{ "paper_id": "O17-1021", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T07:59:36.634465Z" }, "title": "", "authors": [], "year": "", "venue": null, "identifiers": {}, "abstract": "", "pdf_parse": { "paper_id": "O17-1021", "_pdf_hash": "", "abstract": [], "body_text": [], "back_matter": [], "bib_entries": { "BIBREF0": { "ref_id": "b0", "title": "Sequence to Sequence Learning with Neural Networks", "authors": [ { "first": "Ilya", "middle": [], "last": "Sutskever", "suffix": "" }, { "first": "Oriol", "middle": [], "last": "Vinyals", "suffix": "" }, { "first": "Quoc", "middle": [ "V" ], "last": "Le", "suffix": "" } ], "year": 2014, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:1409.3215v3" ] }, "num": null, "urls": [], "raw_text": "Ilya Sutskever, Oriol Vinyals, Quoc V. Le. Sequence to Sequence Learning with Neural Networks, In arXiv:1409.3215v3 [cs.CL] 14 Dec 2014", "links": null }, "BIBREF1": { "ref_id": "b1", "title": "Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation", "authors": [ { "first": "Kyunghyun", "middle": [], "last": "Cho", "suffix": "" }, { "first": "Bart", "middle": [], "last": "Van Merrienboer", "suffix": "" }, { "first": "Caglar", "middle": [], "last": "Gulcehre", "suffix": "" }, { "first": "Dzmitry", "middle": [], "last": "Bahdanau", "suffix": "" }, { "first": "Fethi", "middle": [], "last": "Bougares", "suffix": "" }, { "first": "Holger", "middle": [], "last": "Schwenk", "suffix": "" }, { "first": "Yoshua", "middle": [], "last": "Bengio", "suffix": "" } ], "year": 2014, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:1406.1078v3[cs.CL]3" ] }, "num": null, "urls": [], "raw_text": "Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, Yoshua Bengio. Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation. In arXiv:1406.1078v3 [cs.CL] 3 Sep 2014.", "links": null }, "BIBREF2": { "ref_id": "b2", "title": "Distributed Representations of Words and Phrases and their Compositionality", "authors": [ { "first": "Tomas", "middle": [], "last": "Mikolov", "suffix": "" }, { "first": "Ilya", "middle": [], "last": "Sutskever", "suffix": "" }, { "first": "Kai", "middle": [], "last": "Chen", "suffix": "" }, { "first": "Greg", "middle": [], "last": "Corrado", "suffix": "" }, { "first": "Jeffrey", "middle": [], "last": "Dean", "suffix": "" } ], "year": 2013, "venue": "Advances in Neural Information Processing Systems", "volume": "26", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg Corrado, Jeffrey Dean. Distributed Representations of Words and Phrases and their Compositionality. Advances in Neural Information Processing Systems 26 (NIPS 2013).", "links": null }, "BIBREF3": { "ref_id": "b3", "title": "Distributed Representations of Words and Phrases and their Compositionality", "authors": [ { "first": "Tomas", "middle": [], "last": "Mikolov", "suffix": "" }, { "first": "Ilya", "middle": [], "last": "Sutskever", "suffix": "" }, { "first": "Kai", "middle": [], "last": "Chen", "suffix": "" }, { "first": "Greg", "middle": [], "last": "Corrado", "suffix": "" }, { "first": "Jeffrey", "middle": [], "last": "Dean", "suffix": "" } ], "year": 2013, "venue": "Advances in Neural Information Processing Systems", "volume": "26", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg Corrado, Jeffrey Dean. Distributed Representations of Words and Phrases and their Compositionality, Advances in Neural Information Processing Systems 26 (NIPS 2013).", "links": null }, "BIBREF5": { "ref_id": "b5", "title": "HMM-based Speech Synthesis System", "authors": [], "year": 2016, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "HMM-based Speech Synthesis System (HTS) : http://hts.sp.nitech.ac.jp , 2016, July.", "links": null }, "BIBREF6": { "ref_id": "b6", "title": "Conditional Random Field-based Parser and Language Model for Traditional Chinese Spelling Checker", "authors": [ { "first": "Yih-Ru", "middle": [], "last": "Wang", "suffix": "" }, { "first": "Yuan-Fu", "middle": [], "last": "Liao", "suffix": "" }, { "first": "Yeh-Kuang", "middle": [], "last": "Wu", "suffix": "" }, { "first": "Liang-Chun", "middle": [], "last": "Chang", "suffix": "" } ], "year": 2013, "venue": "Proceedings of the Seventh SIGHAN Workshop on Chinese Language Processing (SIGHAN-7)", "volume": "", "issue": "", "pages": "69--73", "other_ids": {}, "num": null, "urls": [], "raw_text": "Yih-Ru Wang, Yuan-Fu Liao, Yeh-Kuang Wu, Liang-Chun Chang,\"Conditional Random Field-based Parser and Language Model for Traditional Chinese Spelling Checker\" Proceedings of the Seventh SIGHAN Workshop on Chinese Language Processing (SIGHAN-7), pages 69-73,Nagoya, Japan, 14 October 2013.", "links": null }, "BIBREF8": { "ref_id": "b8", "title": "Sequence G2P toolkit", "authors": [], "year": null, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Sequence-to-Sequence G2P toolkit : https://github.com/cmusphinx/g2p-seq2seq", "links": null }, "BIBREF9": { "ref_id": "b9", "title": "Fully Character-Level Neural Machine Translation without Explicit Segmentation", "authors": [ { "first": "Jason", "middle": [], "last": "Lee", "suffix": "" }, { "first": "Kyunghyun", "middle": [], "last": "Cho", "suffix": "" }, { "first": "Thomas", "middle": [], "last": "Hofmann", "suffix": "" } ], "year": null, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:1610.03017v1[cs.CL]10" ] }, "num": null, "urls": [], "raw_text": "Jason Lee, Kyunghyun Cho, Thomas Hofmann. Fully Character-Level Neural Machine Translation without Explicit Segmentation. In arXiv:1610.03017v1 [cs.CL] 10", "links": null } }, "ref_entries": { "TABREF0": { "text": "The 2017 Conference on Computational Linguistics and Speech Processing ROCLING 2017, pp. 213-229 \uf0d3 The Association for Computational Linguistics and Chinese Language Processing", "html": null, "content": "
