0c \u800c\u53f0\u6587\u7684\u300c\u9019\u99ac\u662f\u6697\u6642\u516b\u9ede\u300d\u7ffb\u8b6f\u5f8c\u7684\u300ctsit4 ma1 si7 am3 si5 peh4 tiam2\u300d\uff0c\u624d\u662f\u6b63\u78ba\u7684 \u5716 \u672c\u6587\u4e4b C2T \u6a21\u578b\u5206\u5225\u4ee5\u4e09\u7a2e\u904b\u884c\u6a21\u5f0f\u9032\u884c\u6548\u679c\u6e2c\u8a66\u3002\u6a21\u5f0f\u4e00\u70ba\u4e2d\u6587\u53e5\u5b50\u900f\u904e\u65b7\u8a5e\u5f8c\uff0c\u5224 \u8a9e\u8a9e\u97f3\u5408\u6210\u5de5\u4f5c\u3002 \u53e5\u4e2d\u6587\u53e5\u5b50\u76f4\u63a5\u8f49\u63db\u70ba\u53f0\u7f85\u62fc\u97f3\uff0c\u540c\u6642\u5b78\u7fd2\u53f0\u8a9e\u4e4b\u65b7\u53e5\u898f\u5247\uff0c\u5982\u8868 3 \u6240\u793a\u3002 \u6240\u793a\u3002\u56e0\u6b64\u672c\u6587\u4e4b C2T \u6700\u7d42\u63a1\u7528\u5c07\u4e2d\u6587\u76f4\u63a5\u8f49\u63db\u70ba\u53f0\u7f85\u62fc\u97f3\u7684\u8f49\u63db\u6cd5\uff0c\u4ee5\u5229\u63a5\u4e0b\u4f86\u7684\u53f0 \u8868 1 \u6240\u793a\uff1b\u6a21\u5f0f\u4e8c\u70ba\u5c07\u6574\u53e5\u4e2d\u6587\u53e5\u5b50\u76f4\u63a5\u8f49\u63db\u70ba\u53f0\u7f85\u62fc\u97f3\uff0c\u5982\u8868 2 \u6240\u793a\uff1b\u6a21\u5f0f\u4e09\u70ba\u5c07\u6574 \u53ca 8.699%\u3002\u7d9c\u4e0a\u6578\u64da\u5f97\u77e5\uff0c\u6a21\u5f0f\u4e8c\u7531\u4e2d\u6587\u53e5\u5b50\u76f4\u63a5\u8f49\u63db\u70ba\u53f0\u7f85\u62fc\u97f3\u7684\u6548\u679c\u6700\u4f73\uff0c\u5982\u5716 9 \u65b7\u5404\u500b\u8a5e\u4e4b\u8a5e\u6027\uff0c\u518d\u5c07\u5404\u8a5e\u8f49\u70ba\u53f0\u7f85\u62fc\u97f3\uff0c\u56e0\u6b64\u53ef\u4ee5\u5f97\u5230\u65b7\u8a5e\u5f8c\u542b\u8a5e\u6027\u4e4b\u53f0\u7f85\u62fc\u97f3\uff0c\u5982 \u8868 1. \u4e2d\u6587\u8f49\u53f0\u7f85\u62fc\u97f3^\u65b7\u8a5e/\u8a5e\u6027 \u4e2d\u6587\u53e5\u5b50 \u5085\u9054\u4ec1\u4eca\u5c07\u57f7\u884c\u5b89\u6a02\u6b7b\uff0c\u537b\u7a81\u7136\u7206\u51fa\u81ea\u5df1 20 \u5e74\u524d\u906d\u7def\u4f86\u9ad4\u80b2\u53f0\u5c01\u6bba\uff0c\u4ed6 \u4e0d\u61c2\u81ea\u5df1\u54ea\u88e1\u5f97\u7f6a\u5230\u96fb\u8996\u53f0\u3002 \u65b7\u8a5e \u5085\u9054\u4ec1 \u4eca \u5c07 \u57f7\u884c \u5b89\u6a02\u6b7b \uff0c \u537b \u7a81\u7136 \u7206\u51fa \u81ea\u5df1 20 \u5e74 \u524d \u906d \u7def\u4f86 \u9ad4\u80b2\u53f0 \u5c01\u6bba \uff0c \u4ed6 \u4e0d \u61c2 \u81ea\u5df1 \u54ea\u88e1 \u5f97\u7f6a\u5230 \u96fb\u8996\u53f0\u3002 \u8a5e\u6027 Nb Nd D VC Na COMMACATEGORY D D VJ Nh Neu Nf Ng P Nb Na VC COMMACATEGORY Nh D VK Nh Ncd VJ Nc PERIODCATEGORY \u8a5e/\u8a5e\u6027 hing5^E/VC an^B/Na lok8^I/Na si2^E/Na , khiok^S/D tut8^B/D jian5^E/D pok8^B/VJ chhut^E/VJ ka^B/Nh ki7^E/Nh ji7^B/Neu tsap8^E/Neu ni5^S/Nf tsing5^S/Ng cho^S/P hu7i^B/Nb la5i^E/Nb the2^B/Na iok8^I/Na tai5^E/Na hong^B/VC sat^E/VC , i^S/Nh bo5^S/D tong2^S/VK ka^B/Nh ki7^E/Nh to2^B/Ncd ui7^E/Ncd tioh8^B/VJ choe7^I/VJ kau3^E/VJ tian7^B/Nc si7^I/Nc tai5^E/Nc . \u8868 2. \u4e2d\u6587\u8f49\u53f0\u7f85\u62fc\u97f3 [Table 2. Chinese to TLPA] \u4e2d\u6587\u53e5\u5b50 \u4e2d\u592e\u6d41\u884c\u75ab\u60c5\u6307\u63ee\u4e2d\u5fc3\uff0c\u4eca\u65e5\u8868\u793a\uff0c\u570b\u5167\u7121\u65b0\u589e\u78ba\u8a3a\u500b\u6848\u3002 \u53f0\u7f85\u62fc\u97f3 Tiong iang liu5 heng5 ek8 cheng5 chi2 hui tiong sim , kin a2 jit8 piau2 si7 , kok lai7 bo5 sin cheng7 chin2 ko3 an3 . \u8868 3. \u4e2d\u6587\u8f49\u53f0\u8a9e\u8a5e [Table 3. Chinese to words of TLPA] \u4e2d\u6587\u53e5\u5b50 \u91cc\u9577\u7684\u8a00\u8ad6\u5728 PTT \u5f15\u767c\u71b1\u8b70\u8a31\u591a\u7db2\u53cb\u7d1b\u7d1b\u7559\u8a00\u3002 \u53f0\u8a9e\u8a5e li2-tiunn2-e5 gian5-lun7 ti7 PTT in2-huat4 jiat8-gi7 tsiann5-tse7 bang7-iu2 hun1-ue7 . \u70ba\u6e2c\u8a66\u4ee5\u4e0a\u4e09\u7a2e C2T \u6a21\u5f0f\u7684\u6548\u80fd\uff0c\u6211\u5011\u4ee5 iCorpus \u53f0\u83ef\u5e73\u884c\u8a9e\u6599\u8207\u8fad\u5178\u9032\u884c\u6e2c\u8a66\u3002 \u5be6\u9a57\u8cc7\u6599\u5eab\u5305\u62ec iCorpus(78821 \u53e5)\u8207\u53f0\u83ef\u8fad\u5178\u5408\u96c6(225965 \u8a5e\u689d)\uff0c\u4e26\u5207\u5206\u6210\u4e09\u500b\u5b50\u96c6\uff0c \u5305\u62ec Train 90%\uff0cValid 5%\uff0cTest 5%\u3002\u7cfb\u7d71\u6548\u80fd\u5247\u4ee5 Perplexity\uff0c\u53ca Word error rate(WER) \u53ca 