{"id":23204,"date":"2024-12-16T18:14:23","date_gmt":"2024-12-16T10:14:23","guid":{"rendered":"http:\/\/139.9.1.231\/?p=23204"},"modified":"2024-12-16T18:14:24","modified_gmt":"2024-12-16T10:14:24","slug":"tts-summary","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2024\/12\/16\/tts-summary\/","title":{"rendered":"TTS\u8c03\u7814 | \u8bed\u97f3\u5408\u6210\u7cfb\u5217\u57fa\u7840\u77e5\u8bc6\u53ca\u8bba\u6587\u603b\u7ed3"},"content":{"rendered":"\n

\u539f\u521b\u00a0AI Pulse<\/p>\n\n\n\n

Text-to-Speech\uff08\u901a\u5e38\u7f29\u5199\u4e3aTTS\uff09\u662f\u6307\u4e00\u79cd\u5c06\u6587\u672c\u8f6c\u4e3a\u97f3\u9891\u7684\u6280\u672f\u3002<\/p>\n\n\n\n\n\n

1.\u5386\u53f2<\/strong><\/h2>\n\n\n\n

\u7b2c\u4e00\u53f0\u201c\u4f1a\u8bf4\u8bdd\u7684\u673a\u5668\u201d\u53ef\u80fd\u662f\u5728 18 \u4e16\u7eaa\u540e\u671f\u5236\u9020\u7684\uff08\u636e\u8bf4\u662f\u4e00\u4f4d\u5308\u7259\u5229\u79d1\u5b66\u5bb6\u53d1\u660e\u7684\uff09\u3002\u8ba1\u7b97\u673a\u8f85\u52a9\u521b\u4f5c\u8d77\u6e90\u4e8e20\u4e16\u7eaa\u4e2d\u671f\uff0c\u5404\u79cd\u6280\u672f\u5df2\u7ecf\u4f7f\u7528\u4e86\u5927\u7ea650\u5e74\u3002\u5982\u679c\u5bf9\u65e7\u6280\u672f\u8fdb\u884c\u5206\u7c7b.\u9996\u5148\uff0c<\/p>\n\n\n\n

1\uff09Articulatory Synthesis\uff1a\u8fd9\u662f\u4e00\u79cd\u6a21\u62df\u4eba\u7684\u5634\u5507\u3001\u820c\u5934\u548c\u53d1\u58f0\u5668\u5b98\u7684\u6280\u672f\u3002<\/p>\n\n\n\n

2\uff09\u5171\u632f\u5cf0\u5408\u6210\uff1a\u4eba\u58f0\u53ef\u4ee5\u770b\u4f5c\u662f\u5728\u8bed\u97f3\u5728\u5668\u5b98\u4e2d\u8fc7\u6ee4\u67d0\u4e9b\u58f0\u97f3\u800c\u4ea7\u751f\u7684\u58f0\u97f3\u3002\u8fd9\u5c31\u662f\u6240\u8c13\u7684\u6e90\u6ee4\u6ce2\u5668\u6a21\u578b\uff0c\u5b83\u662f\u4e00\u79cd\u5728\u57fa\u672c\u58f0\u97f3\uff08\u4f8b\u5982\u5355\u4e2a\u97f3\u9ad8\uff09\u4e0a\u6dfb\u52a0\u5404\u79cd\u6ee4\u6ce2\u5668\u4ee5\u4f7f\u5176\u542c\u8d77\u6765\u50cf\u4eba\u58f0\u7684\u65b9\u6cd5\uff08\u79f0\u4e3a\u52a0\u6cd5\u5408\u6210\uff09\u3002<\/p>\n\n\n\n

3) Concatenative Synthesis\uff1a\u73b0\u5728\u4f7f\u7528\u6570\u636e\u7684\u6a21\u578b\u3002\u4e3e\u4e2a\u7b80\u5355\u7684\u4f8b\u5b50\uff0c\u4f60\u53ef\u4ee5\u5f55\u5236 0 \u5230 9 \u7684\u58f0\u97f3\uff0c\u5e76\u901a\u8fc7\u94fe\u63a5\u8fd9\u4e9b\u58f0\u97f3\u6765\u62e8\u6253\u7535\u8bdd\u53f7\u7801\u3002\u7136\u800c\uff0c\u58f0\u97f3\u5e76\u4e0d\u662f\u5f88\u81ea\u7136\u6d41\u7545\u3002<\/p>\n\n\n\n