\u6458\u8981 \u50b3\u7d71\u7684\u8a9e\u97f3\u5408\u6210\u4f7f\u7528\u5148\u6587\u5b57\u5206\u6790\u5f8c\u8a9e\u97f3\u5408\u6210\u7684\u67b6\u69cb\uff0c\u4f46\u662f\u9019\u7a2e\u5169\u968e\u6bb5\u7684\u4f5c\u6cd5\uff0c \u901a\u5e38\u6703\u6709\uff0c\u82e5\u524d\u7d1a\u5206\u6790\u932f\u8aa4\uff0c\u5c31\u6703\u5f71\u97ff\u5f8c\u7d1a\u5408\u6210\uff0c\u4e14\u7121\u6cd5\u633d\u6551\u7684\u554f\u984c\u3002\u56e0\u6b64\uff0c\u5728 \u672c\u8ad6\u6587\u4e2d\u6211\u5011\u5e0c\u671b\u5617\u8a66\u628a\u524d\u5f8c\u7d1a\uff0c\u5168\u90e8\u90fd\u6539\u6210\u4ee5\u985e\u795e\u7d93\u7db2\u8def\u5be6\u73fe\uff0c\u4ee5\u4fbf\u5c07\u4f86\u53ef\u4ee5 \u76f4\u63a5\u5408\u6210\u4e00\u500b\u5927\u7684\u7aef\u5c0d\u7aef\u8a9e\u97f3\u5408\u6210\u985e\u795e\u7d93\u7db2\u8def\u3002\u4e3b\u8981\u7684\u60f3\u6cd5\u662f\uff0c\u76f4\u63a5\u4ee5\u5b57\u5143\u4e32\u70ba \u8f38\u5165\u55ae\u4f4d\uff0c\u4e26\u76e1\u91cf\u7528\u5927\u91cf\u672a\u6a19\u8a18\u8a9e\u6599\uff0c\u9032\u884c\u975e\u76e3\u7763\u5f0f\u985e\u795e\u7d93\u7db2\u8def\u8a13\u7df4\u3002\u6211\u5011\u7684\u7cfb \u7d71\u5305\u542b\u56db\u500b\u5b50\u7db2\u8def\uff0c\u5206\u5225\u662fDNNG\u4ee5sequence-to-sequence[1] [2]\u67b6\u69cb\u4f5c\u5b57\u8f49\u97f3\uff0cDNNC \u4ee5 word2vec[3] \u64f7\u53d6characterclass \uff0c DNNT \u4ee5 recurrent neural networklanguage model (RNNLM)[4]\uff0c\u6c42\u53d6\u5b57\u5143\u6642\u5e8f\u95dc\u4fc2\uff0c\u8207DNNs\u4ee5deep neural network\u9032\u884c\u8a9e\u97f3\u5408\u6210\u3002 \u5be6\u9a57\u8a9e\u6599\u7531\u5c08\u696d\u64ad\u97f3\u54e1\u9304\u88fd\uff0c\u5167\u5bb9\u5305\u62ec\u5b5f\u5fb7\u723e\u50b3\u5168\u66f8\u4ee5\u53ca\u5f9e\u7db2\u8def\u64f7\u53d6\u7d043000\u53e5\u7684 \u4e2d\u82f1\u593e\u96dc\u53e5\u5b50\u3002\u4e26\u4ee5\u76f8\u540c\u6587\u5b57\u8981\u6c42\u65b0\u820a\u7cfb\u7d71\u5404\u81ea\u5408\u6210\u6e2c\u8a66\u8a9e\u6599\uff0c\u8acb10\u4eba\u9032\u884c\u807d\u6e2c \u8a66\uff0c\u5206\u5225\u4ee5\u65b0\u820a\u7cfb\u7d71\u5404\u807d10\u53e5\uff0c\u9032\u884c\uff21/B/X\u504f\u597d\u5ea6\u6e2c\u8a66\uff0c\u8207\u4ee5\u65b0\u820a\u7cfb\u7d71\u5404\u807d20\u53e5\uff0c \u505amean opinion score(MOS)\u8a55\u5206\uff0c\u8a55\u4f30\u65b0\u820a\u7cfb\u7d71\u7684\u53ef\u7406\u89e3\u5ea6\uff0c\u81ea\u7136\u5ea6\u8207\u76f8\u4f3c\u5ea6\u3002 \u5f9e\u5be6\u9a57\u7d50\u679c\u767c\u73fe\uff0c\u5728\u53ef\u7406\u89e3\u5ea6\u3001\u81ea\u7136\u5ea6\u548c\u76f8\u4f3c\u5ea6\u65b9\u9762\uff0c\u5206\u5225\u670972%\u300170%\u548c61%\u7684 \u4eba\u504f\u597d\u65b0\u7cfb\u7d71\u3002\u800c\u4e14\u65b0\u7cfb\u7d71\u7684\u53ef\u7406\u89e3\u5ea6\u3001\u81ea\u7136\u5ea6\u548c\u76f8\u4f3c\u5ea6\u7684MOS\u4e3b\u89c0\u5206\u6578\u5404\u70ba 3.59\u30013.1\u548c3.18\u5206\uff0c\u9ad8\u65bc\u820a\u7cfb\u7d71\u76843.33\u30013.03\u548c2.9\u5206\uff0c\u986f\u793a\u6211\u5011\u6240\u63d0\u51fa\u7684\u7cfb\u7d71\u6548\u80fd \u76f8\u7576\u4e0d\u932f\uff0c\u5370\u8b49\u6211\u5011\u63d0\u51fa\u7684\u60f3\u6cd5\u78ba\u5be6\u53ef\u884c\u3002 \u95dc\u9375\u8a5e\uff1a\u8a9e\u97f3\u5408\u6210\u3001\u6df1\u5ea6\u985e\u795e\u7d93\u7db2\u8def\u3001\u7aef\u5c0d\u7aef \u4e00\u3001\u7c21\u4ecb \u50b3\u7d71\u8a9e\u97f3\u5408\u6210\u7cfb\u7d71\u4e2d\uff0c\u5305\u542b\u5169\u500b\u8655\u7406\u968e\u6bb5\uff0c\u5206\u5225\u662f\u524d\u7aef\u6587\u672c\u5206\u6790\u8207\u5f8c\u7aef\u8072\u97f3\u5408 \u6210(\u5982\u4e0b\u57161)\u3002\u5176\u4e2d\u5728\u524d\u7aef\u6587\u672c\u5206\u6790\u6a21\u7d44\uff0c\u5305\u542b\u4e86\u6587\u5b57\u6b63\u898f\u5316\u3001\u65b7\u8a5e\u3001\u5b57\u8f49\u97f3\u3001\u8a5e \u6027(part of speech\uff0cPOS)\u6a19\u8a3b[5]\u7b49\u6587\u5b57\u5206\u6790\uff0c\u4ee5\u6c42\u53d6\u6587\u8108\u8a0a\u606f\u3002\u53e6\u4e00\u65b9\u9762\u5728\u5f8c\u7aef\u7684 \u8072\u97f3\u5408\u6210\u6a21\u7d44\uff0c\u5247\u900f\u904e\u524d\u7d1a\u6c42\u51fa\u7684\u7684\u6587\u8108\u8a0a\u606f\u7279\u5fb5\u53c3\u6578\uff0c\u9032\u884c\u8a9e\u8a00\u5408\u6210\u6a21\u578b\u8a13\u7df4\uff0c \u4ee5\u5408\u6210\u8072\u97f3\u8a0a\u606f\u3002 \u5716 1 \u50b3\u7d71 TTS \u7cfb\u7d71\u67b6\u69cb\u5716 \u4f46\u5728\u9019\u500b\u67b6\u69cb\u4e0b\uff0c\u524d\u7d1a\u548c\u5f8c\u7d1a\u90fd\u662f\u7368\u7acb\u5efa\u69cb\u7684\uff0c\u56e0\u6b64\uff0c\u5982\u679c\u524d\u7d1a\u51fa\u73fe\u554f\u984c\u7684\u8a71\uff0c \u5f8c\u7d1a\u4e5f\u6703\u53ea\u80fd\u7e7c\u627f\u524d\u7d1a\u7684\u932f\u8aa4\uff0c\u4f46\u537b\u7121\u6cd5\u5f80\u56de\u8abf\u6574\uff0c\u4ee5\u66f4\u6b63\u524d\u7d1a\u7684\u932f\u8aa4\u3002\u800c\u4e14\uff0c \u50b3\u7d71\u4e0a\uff0c\u524d\u7d1a\u6240\u7528parser\uff0c\u90fd\u662f\u4f7f\u7528\u81ea\u7136\u8a9e\u97f3\u5c08\u5bb6\u767c\u5c55\u597d\u7684\u73fe\u6210\u7cfb\u7d71\uff0c\u4e26\u4e0d\u5bb9\u6613\u81ea \u884c\u4fee\u6539\u3002 \u56e0\u6b64\uff0c\u6211\u5011\u5e0c\u671b\u80fd\u628a\u524d\u5f8c\u7d1a\uff0c\u5168\u90e8\u90fd\u6539\u6210\u4ee5\u985e\u795e\u7d93\u7db2\u8def\u5be6\u73fe\uff0c\u4ee5\u4fbf\u5c07\u4f86\u53ef\u4ee5\u76f4 \u63a5\u5408\u6210\u4e00\u500b\u5982\u57162\u7684\u7aef\u5c0d\u7aef\u8a9e\u97f3\u5408\u6210\u985e\u795e\u7d93\u7db2\u8def\uff0c\u907f\u514d\u50b3\u7d71\u5169\u968e\u6bb5\u67b6\u69cb\u7684\u7f3a\u9ede\u3002 \u57162 \u5b57\u5143\u5c64\u7d1a\u7684End-to-End\u67b6\u69cb \u70ba\u9054\u5230\u6b64\u76ee\u7684\uff0c\u6211\u5011\u5b8c\u5168\u7528\u985e\u795e\u7d93\u7db2\u8def\uff0c\u53d6\u4ee3\u524d\u7d1a\u7684\u5b57\u8f49\u97f3\uff0c\u6587\u5b57\u5206\u6790\uff0c\u8207\u5f8c \u7d1a\u7684\u8a9e\u97f3\u5408\u6210\u6a21\u7d44\u3002\u4e3b\u8981\u7684\u60f3\u6cd5\u662f\uff0c\u5148\u5c07\u539f\u672c\u7684\u7cfb\u7d71\u6539\u6210\u76f4\u63a5\u4ee5\u5b57\u5143\u4e32\u70ba\u8655\u7406\u55ae \u4f4d\uff0c\u518d\u4f7f\u7528\u56db\u500b\u5b50\u7db2\u8def\uff0c\u5206\u5225\u662fDNNG\u4ee5sequence-to-sequence\u67b6\u69cb\u4f5c\u5b57\u8f49\u97f3\uff0cDNNC \u4ee5Word2vec\u64f7\u53d6\u6bcf\u500b\u8f38\u5165\u5b57\u5143\u7684\u985e\u5225\u8207\u6587\u6cd5\u5c6c\u6027\uff0cDNNT\u4ee5RNNLM\uff0c\u6c42\u53d6\u6bcf\u500b\u5b57 \u5143\u5728\u6574\u500b\u53e5\u5b50\u4e2d\u7684\u72c0\u614b\u8ddf\u6642\u5e8f\u95dc\u4fc2\uff0c\u8207DNNs\u4ee5deep neural