9.211%\u3002\u4e0d\u8003\u616e\u8072\u8abf\u7684\u60c5\u6cc1\u4e0b\uff0c\u6a21\u5f0f\u4e00\u5230\u6a21\u5f0f\u4e09\u7684 WER \u5206\u5225\u70ba 18.660%\uff0c6.530%\u4ee5 Tacrtron2 \u8207 WaveGlow \u6a21\u578b\u78ba\u5be6\u53ef\u4ee5\u6b63\u78ba\u5408\u6210\u51fa\u63a5\u8fd1\u771f\u4eba\u8072\u97f3\u4e4b\u53f0\u8a9e\u8a9e\u97f3\u3002 \u4f86\u8861\u91cf\u7d50\u679c\u3002\u8003\u616e\u8072\u8abf\u7684\u60c5\u6cc1\u4e0b\uff0c\u6a21\u5f0f\u4e00\u5230\u6a21\u5f0f\u4e09\u7684 WER \u5206\u5225\u70ba 25.265%\uff0c7.102%\u4ee5 \u5716 9. C2T \u932f\u8aa4\u7387\u6bd4\u8f03 [Figure 9. C2T Syllable-level WER] 4.2 Tacotron2+WaveGlow \u5408 \u6210 \u8a9e \u97f3 \u54c1 \u8cea \u5be6 \u9a57 (Tacotron2+WaveGlow synthesized speech quality experiment) \u5c07\u4e8b\u5148\u6e96\u5099\u597d\u7684 15 \u53e5\u5408\u6210\u97f3\u6a94\u653e\u4e0a Google \u8868\u55ae\uff0c\u4e26\u958b\u653e\u4e00\u822c\u4eba\u5c0d\u6bcf\u500b\u53e5\u5b50\u55ae\u7368\u9032\u884c\u8a55 \u5206\u3002\u5ffd\u7565\u8a9e\u97f3\u5167\u5bb9\u7ffb\u8b6f\u932f\u8aa4\u6216\u662f\u8a9e\u610f\u4e0d\u9806\u7b49\u56e0\u7d20\uff0c\u50c5\u6839\u64da\u807d\u5230\u7684\u300c\u54c1\u8cea\u300d\u8a55\u5206 1.0 \u5230 5.0 \u5206\u3002\u6700\u4f4e\u5206\u70ba 1.0 \u5206 \u70ba\u6700\u63a5\u8fd1\u6a5f\u5668\u4eba\u8b1b\u8a71\u7684\u8072\u97f3\uff1b\u6700\u9ad8\u5206\u70ba 5.0 \u5206 \u70ba\u6700\u63a5\u8fd1\u771f\u4eba\u8b1b\u8a71\u7684 \u8072\u97f3\u3002\u8a55\u5206\u5230\u5c0f\u6578\u9ede\u5f8c\u4e00\u4f4d\u3002\u958b\u653e\u8a55\u5206\u6642\u9593\u7d04\u5169\u5929\uff0c\u622a\u6b62\u6642\u6709 20 \u4f4d\u807d\u8005\u8a55\u5206\u3002\u8868 4 \u4e4b S1 \u5230 S15 \u4ee3\u8868 15 \u53e5\u4e2d\u6587\u8a9e\u53e5\u5167\u5bb9\uff0c\u70ba\u4e86\u4e0d\u8b93\u7ffb\u8b6f\u932f\u8aa4\u6216\u662f\u8a9e\u610f\u4e0d\u901a\u9806\u7b49\u56e0\u7d20\u5f71\u97ff\u807d\u8005 \u5c0d\u97f3\u6a94\u54c1\u8cea\u7684\u8a55\u5206\uff0c\u6240\u6709\u53e5\u5b50\u7686\u6709\u900f\u904e\u4eba\u5de5\u6821\u6b63\u53f0\u7f85\u62fc\u97f3\uff0c\u4e14\u9069\u7576\u7684\u52a0\u5165\u65b7\u8a5e\u6a19\u8a18\u4f7f\u53e5 \u5b50\u6574\u9ad4\u8a9e\u6c23\u901a\u9806\uff0c\u8b93\u807d\u8005\u53ef\u5c08\u5fc3\u91dd\u5c0d\u97f3\u6a94\u300c\u54c1\u8cea\u300d\u8a55\u5206\u3002\u5be6\u9a57\u6700\u7d42\u7d50\u679c\u4e4b\u76d2\u9b1a\u5716\u5982\u5716 10 \u6240\u793a\u3002 \u7e3d\u5171 300 \u7b46\u8a55\u5206\u8cc7\u6599\uff0c\u6700\u7d42\u5e73\u5747\u610f\u898b\u5f97\u5206(mean opinion score\uff0cMOS)\u7d04\u70ba 4.30 \u5206\u3002 \u6b64\u5be6\u9a57\u6240\u4f7f\u7528\u4e4b 15 \u500b\u97f3\u6a94\u7686\u7531\u7cfb\u7d71\u96db\u578b\u5c55\u793a\u7db2\u9801\u5408\u6210\uff0c\u7531\u6b64\u9ad8\u5f97\u5206\u53ef\u77e5\u672c\u6587\u6240\u7528\u7684 \u8868 4. \u6e2c\u8a66\u4e2d\u6587\u8a9e\u53e5\u5167\u5bb9 [Table 4. Chinese sentence content for experiment] S1 \u5927\u5bb6\u597d\uff0c\u6211\u662f\u6703\u8aaa\u53f0\u8a9e\u7684\u6a5f\u5668\u4eba S2 \u8acb\u6839\u64da\u807d\u5230\u8072\u97f3\u7684\u97f3\u8cea\u6253\u5206\u6578 S3 \u4eca\u5929\u4e00\u65e9\u8d77\u4f86\uff0c\u5929\u6c23\u5c31\u975e\u5e38\u708e\u71b1 S4 \u4e00\u5343\u5169\u767e\u4e09\u5341\u56db\u842c\u4e94\u5343\u516d\u767e\u4e03\u5341\u4e5d\u9ede\u96f6\u4e00\u7f8e\u5143 S5 \u97d3\u570b\u745c\u662f\u53f0\u7063\u6b77\u53f2\u4e0a\uff0c\u7b2c\u4e00\u4f4d\u88ab\u7f77\u514d\u7684\u7e23\u5e02\u9996\u9577 S6 \u6b66\u6f22\u80ba\u708e\u7684\u51fa\u73fe\uff0c\u8b93\u5168\u4e16\u754c\u7684\u4eba\u90fd\u958b\u59cb\u6234\u53e3\u7f69 S8 \u6b50\u7f8e\u570b\u5bb6\u5982:\u7f8e\u570b\u3001\u52a0\u62ff\u5927\u3001\u5fb7\u570b\u3001\u6cd5\u570b\u3001\u82f1\u570b\u3001\u897f\u73ed\u7259\u3001\u745e\u5178\u3001\u745e\u58eb\u3001\u632a\u5a01\u3001\u82ac \u862d\u7b49\u7b49 S9 \u9019\u500b\u7334\u6b7b\u56dd\u4ed4\uff0c\u7adf\u7136\u5077\u6041\u7238\u7684\u9322\u53bb\u8cb7\u90a3\u500b\u5783\u573e S10 \u5403\u4e88\u80a5\u80a5\uff0c\u88dd\u4e88\u9318\u9318\uff0c\u88dd\u4e88\u6c34\u6c34\uff0c\u7b49\u9818\u85aa\u6c34 S11 \u6628\u5929\u5730\u9707\u6642\uff0c\u6211\u5011\u5bb6\u7684\u82b1\u74f6\u6389\u4e0b\u4f86\u6454\u7834\u4e86 S12 \u9f9c\u7b11\u9c49\u7121\u5c3e\uff0c\u9c49\u7b11\u9f9c\u7c97\u76ae S13 \u6b61\u8fce\u5149\u81e8\uff0c\u8acb\u554f\u6709\u5e7e\u4f4d S14 \u6709\u98b1\u98a8\u5f9e\u592a\u5e73\u6d0b\u4f86\u7684\u6642\u5019\uff0c\u4e2d\u592e\u5c71\u8108\u5e38\u5e38\u5e6b\u53f0\u7063\u7684\u897f\u90e8\u64cb\u53bb\u5f88\u591a\u707d\u60c5 S15 \u8b1d\u8b1d\u4f60\u4ed8\u51fa\u5bf6\u8cb4\u7684\u6642\u9593\uff0c\u53c3\u52a0\u9019\u6b21\u7684\u8abf\u67e5 4.3 WaveGlow\u8a9e\u97f3\u5408\u6210\u901f\u5ea6\u5be6\u9a57 (WaveGlow speech synthesis speed experiment) \u9078\u7528 WaveGlow \u7576\u4f5c\u5408\u6210\u7aef\u7684\u597d\u8655\u5728\u65bc\u5373\u6642\u7684\u8a9e\u97f3\u5408\u6210\u901f\u5ea6\uff0c\u8868 5 \u70ba\u4e00\u500b\u7c21\u55ae\u7684 WaveGlow \u5408\u6210\u97f3\u6a94\u7684\u901f\u5ea6\u5be6\u9a57\uff0c\u6642\u9593\u7684\u55ae\u4f4d\u70ba\u79d2\u3002 \u8868 5. WaveGlow \u5408\u6210\u901f\u5ea6\u5be6\u9a57 [Table 5. WaveGlow synthesis speed experiment] \u97f3\u6a94\u9577\u5ea6 5.83 4.25 7.48 3.96 9.16 \u5408\u6210\u82b1\u8cbb\u6642\u9593 1.90 1.22 1.83 1.08 2.74 \u7531\u8868 5 \u53ef\u4ee5\u5f97\u77e5\uff0c\u672c\u6587\u4e2d\u7684 WaveGlow \u4e00\u79d2\u7d04\u53ef\u5408\u6210 3.5 \u79d2\u7684\u97f3\u6a94\uff0c\u76f8\u6bd4\u539f\u59cb\u5408\u6210 2.1\u6b64\u8655\u6a21\u578b\u8a13\u7df4\u7528\u4e4b\u8a9e\u6599\u5eab\u4f86\u6e90\uff0c\u70ba\u81fa\u5317\u79d1\u6280\u5927\u5b78\u548c\u674e\u6c5f\u78ba\u53f0\u8a9e\u6587\u6559\u57fa\u91d1\u6703\u4ee5\u53ca\u610f\u50b3\u79d1\u6280 \u53f0\u7f85\u62fc\u97f3\u65b7 poo3^B/Nb tat8^I/Nb jin5^E/Nb kim1^S/Nd tsiong3^S/D tsip4^B/VC S7 \u73fe\u5728\u662f\u665a\u4e0a\u516b\u9ede\uff0c\u6709\u4e00\u4e9b\u8001\u4eba\u61c9\u8a72\u60f3\u7761\u4e86 \u901f\u5ea6\u975e\u5e38\u7de9\u6162\u7684 WaveNet\uff0c\u5df2\u7d93\u53ef\u4ee5\u9054\u5230\u5373\u6642\u5408\u6210\u7684\u6548\u76ca\u3002</td></tr><tr><td/><td>\u5716 10. \u5408\u6210\u8a9e\u97f3\u54c1\u8cea\u5be6\u9a57\u7d50\u679c\u76d2\u9b1a\u5716</td><td/></tr></table>", | |
| "html": null | |
| } | |
| } | |
| } | |
| } |