4\uff09\u7edf\u8ba1\u53c2\u6570\u8bed\u97f3\u5408\u6210\uff08SPSS\uff09\uff1a\u901a\u8fc7\u521b\u5efa\u58f0\u5b66\u6a21\u578b\u3001\u4f30\u8ba1\u6a21\u578b\u53c2\u6570\u5e76\u4f7f\u7528\u5b83\u6765\u751f\u6210\u97f3\u9891\u7684\u6a21\u578b\u3002\u5b83\u53ef\u4ee5\u5927\u81f4\u5206\u4e3a\u4e09\u4e2a\u90e8\u5206\u3002<\/p>\n\n\n\n

\u9996\u5148\uff0c\u201c\u6587\u672c\u5206\u6790\u201d \uff0c\u5c06\u8f93\u5165\u6587\u672c\u8f6c\u6362\u4e3a\u8bed\u8a00\u7279\u5f81\uff0c\u201c\u58f0\u5b66\u6a21\u578b\u201d \uff0c\u5c06\u8bed\u8a00\u7279\u5f81\u8f6c\u6362\u4e3a\u58f0\u5b66\u7279\u5f81\uff0c\u6700\u540e\u662f\u58f0\u5b66\u7279\u5f81\u3002\u8fd9\u662f\u58f0\u7801\u5668\u3002\u8be5\u9886\u57df\u4f7f\u7528\u6700\u5e7f\u6cdb\u7684\u58f0\u5b66\u6a21\u578b\u662f\u9690\u9a6c\u5c14\u53ef\u592b\u6a21\u578b\uff08HMM\uff09\u3002\u4f7f\u7528 HMM\uff0c\u80fd\u591f\u521b\u5efa\u6bd4\u4ee5\u524d\u66f4\u597d\u7684\u58f0\u5b66\u7279\u5f81\u3002\u4f46\u662f\uff0c\u5927\u90e8\u5206\u751f\u6210\u7684\u97f3\u9891\u6bd4\u8f83\u673a\u68b0\uff0c\u4f8b\u5982\u673a\u5668\u4eba\u58f0\u97f3\u7b49\u3002<\/p>\n\n\n\n

5)\u795e\u7ecf TTS\uff1a\u968f\u7740\u6211\u4eec\u5728 2010 \u5e74\u4ee3\u8fdb\u5165 \u6df1\u5ea6\u5b66\u4e60\u65f6\u4ee3\uff0c\u5df2\u7ecf\u5f00\u53d1\u4e86\u57fa\u4e8e\u51e0\u79cd\u65b0\u795e\u7ecf\u7f51\u7edc\u7684\u6a21\u578b\u3002\u8fd9\u4e9b\u9010\u6e10\u53d6\u4ee3\u4e86HMM\uff0c\u5e76\u88ab\u7528\u4e8e\u201c\u58f0\u5b66\u6a21\u578b\u201d\u90e8\u5206\uff0c\u9010\u6e10\u63d0\u9ad8\u4e86\u8bed\u97f3\u751f\u6210\u7684\u8d28\u91cf\u3002\u4ece\u67d0\u79cd\u610f\u4e49\u4e0a\u8bf4\uff0c\u5b83\u53ef\u4ee5\u770b\u4f5c\u662fSPSS\u7684\u4e00\u6b21\u8fdb\u5316\uff0c\u4f46\u968f\u7740\u6a21\u578b\u6027\u80fd\u7684\u9010\u6e10\u63d0\u9ad8\uff0c\u5b83\u671d\u7740\u9010\u6e10\u7b80\u5316\u4e0a\u8ff0\u4e09\u4e2a\u7ec4\u6210\u90e8\u5206\u7684\u65b9\u5411\u53d1\u5c55\u3002\u6bd4\u5982\u4e0b\u56fe\u4e2d\uff0c\u53ef\u4ee5\u770b\u51fa\u5b83\u662f\u5728\u4ece\u4e0a\uff080\uff09\u5230\u4e0b\uff084\uff09\u7684\u65b9\u5411\u53d1\u5c55\u7684\u3002<\/p>\n\n\n\n