network\u63a5\u6536\u5404\u500b\u7db2\u8def\u7684 \u96b1\u85cf\u5c64\u795e\u7d93\u5143\u6fc0\u767c\u8cc7\u8a0a\uff0c\u9032\u884c\u8a9e\u97f3\u5408\u6210\u3002 \u5728\u6b64\u67b6\u69cb\u4e2d\u7684DNNG\u56e0\u70ba\u662f\u4f7f\u7528sequence-to-sequence\u6a21\u578b\uff0c\u53ef\u4ee5\u8655\u7406\u672a\u66fe\u770b\u904e\u7684 \u5b57\u8a5e\u7684\u767c\u97f3\uff0cDNNC\u8207DNNT\u6240\u4f7f\u7528\u7684Word2vec\u8207RNNLM\u66f4\u53ef\u4ee5\u5584\u7528\u5927\u91cf\u96a8\u624b\u53ef\u5f97 End-to-End \u67b6\u69cb \u5408\u6210\u8a9e\u97f3 \u4eca \u5929 \u5929 \u6c23 \u6674 C0 C1 C2 C3 C4 \u7684\u672a\u6a19\u8a18\u6587\u5b57\u8a9e\u6599\uff0c\u9032\u884c\u975e\u76e3\u7763\u5f0f\u8a13\u7df4\uff0c\u5145\u5206\u8a13\u7df4\u6574\u500b\u985e\u795e\u7d93\u7db2\u8def\uff0c\u907f\u958b\u50b3\u7d71 parser\uff0c\u9700\u8981\u4f9d\u8cf4\u4eba\u5de5\u6a19\u8a18\u8a9e\u6599\uff0c\u624d\u80fd\u9032\u884c\u624d\u80fd\u9032\u884c\u8a13\u7df4\u7684\u554f\u984c\u3002 \u4e8c\u3001\u671d\u5411\u7aef\u5c0d\u7aef\u8a9e\u97f3\u5408\u6210\u67b6\u69cb \u70ba\u4e86\u671d\u5411\u4ee5\u57fa\u65bc\u5b57\u5143\u5c64\u7d1a\u7684\u7aef\u5c0d\u7aef\u8a9e\u97f3\u5408\u6210\u7cfb\u7d71\uff0c\u6211\u5011\u5c07\u6574\u500b\u7aef\u5c0d\u7aef\u8a9e\u97f3\u5408\u6210 \u7cfb\u7d71\uff0c\u4ee5\u57163\u7684\u65b9\u5f0f\uff0c\u5207\u5206\u62104\u584aDNNs\u3002\u5206\u5225\u662f\u6587\u5b57\u8f49\u62fc\u97f3\u7db2\u8defDNNG\u3001\u5b57\u5143\u5c6c\u6027 \u8207\u89d2\u8272\u5206\u985eDNNc\u3001\u5b57\u5143\u6642\u5e8f\u95dcDNNT\u548c\u8a9e\u97f3\u5408\u6210DNNs\u3002 \u57163 One-Stage\u67b6\u69cb\u5167\u90e8\u529f\u80fd\u65b9\u584a \u5176\u4e2d\u5b57\u8f49\u97f3(DNN\uff27)\u4e3b\u8981\u662f\u5229\u7528sequence-to-sequence\u6a21\u578b\u8f49\u63db\u6587\u672c\u7684\u62fc\u97f3\u548c\u97f3\u8abf\uff0c DNNc\u4f7f\u7528 Word2vec \u4f86\u6293\u53d6\u5b57\u5143\u7279\u6027\uff0c\u800cDNNT \u900f\u904eRNNLM\u4f86\u64f7\u53d6\u5b57\u5143\u524d\u5f8c\u6642 \u5e8f\u8cc7\u8a0a\u95dc\u4fc2\u3002\u6700\u5f8c\u5f62\u6210\u4e00\u500b\u6240\u6709\u7684\u6587\u8108\u8a0a\u606f\uff0c\u90fd\u662f\u7531\u985e\u795e\u7d93\u7db2\u8def\u81ea\u52d5\u7522\u751f\u3002\u9019\u6a23 \u4e00\u4f86\u4fbf\u80fd\u907f\u958b\u50b3\u7d71\u6587\u672c\u5206\u6790\u7684\u8af8\u591a\u4e0d\u4fbf\u3002\u6700\u5f8c\u5f8c\u7d1a\u8072\u97f3\u5408\u6210\u90e8\u5206\u5247\u4f7f\u7528DNNs\u4f86 \u63a5\u6536DNNG\u3001DNNc\u8207DNNT\u64f7\u53d6\u51fa\u7684\u6587\u8108\u8a0a\u606f\uff0c\u4ee5\u5408\u6210\u8a9e\u97f3\u3002\u4ee5\u4e0b\u9032\u4e00\u6b65\u8a73\u7d30\u6558\u8ff0 \u5404\u5b50\u7db2\u8def\u7684\u5be6\u969b\u4f5c\u6cd5\u3002 (\u4e00)\u3001\u5b57\u8f49\u97f3(DNNG) Seq2Seq \u5168\u540d\u662f Sequence to Sequence\uff0cSeq2Seq \u5c31\u50cf\u4e00\u500b\u7ffb\u8b6f\u6a21\u578b\uff0c\u6bd4\u5982\u8f38\u5165 \u5e8f\u5217\u662f\u82f1\u6587(hello)\uff0c\u8f38\u51fa\u5e8f\u5217\u662f\u4e2d\u6587(\u4f60\u597d)\uff0c\u8a72\u6280\u8853\u6539\u5584\u4e86\u50b3\u7d71\u8f38\u5165\u5e8f\u5217\u548c\u8f38\u51fa\u5e8f \u5217\u9577\u5ea6\u9700\u8981\u4e00\u6a23\u7684\u554f\u984c\uff0c\u958b\u59cb\u4e86\u5c07\u6df1\u5ea6\u795e\u7d93\u7db2\u8def\u6a21\u578b(DNN)\u904b\u7528\u5728\u6a5f\u5668\u7ffb\u8b6f\u9019\u985e \u5408\u6210\u8a9e\u97f3 DNNc DNNs \u6587\u672c\u5206\u6790 \u8072\u97f3\u5408\u6210 DNNT DNNG \u4eca \u5929 \u5929 \u6c23 \u6674 DNNP C0 C1 C2 C3 C4 \u578b\u7684\u4efb\u52d9\u3002Seq2Seq \u6700\u65e9\u662f\u7531\u5169\u7bc7\u6587\u7ae0\u95e1\u8ff0\u4ed6\u7684\u4e3b\u8981\u601d\u60f3\uff0c\u5206\u5225\u662f Google \u7684 Sequence to Sequence Learning with Neural Networks[1]\u548c Yoshua Bengio \u5718\u968a\u7684 Learning Phrase Representation using RNN Encoder-Decoder for Statistical Machine Translation[2]\uff0c\u9019\u5169\u7bc7\u6587\u7ae0\u91dd\u5c0d\u6a5f\u5668\u7ffb\u8b6f\u7684\u554f\u984c\u4e0d\u7d04\u800c\u540c\u7684\u63d0\u51fa\u76f8\u4f3c\u7684\u89e3\u6c7a\u60f3\u6cd5\uff0c Seq2Seq \u7531\u6b64\u7522\u751f\u3002 \u5728\u5b57\u8f49\u97f3\u65b9\u9762\u5229\u7528Seq2Seq\u6280\u8853\uff0cSeq2Seq\u5168\u540d\u662fSequence to Sequence\uff0cSeq2Seq\u7684 \u6838\u5fc3\u60f3\u6cd5\u5c31\u662f\u900f\u904e\u6df1\u5ea6\u795e\u7d93\u7db2\u8def\u6a21\u578b(\u5e38\u7528\u7684\u662fLong-Short Term Memory\uff0cLSTM)\uff0c \u5c07\u4e00\u500b\u8f38\u5165\u7684\u5e8f\u5217\u6620\u5c04\u5230\u4e00\u500b\u8f38\u51fa\u7684\u5e8f\u5217\u3002\u800c\u9019\u904e\u7a0b\u5305\u542b\u5169\u500b\u74b0\u7bc0\uff0c\u5206\u5225\u662f\u5c07\u8f38 \u5165\u7de8\u78bc\u548c\u89e3\u78bc\u7522\u751f\u8f38\u51fa\u3002\u5728\u9019\u500b\u6a21\u578b\u4e2d\u6bcf\u4e00\u500b\u6642\u9593\u7684\u8f38\u5165\u548c\u8f38\u51fa\u662f\u4e0d\u4e00\u6a23\u7684\uff0c\u4f8b \u5982\u73fe\u5728\u7684\u8f38\u5165\u7de8\u78bc\u5e8f\u5217\u662f\u300c\u4e0a\u73ed\u65cfEOS\u300d\uff0c\u5176\u4e2dEOS(End of Sentence)\u70ba\u53e5\u5c3e\u8b58\u5225 \u7b26\u865f\uff0c\u4f9d\u5e8f\u5c07\u300c\u4e0a\u300d\u3001\u300c\u73ed\u300d\u3001\u300c\u65cf\u300d\u3001\u300cEOS\u300d\u50b3\u5165\u6a21\u578b\u4e2d\uff0c\u5c07\u8f38\u5165\u5e8f\u5217\u6620\u5c04 \u70ba\u89e3\u78bc\u8f38\u51fa\u5e8f\u5217\u300css_ch-A:_ch-N_chp_ch-A:_ch-n_chts_ch-u:_ch<EOS>\u300d\u3002 \u57164 Google\u7684Sequence to Sequence\u67b6\u69cb[1] \u6b64\u5916\u70ba\u4f7fSeq2Seq\u7684G2P\u67b6\u69cb\u53ef\u4ee5\u8003\u616e\u5230\u524d\u5f8c\u6587\u7684\u5167\u5bb9\uff0c\u9032\u800c\u7d66\u4e88\u7576\u524d\u5b57\u4e00\u500b\u8f03 \u70ba\u53ef\u80fd\u7684\u767c\u97f3\uff0c\u4ee5\u8655\u7406\u591a\u97f3\u5b57\u7684\u554f\u984c\uff0c\u56e0\u6b64\u6211\u5011\u4e00\u6b21\u4e0d\u662f\u53ea\u8f38\u5165\u4e00\u500b\u5b57\uff0c\u800c\u662f\u540c \u6642\u5305\u542b\u5176\u524d\u5f8c\u95dc\u4fc2\u3002 \u88681\u70ba\u9032\u884cSeq2Seq\u7684G2P\u8a13\u7df4\u6642\u7684\u8f38\u5165\u8cc7\u6599\u5f62\u5f0f[] []\uff0c\u4e3b\u8981\u6a21\u4effCNN\u958b\u4e00\u500b sliding window\u53bb\u6383\u524d\u5f8c\u7684\u5b57\uff0c\u8b93\u5b83\u80fd\u5f80\u524d\u8207\u5f80\u5f8c\u591a\u770b5\u500b\u5b57\uff0c\u4ee5\u7372\u5f97\u66f4\u591a\u8a0a\u606f\uff0c\u800c \u80fd\u5b78\u5f97\u66f4\u597d\u3002 \u88681 G2P\u8a13\u7df4\u8cc7\u6599\u683c\u5f0f Decode Encode \u4e0a \u73ed \u65cf <EOS> ss_ch-A:_ch-N_ch p_ch-A:_ch-n_ch ts_ch-u:_ch <EOS> ss_ch-A:_ch-N_ch p_ch-A:_ch-n_ch ts_ch-u:_ch (\u4e8c)\u3001\u5b57\u5143\u8a9e\u610f\u8207\u6587\u6cd5\u5c6c\u6027(DNNc) \u7a7a\u9593\u4e2d\u767c\u73fe\uff0c\u76f8\u805a\u5728\u4e00\u8d77\u7684\u5411\u91cf\u8f49\u63db\u56de\u6587\u5b57\u5f8c\uff0c\u6703\u662f\u76f8\u8fd1\u5c6c\u6027\u7684\u8a5e\u5f59\u3002word2vec \u80fd\u5920\u5c07\u5b57\u8a5e\u8a9e\u610f\u548c\u6587\u6cd5\u89d2\u8272\u505a\u5206\u985e\uff0c\u800c\u4e14\u5b83\u4e0d\u9700\u8981\u7d66\u6a19\u8a3b\u904e\u7684\u6587\u5b57\u8a9e\u6599\u5c31\u80fd\u8a13\u7df4\uff0c \u9019\u53ef\u4ee5\u907f\u958b\u50b3\u7d71 parser \u9700\u8981\u4eba\u5de5\u6a19\u8a3b\u7684\u7e41\u96dc\u5de5\u4f5c\uff0c\u4e5f\u80fd\u5920\u505a\u5230\u985e\u4f3c POS \u7684\u529f\u80fd\u3002 \u6211\u5011\u5229\u7528Word2Vec\uff0c\u8a13\u7df4\u5982\u57165\u7684\u985e\u795e\u7d93\u7db2\u8def\uff0c\u5c07\u5b57\u5143\u8f49\u5230\u5411\u91cf\u7a7a\u9593\u3002\u8a13\u7df4\u5b8c \u6210\u5f8c\uff0c\u518d\u64f7\u53d6\u96b1\u85cf\u5c64\u795e\u7d93\u5143\u7684word vector\u8f38\u51fa\u5411\u91cf\uff0c\u7576\u4f5c\u6bcf\u500b\u5b57\u5143\u7684\u8a9e\u610f\u8207\u6587\u6cd5 \u89d2\u8272\u8cc7\u8a0a\u3002\u4e3b\u8981\u662f\u5c07\u5927\u91cf\u672a\u6a19\u8a3b\u8a9e\u6599\u5012\u5165Word2vec\uff0c\u8b93\u4ed6\u81ea\u884c\u8a13\u7df4\uff0c\u518d\u5229\u7528\u6c42\u51fa \u4e4b\u5b57\u5143\u5411\u91cf\uff0c\u754c\u5b9a\u6bcf\u500b\u5b57\u5143\u7684\u5c6c\u6027\u8207\u6587\u6cd5\u89d2\u8272\u95dc\u7cfb\u3002 \u4e5f\u56e0\u70baWord2vec\u6240\u8a13\u7df4\u51fa\u4f86\u7684\u5b57\u5411\u91cf\u7a7a\u9593\uff0c\u80fd\u6709\u610f\u7fa9\u7684\u8868\u793a\u5b57\u7684\u5c6c\u6027\uff0c\u4e26\u4e14\u80fd \u5920\u5c07\u8a13\u7df4\u51fa\u4f86\u7684\u5b57\u5411\u91cf\u9032\u884c\u6392\u5217\uff0c\u8b93\u985e\u4f3c\u5c6c\u6027\u7684\u5b57\u805a\u985e\u5728\u4e00\u8d77\uff0c\u6240\u4ee5\u6211\u5011\u89ba\u5f97\u7528 \u5b83\u4f86\u66ff\u4ee3\u50b3\u7d71Parser\u4e2d\u7684\u6a19\u8a3b\u8a5e\u6027\u529f\u80fd\u53ef\u80fd\u662f\u884c\u5f97\u901a\u7684\u3002 \u5716 5 \u64f7\u53d6\u96b1\u85cf\u5c64\u5206\u985e\u8cc7\u8a0a W(t) W(t+2) W(t-1) W(t-2) W(t+1) input projection output SUM CBOW \u5c07\u5411\u91cf\u7a7a\u9593\u5206\u985e\u5b8c\u7684\u8cc7\u8a0a\u53d6\u51fa RNNLM\u80fd\u5920\u76f4\u63a5\u4f7f\u7528\u7121\u6a19\u8a3b\u6587\u7ae0\u8a9e\u6599\uff0c\u9032\u884c\u8a13\u7df4\uff0c\u4e26\u56e0\u5176\u64c1\u6709\u8a18\u61b6\u80fd\u529b\uff0c\u80fd\u5920 \u5b78\u7fd2\u5230\u8f03\u9577\u6642\u9593\u7684\u6587\u7ae0\u8108\u7d61\u3002\u6240\u4ee5\u6211\u5011\u7528\u5927\u91cf\u7121\u6a19\u8a18\u8a9e\u6599\uff0c\u8a13\u7df4\u5b8cRNNLM\u5f8c\uff0c \u85c9\u7531\u64f7\u53d6RNNLM\u96b1\u85cf\u5c64\u795e\u7d93\u5143\u7684\u6fc0\u767c\u72c0\u614b\u503c(\u5982\u57166\u6240\u793a)\uff0c\u7576\u4f5c\u67d0\u4e00\u5b57\u5143\u5728\u6587 \u7ae0\u6bb5\u843d\u4e2d\u7684\u6642\u5e8f\u72c0\u614b\u8cc7\u8a0a\u3002\u6b64\u5916\uff0c\u6211\u5011\u4e26\u9032\u4e00\u6b65\u9032\u884c\u91cf\u5316\uff0c\u6574\u7406\u62100\u82071\u7684\u503c\uff0c\u7528 \u4f86\u8868\u793a\u67d0\u4e00\u5b57\u5143\u5728\u53e5\u5b50\u4e2d\u7684\u6642\u5e8f\u72c0\u614b\u8cc7\u8a0a\u3002 \u57166 \u64f7\u53d6\u96b1\u85cf\u5c64\u6642\u5e8f\u8cc7\u8a0a \u56db\u3001DNN \u67b6\u69cb(DNNS) \u5728\u672c\u8ad6\u6587\u4e2d\uff0c\u70ba\u4e86\u5c07\u6574\u500b\u67b6\u69cb\u7528\u6210 One-Stage\uff0c\u5f8c\u7d1a\u8072\u97f3\u5408\u6210\u52e2\u5fc5\u4e5f\u8981\u66ff\u63db\u6210 \u985e\u795e\u7d93\u7db2\u8def\u7684\u67b6\u69cb\uff0c\u5728\u6211\u5011\u7684 One-Stage \u4e2d\u5229\u7528\u4e86 HTS[6] version 2.3.1 \u4e2d\u65b0\u6dfb\u52a0\u7684 frame-by-frame modeling option using DNN based on HMM state alignment\uff0c\u4f86\u63a5\u6536\u524d \u9762 3 \u500b DNN \u6240\u8403\u53d6\u51fa\u7684\u8cc7\u8a0a\u4e26\u4e14\u8a13\u7df4\u5408\u6210\u8a9e\u97f3\u3002 \u524d\u97623\u500bDNN\u6240\u8403\u53d6\u51fa\u7684\u8cc7\u8a0a\uff0c\u53ef\u4ee5\u9032\u4e00\u6b65\u5408\u4f75\u6210\u70ba\u5b8c\u6574\u7684\u6587\u8108\u8a0a\u606f\u3002DNNS\u8981 Wikipedia\u3002Gigaword\u8207Wikipedia\u8a9e\u6599\u7684\u7d71\u8a08\u8cc7\u6599\u5982\u88683\u6240\u793a\u3002 4\u3001\u6587\u8108\u8a0a\u606f\u6c42\u53d6\u65b9\u6cd5\u8207\u8a2d\u5b9a \u886811 \u4f7f\u7528RNNLM\u7522\u751f\u4e4b\u6587\u53e5 POS\u8cc7\u8a0a\u4f86\u505a\u6bd4\u8f03\uff0c\u89c0\u5bdf\u5b57\u5143\u7684\u8a9e\u610f\u6587\u6cd5\u8cc7\u8a0a\u7684\u53d6\u4ee3\u8207\u50b3\u7d71\u65b9\u5f0f\u7684\u5dee\u7570\u3002 \u5716 12 Dnnp+HTS \u67b6\u69cb\u8207 Dnnp+DNNs \u67b6\u69cb\u504f\u597d\u5ea6\u6bd4\u8f03 \u5b78\u7fd2\u7684\u5c31\u662fDNNp\u8f38\u51fa\u7684\u6587\u8108\u8a0a\u606f\uff0c\u8207\u8a13\u7df4\u8a9e\u6599\u9593\u7684\u5c0d\u61c9\u95dc\u4fc2\u3002\u5728DNNS\u9019\u908a\uff0c Last word Context(t-1) Input(t) Context(t) Output(t) Next word \u53d6\u51fa\u96b1\u85cf\u5c64 nerual \u8cc7\u8a0a \u4e09\u3001\u8a9e\u97f3\u5408\u6210\u5be6\u9a57\u7d50\u679c\u8207\u5206\u6790 \u70ba\u4e86\u8207\u50b3\u7d71\u65b9\u6cd5\u4f5c\u6bd4\u8f03\uff0c\u820a\u7cfb\u7d71\u4f7f\u7528Parser\u505a\u6587\u672c\u5206\u6790\u548c\u67e5\u8868\u5b57\u8f49\u97f3\uff0c\u800c\u65b0\u7cfb \u7d71\u5247\u662f\u4f7f\u7528DNNP\u3001DNNc\u8207DNNT\u4f86\u5206\u6790\u6587\u672c\uff0cDNNS\u4f86\u5408\u6210\u8a9e\u97f3\u3002\u65b0\u820a\u7cfb\u7d71\u7686\u4f7f\u7528 \u76f8\u540c\u7684\u4e2d\u82f1\u593e\u96dc\u8a9e\u6599\u8a13\u7df4\u3002\u6211\u5011\u8981\u6bd4\u8f03\u7684\u662f\u65b0\u820a\u7cfb\u7d71\u5408\u6210\u97f3\u6a94\u7684\u76f8\u4f3c\u5ea6\u3001\u53ef\u7406\u89e3 \u5ea6\u3001\u81ea\u7136\u5ea6\u7684\u504f\u597d\u5ea6\u8207MOS\u5206\u6578\u3002\u70ba\u6c42\u516c\u5e73\u6bd4\u8f03\uff0c\u9078\u7528\u6c92\u6709\u5728\u8a13\u7df4\u904e\u7a0b\u4e2d\u51fa\u73fe\u7684 \u8a9e\u6599\u4f86\u5408\u6210\uff0c\u6700\u5f8c\u6bd4\u8f03\u7684\u97f3\u6a94\u7686\u70ba\u540c\u6a23\u6587\u5b57\uff0c\u4f46\u53d7\u6e2c\u8005\u4e0d\u77e5\u9053\u90a3\u500b\u97f3\u6a94\u70ba\u65b0\u6216\u820a \u7cfb\u7d71\u6240\u5408\u6210\uff0c\u4ee5\u76f2\u6e2c\u65b9\u5f0f\u9032\u884c (\u4e00)\u3001\u5be6\u9a57\u8a2d\u5b9a 1\u3001\u5b57\u8f49\u97f3\u8a9e\u6599 \u5728DNNG\u90e8\u5206\uff0cG2P\u8a13\u7df4\u8a9e\u6599\u5206\u5225\u70baTCC300\u8207\u6211\u5011\u5be6\u9a57\u5ba4\u6709\u768410\u842c\u5b57\u8a5e\u5b57\u5178\uff0c TCC300\u662f\u6587\u7ae0\u6027\u8cea\u7684\u8a9e\u6599\uff0c\u800c10\u842c\u5b57\u5b57\u5178\u5247\u662f\u4ee5\u55ae\u5b57\u8a5e\u548c\u591a\u5b57\u8a5e\u70ba\u4e3b\uff0c\u5982\u88682\u3002 \u5b57\u6578 \u6027\u8cea TCC300 \u6587\u5b57\u8a9e\u6599\u5eab 286685 \u6587\u7ae0 10 \u842c\u5b57\u8a5e\u5b57\u5178\u8a9e\u6599 231477 \u55ae\u5b57\u3001\u8a5e 2\u3001\u6587\u672c\u8a9e\u6599 \u5728DNNc\u8207DNNT\u90e8\u5206\uff0c\u6211\u5011\u4f7f\u7528Mikolov\u5718\u968a\u7684open source\uff0c\u5206\u5225\u662fword2vec\u8207 RNNLM Toolkit \uff0c\u8a13\u7df4\u9019\u5169\u500b\u6a21\u578b\u7684\u8a9e\u6599\u70ba Chinese Gigaword Second Edition + Number of Hidden&Units 3layers1024units activation Sigmoid optimizer Adam Batch size 256 learnRate 0.001 \u88688 \u5927\u9678\u6587\u7ae0\u4e2d\u65b0\u820aG2P\u7684\u97f3\u7bc0\u6b63\u78ba\u7387 \u4e00\u822c\u6587\u7ae0 \u7e3d\u5171 5880 \u53e5(\u7d04 31 \u842c\u500b\u97f3\u7bc0) Accuracy \u820a G2P \u932f 3096 \u500b\u97f3\u7bc0 99.0% \u65b0 G2P \u932f 8488 \u500b\u97f3\u7bc0 97.3% \u8a13\u7df4\u51fa\u4f86\u7684 RNNLM\uff0c\u5728\u5229\u7528 RNNLM \u6a21\u578b\u7522\u751f\u51fa\u7684\u53e5\u5b50\u4e2d\uff0c\u8a9e\u610f\u6216\u662f\u6642\u9593\u9806\u5e8f \u4e0a\u90fd\u8cbc\u8fd1\u65b0\u805e\u6587\u5b57\u3002\u56e0\u6b64\u61c9\u7528 RNNLM \u80fd\u5b78\u5230\u6642\u5e8f\u95dc\u4fc2\u7684\u7279\u6027\uff0c\u525b\u597d\u8207\u50b3\u7d71\u6587\u672c \u672c\u5be6\u9a57\u5229\u7528Word2vec\u4f86\u6c42\u53d6\u5b57\u5143\u7684\u8a9e\u610f\u548c\u6587\u6cd5\u89d2\u8272\u8cc7\u8a0a\u3002\u5728\u6b64\u6211\u5011\u8207\u50b3\u7d71Parser\u7684 \u81ea\u7136\u5ea6\u8a55\u5206 3 \u7684\u8a9e\u97f3\u5408\u6210\u7cfb\u7d71\u3002\u56e0\u6b64\u4ee5\u5404\u65b9\u9762\u4f86\u770b\uff0c\u6574\u500b\u67b6\u69cb\u4f7f\u7528\u795e\u7d93\u7db2\u8def\u4f86\u5be6\u505a\u7684\u78ba\u6703\u6bd4 3.1 \u5206\u6790\u4e2d\u7684\u6642\u9593\u9806\u5e8f\u95dc\u4fc2\u76f8\u4f3c\uff0c\u6240\u4ee5\u5c07\u5176\u61c9\u7528\u5728\u6211\u5011\u7684\u60f3\u6cd5\u4e0a\u61c9\u8a72\u662f\u53ef\u884c\u7684\u3002 \u53ef\u7406\u89e3\u5ea6\u8a55\u5206 3.1 3.3 \u76f8\u4f3c\u5ea6\u8a55\u5206 3.1 \u81ea\u7136\u5ea6\u8a55\u5206 2.9 3.1 \u76f8\u4f3c\u5ea6\u8a55\u5206 3.2 3 43% 57% \u81ea\u7136\u5ea6\u504f\u597d \u7684\u8072\u97f3\u660e\u986f\u597d\u5f88\u591a\u3002 \u8868 16 \u8207\u5716 16\u3002\u7d50\u679c\u986f\u793a\u65b0\u7cfb\u7d71\u6240\u5408\u6210\u7684\u8072\u97f3\u5728\u5404\u65b9\u9762\u90fd\u52dd\u904e\u65bc\u50b3\u7d71 Paser+HTS b\u3001\u66ff\u63dbParser\u8a5e\u53caPOS\u8cc7\u8a0a\u8207Word2vec\u5f71\u97ff\u4e4b\u6bd4\u8f03 \u7d14\u4e2d\u6587 Parser \u53ef\u7406\u89e3\u5ea6\u8a55\u5206 3.1 3.3 51% 49% \u76f8\u4f3c\u5ea6\u504f\u597d \u7684Two-Stage\u9084\u8981\u597d\u3002\u5be6\u9a57\u7d50\u679c\u5982\u4e0b\u571615\u6240\u793a\u3002\u986f\u7136\u4f7f\u7528DNNp+DNNs\u67b6\u69cb\uff0c\u6240\u5408\u6210 \u91cf\u6e2c\u8f38\u51fa\u5408\u6210\u8a9e\u97f3\u7684\u932f\u8aa4\u6210\u672c\u51fd\u6578\uff0c\u56de\u904e\u982d\u4f86\u8a13\u7df4\u6574\u500b\u7cfb\u7d71\u3002\u5be6\u9a57\u7d50\u679c\u7e3d\u7d50\u5728 RNNLM 48% 52% \u53ef\u7406\u89e3\u5ea6\u504f\u597d 3 \u886814 \u65b0\u820a\u6642\u5e8f\u95dc\u4fc2\u4e4bMOS\u4e3b\u89c0\u5206\u6578 Dnnp+HTS DNNp\u3001DNNc\u8207DNNT\uff0c\u5f8c\u7d1a\u4f7f\u7528DNNS)\uff0c\u4f86\u770b\u6211\u5011\u6240\u63d0\u51fa\u7684\u65b0\u67b6\u69cb\u6709\u6c92\u6709\u6bd4\u50b3\u7d71 \u6709\u5b50\u7db2\u8def\uff0c\u4e26\u4ee5\u76ee\u524d\u7684\u5b50\u7db2\u8def\u7684\u8a13\u7df4\u7d50\u679c\u7576\u5927\u7db2\u8def\u7684\u4fc2\u6578\u7684\u521d\u59cb\u503c\u3002\u6700\u5f8c\u76f4\u63a5 Dnnp+DNNs \u5728\u8868 11 \u53ef\u4ee5\u770b\u898b\u5229\u7528 Chinese Gigaword Second Edition + Wikipedia \u6587\u5b57\u8a9e\u6599\u5eab \u88686 HTS\u7684DNN\u53c3\u6578 \u820a G2P \u932f 26 \u5b57 99.64% \u65b0 G2P \u932f 32 \u5b57 99.56% 6 \u5427\u3001\u5440\u3001\u5462\u3001\u54e6\u3001\u554a\u3001\u5566\u3001\u5594\u3001\u55ce\u3001\u55ef\u3001\u5475 \u7d14\u4e2d\u6587 \u820a G2P \u65b0 G2P \u571610 \u65b0\u820a\u67b6\u69cb\u4e4b\u6642\u5e8f\u95dc\u4fc2\u504f\u597d\u5ea6\u6bd4\u8f03 \u69cb(\u50b3\u7d71Two-Stage\u8a9e\u97f3\u5408\u6210\u7cfb\u7d71)\u8207DNNP+DNNS\u865f\u67b6\u69cb(\u65b0\u7684DNN\u67b6\u69cb\uff0c\u524d\u7d1a\u63db\u6210 \u5f8c\u7d1a\u5247\u4f7f\u7528 DNNs \u4f86\u505a\u8a9e\u97f3\u5408\u6210\uff0c\u9019\u662f\u521d\u6b65\u5617\u8a66\uff0c\u4ee5\u5f8c\u6703\u5efa\u4e00\u500b\u5927\u7db2\u8def\uff0c\u5305\u542b\u6240 \u5716 11 Parser+HTS \u67b6\u69cb\u8207 Parser+DNNS c.RNNLM \u5be6\u9a57 \u886812 \u65b0\u820aG2P\u4e4bMOS\u4e3b\u89c0\u5206\u6578\u6bd4\u8f03 48% 52% \u81ea\u7136\u5ea6\u504f\u597d \u6574\u9ad4\u65b0\u820a\u7cfb\u7d71\u7684\u6548\u80fd\u6bd4\u8f03\u70ba\u672c\u8ad6\u6587\u7684\u91cd\u9ede\uff0c\u5728\u6b64\u6211\u5011\u9032\u4e00\u6b65\u6bd4\u8f03Parser+HTS\u67b6 \u5728\u672c\u8ad6\u6587\u4e2d\uff0c\u6211\u5011\u5c07\u50b3\u7d71\u524d\u7d1a\u6587\u672c\u5206\u6790\u62c6\u6210 DNNG\u3001DNNc \u8207 DNNT \u4e09\u500b\u90e8\u5206\uff0c \u88682 G2P\u8a13\u7df4\u8a9e\u6599\u5eab \u88685\u5247\u70ba\u5408\u6210\u8a9e\u6599\u7684\u8cc7\u6599\u3002 \u88684 \u8a13\u7df4\u8a9e\u6599\u8cc7\u6599\u8868 \u4e2d\u6587\u8a9e\u6599 \u4e2d\u82f1\u593e\u96dc\u8a9e\u6599 \u82f1\u6587\u8a9e\u6599 \u8a9e\u6599\u5167\u5bb9\u51fa\u8655 \u751f\u547d\u79d1\u5b78\u5927\u5e2b\uff1a\u907a\u50b3\u5b78\u4e4b\u7236 \u5b5f\u5fb7\u723e\u7684\u6545\u4e8b(\u5f35\u6587\u4eae\u8457) \u7dda\u4e0a\u6587\u672c (\u5de5\u7814\u9662\u63d0\u4f9b) CMU \u8a9e\u6599\u53e5\u6578 \u7d04 4800 \u53e5 \u7d04 3500 \u53e5 \u7d04 990 \u53e5 \u6bcf\u53e5\u8a5e\u6578 20-35 \u8a5e 10-30 \u8a5e 5-15 \u500b\u55ae\u5b57 \u6642\u9593\u9577\u5ea6 \u7d04 170 \u5206\u9418 \u7d04 200 \u5206\u9418 \u7d04 79 \u5206\u9418 \u88685 \u5408\u6210\u8a9e\u6599\u8cc7\u6599\u8868 \u4e2d\u6587\u8a9e\u6599 \u4e2d\u82f1\u593e\u96dc\u8a9e\u6599 \u8a9e\u6599\u5167\u5bb9\u51fa\u8655 \u751f\u547d\u79d1\u5b78\u5927\u5e2b\uff1a\u907a\u50b3\u5b78\u4e4b\u7236\u5b5f\u5fb7 \u723e\u7684\u6545\u4e8b(\u5f35\u6587\u4eae\u8457) \u7dda\u4e0a\u6587\u672c (\u5de5\u7814\u9662\u63d0\u4f9b) \u6e2c\u8a66\u8a9e\u6599\u97f3\u6a94\u6578 2 \u500b 2 \u500b \u6e2c\u8a66\u8a9e\u6599\u7e3d\u53e5\u6578 192 \u53e5 100 \u53e5 \u6e2c\u8a66\u8a9e\u6599\u6bcf\u53e5\u5b57\u6578 \u4f9d\u6587\u7ae0\u70ba\u6e96 10-30 \u5b57 \u662f2\u90781\u7684\u65b9\u5f0f\uff0c\u70ba\u6a19\u6e96\u7684A/B/X\u6e2c\u8a66\uff0c\u65b0\u820a\u7cfb\u7d71\u8acb\u6bcf\u4eba\u5404\u807d10\u53e5\uff1b\u800c\u5e73\u5747\u4e3b\u89c0\u503c \u5206\u6578\u8acb\u6bcf\u4eba\u5404\u807d20\u53e5\uff0c\u8a55\u5206\u65b9\u5f0f\u70ba1~5\u5206\uff0c\u5206\u6578\u8d8a\u9ad8\u5247\u70ba\u8d8a\u597d\u3002 (\u4e09)\u3001\u8072\u97f3\u5408\u6210\u5be6\u9a57\u7d50\u679c 1\u3001\u524d\u7d1a\u6587\u672c\u5206\u6790\u5be6\u9a57\u7d50\u679c a.G2P \u70ba\u4e86\u8a55\u4f30\u6211\u5011DNNG\u4f7f\u7528Seq2Seq\u7684G2P\u80fd\u5426\u66ff\u63db\u6389\u539f\u672c\u7684\u5b57\u8f49\u97f3\u65b9\u5f0f\uff0c\u6211\u5011\u4ee5 \u4e2d\u82f1\u593e\u96dc\u6587\u5b57\u8a9e\u6599(\u5de5\u7814\u9662\u63d0\u4f9b\u4e4b\u7dda\u4e0a\u6587\u672c)\u3001\u5927\u9678\u6587\u7ae0(Blizzard Challenge 2010\u7684 \u6e2c\u8a66\u8a9e\u6599)\u8207\u64f7\u53d6\u65bc\u570b\u5bb6\u6587\u5b78\u535a\u58eb/\u570b\u7acb\u5e2b\u5927\u6559\u6388\u8a31\u931f\u8f1d\u4e3b\u7de8\u7684\u300c\u5e38\u898b\u7834\u97f3\u5b57\u300d\u4e00 \u66f8\uff0c\u548c<\u859b\u610f\u6885>\u5e38\u7528\u7684100\u500b\u7834\u97f3\u5b57\u4e2d\u7684\u7834\u97f3\u5b57\u96c6\u4f86\u505a\u6bd4\u8f03\uff0c\u770b\u4e0d\u540c\u60c5\u6cc1\u4e0b\u65b0\u820a G2P\u5404\u81ea\u7684\u6b63\u78ba\u7387\u70ba\u4f55\u3002\u5be6\u9a57\u7d50\u679c\u5982\u88687\u3001\u88688\u8207\u88689\u6240\u793a\u3002 \u88687 \u4e2d\u82f1\u593e\u96dc\u6587\u5b57\u8a9e\u6599\u65b0\u820aG2P\u7684\u6b63\u78ba\u7387 \u4e00\u822c\u6587\u7ae0 \u7e3d\u5171 450 \u53e5(\u7d04 7330 \u5b57) Accuracy \u77e5\u5c07\u5b57\u5143\u8f49\u63db\u5230\u5411\u91cf\u7a7a\u9593\u8655\u7406\u5206\u6790\uff0c\u53ef\u4ee5\u6709\u6548\u7684\u5c07\u5b57\u5143\u8a9e\u610f\u3001\u5b57\u5143\u6587\u6cd5\u89d2\u8272\u7b49\u5c6c \u886810 character embedding \u7522\u751f\u7684\u5b57\u5206\u985e \u5206\u985e \u5b57\u5143 1 2 3 \u7d6e\u3001\u7d79\u3001\u7da2\u3001\u7db4\u3001\u7dbf\u3001\u7dde\u3001\u7dfb\u3001\u7e37\u3001\u7e61\u3001\u7e6a 4 \u3044\u3001\u304c\u3001\u3068\u3001\u304d\u3001\u308a\u3001\u3067\u3001\u308b\u3001\u304f\u3001\u3057\u3001\u3082 5 \u500b\u3001\u5206\u3001\u544e\u3001\u54e9\u3001\u5c3a\u3001\u65a4\u3001\u91d0\u3001\u79d2\u3001\u5e74\u3001\u9803 \u57168 \u65b0\u820aG2P\u504f\u597d\u5ea6\u6bd4\u8f03 47% 53% 50% 48% 50% 44% 56% \u81ea\u7136\u5ea6\u504f\u597d \u4e94\u3001\u7d50\u8ad6 c\u3001Parser+HTS\u8207DNNp\uff0bDNNs\u8a9e\u97f3\u5408\u6210\u504f\u597d\u5ea6\u6bd4\u8f03 \u76f8\u4f3c\u5ea6\u504f\u597d 52% 48% \u76f8\u4f3c\u5ea6\u504f\u597d \u571614 Parser+DNNS\u8207DNNP+DNNS\u67b6\u69cb\u504f\u597d\u5ea6\u6bd4\u8f03 52% \u53ef\u7406\u89e3\u5ea6\u504f\u597d 48% 52% \u53ef \u7406 \u89e3 \u5ea6 \u504f \u597d \u81ea\u7136\u5ea6\u8a55\u5206 2.85 3.05 3.1 3.23 \u81ea\u7136\u5ea6\u504f\u597d \u58be\u3001\u58e9\u3001\u5ca9\u3001\u5cb8\u3001\u5cf6\u3001\u5cfd\u3001\u5d17\u3001\u5dba\u3001\u5dbc\u3001\u5dd2 56% 44% Parser rnnlm Parser+HTS Parser+DNNs 41% 59% \u81ea\u7136\u5ea6\u504f\u597d \u76f8\u4f3c\u5ea6\u8a55\u5206 3.05 3.21 2.93 3.1 \u76f8\u4f3c\u5ea6\u504f\u597d \u5f0a\u3001\u6582\u3001\u6848\u3001\u6d89\u3001\u7006\u3001\u7591\u3001\u8490\u3001\u8caa\u3001\u8cc2\u3001\u8cc4 \u820a\u7cfb\u7d71\u5169\u8005\u5176\u5be6\u76f8\u5dee\u4e0d\u6703\u5f88\u591a\u3002 44% 56% \u53ef\u7406\u89e3\u5ea6\u504f\u597d \u820aG2P \u65b0G2P \u6642\u5e8f\u95dc\u4fc2\u7684\u5be6\u9a57\u7d50\u679c\u6bd4\u8f03\u986f\u793a\u5728\u571610\u8207\u886814\u4e2d\u3002\u7e3d\u9ad4\u4f86\u8aaa\u4e5f\u662f\u6539\u7528RNNLM\u7684\u8cc7\u8a0a \u7684\u65b0\u67b6\u69cb\u6bd4\u50b3\u7d71\u4f7f\u7528Parser\u7684\u820a\u67b6\u69cb\u7a0d\u597d\u4e00\u4e9b\u3002 \u8072\u97f3\u6703\u7a0d\u5fae\u81ea\u7136\u4e00\u4e9b\uff0c\u7279\u5225\u662f\u5728\u82f1\u6587\u4e0a\uff0c\u4f46\u5dee\u7570\u4e0d\u5927\u3002 46% \u53ef\u7406\u89e3\u5ea6\u8a55\u5206 3.28 3.3 3.56 3.62 54% \u76f8\u4f3c\u5ea6\u504f\u597d \u4e2d\u82f1\u593e\u96dc Parser+HTS Parser+DNNS DNNP+HTS DNNP+DNNS \u5be6\u9a57\u7d50\u679c\u5982\u571611\u548c12\u6240\u793a\u3002\u5f9e\u571611\u548c12\u53ef\u4ee5\u770b\u51fa\u5728\u8aaa\u8a71\u81ea\u7136\u5ea6\u4e0a\u4ee5DNN\u5408\u6210\u7684 42% 58% \u53ef\u7406\u89e3\u5ea6\u504f\u597d c\u3001\u66ff\u63dbParser\u6642\u5e8f\u4f4d\u7f6e\u8cc7\u8a0a\u8207RNNLM\u5f71\u97ff\u4e4b\u6bd4\u8f03 DNNp+DNNs\u67b6\u69cb\u6bd4\u8f03\uff0c\u4f86\u770b\u4ee5\u50b3\u7d71HMM\u548c\u4ee5DNN\u4f5c\u8a9e\u97f3\u5408\u6210\u7684\u5dee\u7570\u6027\u3002 Parser+DNNs \u886815 \u56db\u7a2e\u67b6\u69cb\u69cb\u4e2d\u82f1\u593e\u96dc\u8072\u97f3MOS\u4e3b\u89c0\u5206\u6578\u6bd4\u8f03 DNNp+DNNs \u6027\u64f7\u53d6\u51fa\u4f86\u3002 \u5169\u8005\u7684\u504f\u597d\u5ea6\u8207MOS\u5206\u6578\u3002 \u5be6\u9a57\u7d50\u679c\u986f\u793a\u5728\u57168\u8207\u886812\u4e2d\uff0c\u6211\u5011\u53ef\u4ee5\u770b\u898b\u55ae\u55ae\u53ea\u6709\u63db\u6389G2P\u5143\u4ef6\u7684\u6642\u5019\uff0c\u65b0 \u53ef\u7406\u89e3\u5ea6\u8a55\u5206 3 3.2 \u76f8\u4f3c\u5ea6\u8a55\u5206 3.1 \u81ea\u7136\u5ea6\u8a55\u5206 3 3 \u6b64 \u90e8 \u4efd \u6211 \u5011 \u4ee5 \u6bd4 \u8f03 Parser+HTS \u8207 Parser+DNNs \u67b6\u69cb\u8ddf\u6bd4\u8f03DNNp+HTS \u8207 \u81ea\u7136\u5ea6\u8a55\u5206 2.95 3.14 3.16 3.13 \u571613 Parser+hts\u8207DNNP+HTS\u67b6\u69cb\u504f\u597d\u5ea6\u6bd4\u8f03 \u76f8\u4f3c\u5ea6\u8a55\u5206 3.01 3.1 3.05 3.1 3 a\u3001DNN vs HMM\u8a9e\u97f3\u5408\u6210\u504f\u597d\u5ea6\u6bd4\u8f03 \u53ef\u7406\u89e3\u5ea6\u8a55\u5206 3.38 3.47 3.6 3.56 Wikipedia\u6587\u5b57\u8a9e\u6599\u5eab\uff0c\u5c07\u5b57\u5143\u8f49\u6210\u5411\u91cf\u5f8c\uff0c\u518d\u4ee5\u5b57\u5143\u5411\u91cf\u4f5c\u5206\u7fa4\u7684\u7d50\u679c\u3002\u7531\u6b64\u53ef \u5927\u7684\u5dee\u7570\u6027\uff0c\u4ee5\u53ca\u66ff\u63db\u5f8c\u7684\u8072\u97f3\u5408\u4e0d\u5408\u7406\uff0c\u5728\u6b64\u6211\u5011\u4f7f\u7528\u5c11\u91cf\u5408\u6210\u97f3\u6a94\u6e2c\u8a66\u4e00\u4e0b \u7d14\u4e2d\u6587 Parser \uf06c\uf020DNNP+DNNS= \u524d\u7d1aDNNG + DNNC + DNNT\uff0c\u5f8c\u7d1aHTS\u7684DNN(End-to-End)\u3002 41% 59% \u81ea\u7136\u5ea6\u504f\u597d \u7d14\u4e2d\u6587 Parser+HTS Parser+DNNS DNNP+HTS DNNP+DNNS Word2vec \u5408 \u6211 \u5011 \u7684 \u9810 \u671f \uff0c \u8868 10 \u70ba\u5229\u7528Word2vec \u5206\u6790Chinese Gigaword Second Edition + \u5728\u6b64\u5148\u55ae\u7368\u63a2\u8a0e\u55ae\u7d14\u7684G2P\u5143\u4ef6\uff0c\u4f7f\u7528Seq2Seq\u7684G2P\u6703\u4e0d\u6703\u8207\u50b3\u7d71\u7684G2P\u6709\u5f88 \u886813 \u65b0\u820a\u6587\u6cd5\u89d2\u8272\u4e4bMOS\u4e3b\u89c0\u5206\u6578 44% 56% \u76f8\u4f3c\u5ea6\u504f\u597d \u886814 \u56db\u7a2e\u67b6\u69cb\u7684\u4e2d\u6587\u8072\u97f3MOS\u4e3b\u89c0\u5206\u6578\u6bd4\u8f03 \uf06c\uf020DNNP+HTS = \u524d\u7d1aDNNG + DNNC + DNNT\uff0c\u5f8c\u7d1aHTS\u3002 \u5728DNNc\u8207DNNT\u9019\u90e8\u5206\u6211\u5011\u5206\u5225\u505a\u5404\u81ea\u7684\u5be6\u9a57\uff0c\u5f9e\u5be6\u9a57\u7d50\u679c\u4f86\u770b\u5230\u5e95\u6709\u6c92\u6709\u7b26 a\u3001\u66ff\u63db\u820aG2P\u8207\u65b0G2P\u8a9e\u97f3 \uf06c\uf020Parser+DNNS= \u524d\u7d1aParser\uff0c\u5f8c\u7d1aHTS\u7684DNN\u3002 39% 61% 13\u548c14\u4f86\u770bDNNp+DNNs\u67b6\u69cb\uff0c\u6240\u5408\u6210\u7684\u8072\u97f3\u78ba\u5be6\u6bd4\u8f03\u597d\u3002 \u53ef\u7406\u89e3\u5ea6\u504f\u597d b. Word2vec\u5be6\u9a57 \u8f38\u51fa\u4e4b\u96b1\u85cf\u5c64\u72c0\u614b\uff0c\u4f86\u505a\u6574\u9ad4\u8a9e\u97f3\u5408\u6210\u8a55\u4f30\u3002 \u57169 \u65b0\u820a\u67b6\u69cb\u4e4b\u8a9e\u610f\u6587\u6cd5\u504f\u597d\u5ea6\u6bd4\u8f03 Parser+HTS DNNp+HTS \u67b6\u69cb\u8207(4)DNNp+DNNs\u67b6\u69cb\uff0c\u5728\u7d14\u4e2d\u6587\u53ca\u4e2d\u82f1\u593e\u96dc\u8a9e\u97f3\u5408\u6210\u7684MOS\u5206\u6578\uff0c\u7531\u8868 word2vec[3]\u80fd\u5920\u5c07\u8f38\u5165\u7684\u5b57\u8a5e\u8f49\u63db\u5230\u5411\u91cf\u7a7a\u9593\u9032\u884c\u8a08\u7b97\uff0c\u5206\u6790\u5f8c\u53ef\u4ee5\u5728\u5411\u91cf \u53e5\u5b50/\u55ae\u5b57 \u62fc\u97f3 \u5404\u884c\u5404\u696d\u7686\u986f\u8457\u7684\u6539\u5584\u4e0d xian3 \u884c\u5404\u696d\u7686\u986f\u8457\u7684\u6539\u5584\u4e0d\u5c11 zhu4 \u5404\u696d\u7686\u986f\u8457\u7684\u6539\u5584\u4e0d\u5c11, de0 \u25cb\u25cb\u25cb\u25cb\u25cb\u9694\u95a1\u25cb\u25cb\u25cb\u25cb ge2 \u25cb\u25cb\u25cb\u25cb\u9694\u95a1\u25cb\u25cb\u25cb\u25cb\u25cb he2 \u25cb\u25cb\u25cb\u25cb\u25cb\u4e09\u25cb\u25cb\u25cb\u25cb\u25cb san1 (\u4e09)\u3001\u5b57\u5143\u6642\u5e8f\u72c0\u614b(DNNT) \u5728\u6211\u5011\u7684\u67b6\u69cb\u4e0b\uff0c\u6211\u5011\u5e0c\u671b\u80fd\u4e0d\u4f7f\u7528 parser \u5c31\u5f9e\u6587\u672c\u7372\u5f97\u6bcf\u500b\u5b57\u5143\u5728\u7576\u524d\u8a9e\u53e5 \u7684\u72c0\u614b\uff0c\u8b93\u6a5f\u5668\u81ea\u52d5\u5b78\u7fd2\u6587\u7ae0\u7684\u8108\u7d61\uff0c\u9032\u800c\u80fd\u5f9e\u76ee\u524d\u8a9e\u53e5\u9810\u6e2c\u5230\u4e0b\u4e00\u53e5\u53ef\u80fd\u70ba\u4f55\u3002 \u905e \u8ff4 \u795e \u7d93 \u7db2 \u8def \u5728 Distributed Representations of Words and Phrases and their Compositionality [4]\u6b64\u7bc7\u8ad6\u6587\u4e5f\u6307\u51fa\u4f7f\u7528\u905e\u8ff4\u795e\u7d93\u7db2\u8def\u6a21\u578b\u9032\u884c\u8a13\u7df4\u80fd\u5f9e\u96b1\u85cf\u5c64 \u4e2d\u7684\u9023\u7e8c\u8f38\u51fa\u5411\u91cf\u7372\u5f97\u5b57\u8a5e\u5728\u8a9e\u53e5\u4e2d\u7684\u72c0\u614b\u3002 \u672c\u6587\u6240\u4f7f\u7528\u7684\u662fMikolov\u7684RNNLM\uff0c\u4e0d\u904e\u6211\u5011\u662f\u4f7f\u7528\u5b57\u5143\u968e\u5c64\u4f86\u9032\u884c\u8a13\u7df4\u3002 \u6703\u5148\u5f9e\u8072\u97f3\u8a9e\u6599\u4e2d\u64f7\u53d6Frame-by-Frame\u7684\u8072\u97f3\u7279\u5fb5\u53c3\u6578\uff0c\u5148\u524d3\u500bDNN\u6240\u64f7\u53d6\u7684\u8cc7 \u8a0a\u4e5f\u6703\u505a\u6210Frame-by-Frame\u7684\u683c\u5f0f\uff0c\u518d\u4ee5\u5982\u57167\u6240\u793a\u7684\u67b6\u69cb\uff0c\u9032\u884c\u548c\u6210\u6a21\u578b\u8a13\u7df4\u3002 Frame-by-Frame \u7684\u6587\u8108\u8a0a\u606f \u2026 \u2026 \u2026 \u2026 Frame-by-Frame \u7684\u8072\u5b78\u7279\u5fb5 \u88683 word2vec\u8207RNNLM\u8a13\u7df4\u8a9e\u6599 Chinese Gigaword Second Edition wikipedia \u8a9e\u6599\u6027\u8cea \u4e3b\u984c\u5f0f\u6587\u7ae0 \u540d\u8a5e\u89e3\u91cb \u53e5\u5b50\u7e3d\u6578 \u5171\u7d04 1200 \u842c\u53e5 3\u3001\u8a9e\u97f3\u5408\u6210\u8a9e\u6599 \u672c\u5be6\u9a57\u4f7f\u7528\u7684\u8a9e\u97f3\u5408\u6210\u8a13\u7df4\u8a9e\u6599\uff0c\u662f\u6211\u5011\u8207\u53f0\u7063\u6578\u4f4d\u6709\u8072\u66f8\u5354\u6703\u5408\u4f5c\u9304\u88fd \u7684\"NTUT Audiobook Corpus Vol.2\"\u3002\u5408\u6210\u7684\u6e2c\u8a66\u8a9e\u6599\u5247\u662f\u5f9e\u4e2d\u62bd\u53d6\u4e2d\u6587100\u53e5\u53ca\u4e2d \u82f1\u593e\u96dc100\u53e5\u4f86\u4f5c\u5408\u6210\uff0c\u62bd\u51fa\u7684\u53e5\u5b50\u7686\u4e0d\u5728\u8a13\u7df4\u8a9e\u6599\u4e4b\u4e2d\u3002\u88684\u70ba\u8a13\u7df4\u8a9e\u6599\u8cc7\u6599\u8868\uff0c \u65b0\u65b9\u6cd5\u8207\u820a\u65b9\u6cd5\u4e2d\u6574\u500b\u524d\u7d1a\u6587\u672c\u5206\u6790\u5b8c\u5168\u4e0d\u540c\uff0c\u820a\u6587\u8108\u8a0a\u606f\u4f9d\u7136\u63a1\u7528Parser\u4f86 \u9032\u884c\u6587\u672c\u5206\u6790\u548c\u67e5\u8868\u5b57\u8f49\u97f3\uff1b\u65b0\u7cfb\u7d71\u5247\u662f\u5c07\u820a\u7cfb\u7d71\u6c42\u53d6\u6587\u8108\u8a0a\u606f\u7684\u65b9\u5f0f\u90fd\u53bb \u9664\uff0c\u5b57\u8f49\u97f3\u90e8\u5206\u63a1\u7528DNNG\u8f49\u63db\u5b57\u8a5e\u62fc\u97f3\u8a9e\u8abf\uff0c\u539f\u672cParser\u90e8\u5206\u5247\u662f\u4f7f\u7528DNNc\u8207 DNNT\u4f86\u5206\u6790\uff0c\u4e26\u5229\u7528DNNT\u7684\u96b1\u85cf\u5c64\u72c0\u614b\u4f86\u4ee3\u8868\u5b57\u5143\u5728\u8a9e\u53e5\u4e2d\u7684\u6642\u5e8f\u95dc\u4fc2\u3002 (\u4e8c)\u3001\u8a55\u4f30\u65b9\u6cd5 \u8a55\u4f30\u6e2c\u8a66\u5305\u62ec\u5408\u6210\u97f3\u6a94\u53ef\u7406\u89e3\u5ea6\uff0c\u76f8\u4f3c\u5ea6\u8207\u81ea\u7136\u5ea6\u7684\u504f\u597d\u5ea6\u8207MOS\u4e3b\u89c0\u5206 \u6578\uff0c\u6211\u5011\u5c07\u6e2c\u8a66\u97f3\u6a94\u7d6610\u4f4d\u6bcd\u8a9e\u70ba\u570b\u8a9e\u7684\u4eba\u58eb\u9032\u884c\u8a55\u5206\uff0c\u65b0\u820a\u7cfb\u7d71\u504f\u597d\u5ea6\u6e2c\u8a66 \u88689 \u7834\u97f3\u5b57\u53e5\u5b50\u4e2d\u65b0\u820aG2P\u7684\u7834\u97f3\u5b57\u6b63\u78ba\u7387 \u7834\u97f3\u5b57\u53e5\u5b50 \u7e3d\u5171 227 \u53e5(7988 \u5b57,\u7834\u97f3\u5b57 509 \u5b57) \u820a G2P \u932f 65 \u5b57 87.22% \u65b0 G2P \u932f 58 \u5b57 88.6% \u7ae0\u4e2d\uff0c\u65b0G2P\u7684\u8868\u73fe\u4e26\u6c92\u6709\u5f88\u597d\uff1b\u800c\u5728\u7834\u97f3\u5b57\u6e2c\u8a66\u4e2d\uff0c\u65b0G2P\u7a0d\u5fae\u597d\u4e9b\uff0c\u7e3d\u4e4b\u65b0\u7684 \u6539\u5584\u6b63\u78ba\u7387\u3002 \uf06c\uf020Parser[7][8]+HTS = \u524d\u7d1aParser\uff0c\u5f8c\u7d1aHTS(\u50b3\u7d71Two-Stage\u8a9e\u97f3\u5408\u6210\u7cfb\u7d71)\u3002 \u820aG2P\u8207\u65b0G2P\uff0c( 2) ParserPOS\u8cc7\u8a0a\u8207Word2vec\uff0c( 3) Parser\u65b7\u8a5e\u6642\u5e8f\u4f4d\u7f6e\u8207RNNLM 48% 52% \u81ea\u7136\u5ea6\u504f\u597d \u524d\u7d1a\u63db\u6210DNNp\u5c0d\u8a9e\u97f3\u5408\u6210\u5f71\u97ff\u76f8\u7576\u660e\u986f\u3002 \u886814\u820715\u5247\u5206\u5225\u70ba(1)Parser+HTS\u67b6\u69cb\u3001 (2) Parser+DNNs\u67b6\u69cb\u3001 (3) DNNp+HTS \u4ee5\u4e0b\u56db\u7a2e\u67b6\u69cb\uff0c\u63a2\u8a0e\u524d\u5f8c\u7d1a\u7684\u4e0d\u540c\u7d44\u5408\u6703\u6709\u751a\u9ebc\u5f71\u97ff\uff1a G2P\u6548\u80fd\u8207\u539f\u672c\u820a\u7cfb\u7d71\u7684\u6548\u679c\u76f8\u7576\uff0c\u672a\u4f86\u53ef\u518d\u591a\u6dfb\u52a0\u7834\u97f3\u5b57\u8a13\u7df4\u8a9e\u6599\uff0c\u624d\u9032\u4e00\u6b65 \u7d71 \u7e3d \u674e \u8a9e\u97f3\u5408\u6210\u5be6\u9a57\u7684\u90e8\u5206\u6211\u5011\u900f\u904e\u66ff\u63db\uff0c\u6bd4\u8f03\u5404\u5143\u4ef6\uff0c\u5206\u70ba\u4ee5\u4e0b\u4e09\u90e8\u5206\uff0c\u5305\u62ec(1) 54% 46% \u76f8\u4f3c\u5ea6\u504f\u597d \u4f7f\u7528DNNP\u7684\u67b6\u69cb\u4e0d\u7ba1\u662f\u5c0dHMM\u6216\u662fDNNs\u90fd\u6703\u6bd4\u8f03\u597d\uff0c\u800c\u4e14\u6709\u76f8\u7576\u5dee\u8ddd\u3002\u986f\u7136\u5c07 d\u3001\u56db\u7a2e\u67b6\u69cb\u7684MOS\u4e3b\u89c0\u5206\u6578\u6bd4\u8f03 \u5728\u8072\u97f3\u5408\u6210\u5be6\u9a57\u7d50\u679c\u90e8\u5206\uff0c\u6211\u5011\u63a1\u7528\u504f\u597d\u8a55\u6bd4\u4ee5\u53ca\u5e73\u5747\u4e3b\u89c0\u503c\u5206\u6578\uff0c\u5206\u5225\u6e2c\u8a66 2\u3001\u66ff\u63db\u6587\u5b57\u5206\u6790\u6a21\u7d44\uff0c\u5c0d\u8a9e\u8a00\u5408\u6210\u7684\u5f71\u97ff 46% 54% \u53ef\u7406\u89e3\u5ea6\u504f\u597d \u571613\u548c14\u70ba\u5be6\u9a57\u7d50\u679c\uff0c\u53ef\u4ee5\u770b\u51fa\u6587\u672c\u5206\u6790\u5c0d\u5408\u6210\u8a9e\u97f3\u7684\u5f71\u97ff\u3002\u5be6\u9a57\u7d50\u679c\u986f\u793a\uff0c \u571615 Parser+HTS\u67b6\u69cb\u8207DNNp+DNNs\u67b6\u69cb\u504f\u597d\u5ea6\u6bd4\u8f03 \u9023\u63a5\uff0c\u5982\u4f55\u6dfb\u52a0 label \u4e2d\u7684\u6587\u8108\u8cc7\u8a0a\u3002 \u7531\u88687\u30018\u30019\u53ef\u4ee5\u770b\u898b\u5728\u4e00\u822c\u6587\u7ae0\u4e2d\uff0c\u65b0G2P\u8207\u820aG2P\u5dee\u8ddd\u4e0d\u5927\uff0c\u4f46\u662f\u5728\u5927\u9678\u6587 \u80fd \u6703 \u5c0d \u6211 \u5011 \u7684 \u58d3 \u529b . Backward . \u9047 \u5f85 \u570b \u60e0 \u6700 \u7684 \u570b \u7f8e \u958b \u96e2 \u5b9a \u6c7a \u5df2 , \u5224 \u8ac7 \u7684 \u570b \u7f8e \u8207 \u570b \u5408 \u806f \u5728 \u570b \u7f8e \u517c \u7406 \u7e3d \u526f \u570b \u7f8e , \u793a \u8868 \u9ad4\u4f86\u8aaa\u9084\u662f\u6539\u7528Word2vec\u7684\u8cc7\u8a0a\u5f8c\u7684\u65b0\u67b6\u69cb\u6bd4\u50b3\u7d71\u4f7f\u7528Parser\u7684\u820a\u67b6\u69cb\u7a0d\u597d\u4e00\u4e9b\u3002 Parser Word2vec \u5408\u6210\u8072\u97f3\u7684\u5f71\u97ff\u5dee\u7570\u3002 \u7576\u6587\u8108\u8cc7\u8a0a\uff0c\u5efa\u7acb\u5f8c\u7d1a\u8981\u7528\u7684 frame-by-frame \u5408\u6210\u8cc7\u8a0a\u3002\u4e3b\u8981\u662f\u524d\u7d1a\u8ddf\u5f8c\u7d1a\u5982\u4f55 30% 70% \u81ea\u7136\u5ea6\u504f\u597d Parser+DNNS\u8207DNNP+DNNS\u67b6\u69cb\uff0c\u4f86\u4e86\u89e3\u4ee5Parser\u65b9\u5f0f\u548cDNNP\u65b9\u5f0f\u6c42\u53d6\u6587\u8108\u5c0d\u8a9e\u8a00 \u524d\u662f\u5148\u7528 HTS \u7684 duraiton model \u4f30\u7b97\u5408\u6210\u9577\u5ea6\u3002\u518d\u52a0\u4e0a\u64f7\u53d6\u524d\u7d1a\u4e09\u500b\u7db2\u8def\u7684\u8f38\u51fa 39% 61% \u76f8\u4f3c\u5ea6\u504f\u597d \u7834\u97f3\u5b57 Accuracy \u6642\u5e8f\u65b9\u5411 \u7bc4\u4f8b Forward \u5728 \u6c11 \u4e3b \u9ee8 \u7684 \u8868 \u73fe , \u4ed6 \u5011 1 \u81f4 \u8a8d \u70ba , 1 \u5207 \u4e0d \u53ef \u5728\u6bd4\u8f03\u65b0\u820a\u7cfb\u7d71\u7684POS\u8cc7\u8a0a\u6642\uff0c\u56e0\u70baG2P\u9084\u662f\u5fc5\u9700\u8981\u6709\u7684\uff0c\u6240\u4ee5\u5728\u6bd4\u8f03\u4e0a\u53ef\u80fd\u6703\u53d7 \u5230\u65b0\u820aG2P\u548c\u65b0\u820a\u6587\u6cd5\u8173\u8272\u8cc7\u8a0a\u7684\u4e92\u76f8\u62c9\u626f\uff0c\u4e0d\u904e\u7531\u57169\u8207\u886813\u7684\u5be6\u9a57\u7d50\u679c\u4f86\u770b\uff0c\u7e3d 3\u3001\u6574\u9ad4\u65b0\u820a\u7cfb\u7d71\u67b6\u69cb\u8072\u97f3\u5408\u6210\u504f\u597d\u5ea6\u6bd4\u8f03 \u56e0\u70ba\u6211\u5011\u9084\u7f3a\u5c11 duration model\uff0c\u9084\u4e0d\u80fd\u76f4\u63a5\u7b97\u51fa\u6bcf\u500b\u8072\u97f3\u8981\u5408\u6210\u591a\u9577\u3002\u6240\u4ee5\u76ee \u5728\u6587\u672c\u5206\u6790\u65b9\u5f0f\u7684\u504f\u597d\u6bd4\u8f03\u4e2d\uff0c\u6211\u5011\u6bd4\u8f03Parser+HTS\u8207DNNP+HTS\u67b6\u69cb\uff0c\u8ddf\u6bd4\u8f03 28% \u597d 72% \u53ef\u7406\u89e3\u5ea6\u504f b\u3001Parser vs DNNP\u8a9e\u97f3\u5408\u6210\u504f\u597d\u5ea6\u6bd4\u8f03 Parser+HTS Dnnp+DNNs
", "num": null, "type_str": "table" } } } }