\u73b0\u5728\u63a8\u51fa\u7684\u5927\u81f4\u5206\u4e3a\u4e09\u79cd\u6a21\u578b\uff1a<\/p>\n\n\n\n

\u2728\u58f0\u5b66\u6a21\u578b<\/strong>\uff1a\u4ee5\u5b57\u7b26<\/strong>\uff08\u6587\u672c\uff09\u6216\u97f3\u7d20\uff08\u97f3\u7d20\uff1b\u53d1\u97f3\u5355\u4f4d\uff09\u4e3a\u8f93\u5165\u5e76\u521b\u5efa\u4efb\u4f55\u58f0\u5b66\u7279\u5f81\u7684\u6a21\u578b\u3002\u5982\u4eca\uff0c\u5927\u591a\u6570\u58f0\u5b66\u7279\u5f81\u90fd\u662f\u6307\u6885\u5c14\u9891\u8c31\u56fe\u3002<\/strong><\/p>\n\n\n\n

\u2728\u58f0\u7801\u5668<\/strong>\uff1a\u4e00\u79cd\u5c06\u6885\u5c14\u9891\u8c31\u56fe\uff08\u548c\u7c7b\u4f3c\u7684\u9891\u8c31\u56fe\uff09\u4f5c\u4e3a\u8f93\u5165\u5e76\u751f\u6210\u771f\u5b9e\u97f3\u9891\u7684\u6a21\u578b\u3002<\/p>\n\n\n\n

\u2728\u5b8c\u5168\u7aef\u5230\u7aef\u7684 TTS \u6a21\u578b<\/strong>\uff1a\u63a5\u6536\u5b57\u7b26\u6216\u97f3\u7d20\u4f5c\u4e3a\u8f93\u5165\u5e76\u7acb\u5373\u751f\u6210\u97f3\u9891\u7684\u6a21\u578b\u3002<\/p>\n\n\n\n

2.\u6587\u672c\u5206\u6790<\/strong><\/h2>\n\n\n\n

\u6587\u672c\u5206\u6790\u662f\u5c06\u5b57\u7b26\u6587\u672c\u8f6c\u6362\u4e3a\u8bed\u8a00\u7279\u5f81\u3002\u8981\u8003\u8651\u4ee5\u4e0b\u95ee\u9898\uff1a<\/p>\n\n\n\n

  1. \u6587\u672c\u89c4\u8303\u5316<\/strong>\uff1a\u5c06\u7f29\u5199\u6216\u6570\u5b57\u66f4\u6539\u4e3a\u53d1\u97f3\u3002\u4f8b\u5982\u628a1989\u6539\u6210\u2018\u4e00\u4e5d\u516b\u4e5d\u2019<\/li>
  2. \u5206\u8bcd<\/strong>\uff1a\u8fd9\u5728\u4e2d\u6587\u7b49\u57fa\u4e8e\u5b57\u7b26\u7684\u8bed\u8a00\u4e2d\u662f\u5fc5\u987b\u7684\u90e8\u5206\u3002\u4f8b\u5982\uff0c\u5b83\u6839\u636e\u4e0a\u4e0b\u6587\u5224\u65ad\u662f\u628a\u201c\u5305\u201d\u770b\u6210\u5355\u4e2a\u8bcd\u8fd8\u662f\u628a’\u4e66\u5305’\u548c’\u5305\u5b50’\u5206\u5f00\u770b.<\/li>
  3. \u8bcd\u6027\u6807\u6ce8\uff1a\u628a\u52a8\u8bcd\u3001\u540d\u8bcd\u3001\u4ecb\u8bcd\u7b49\u5206\u6790\u51fa\u6765\u3002<\/li>
  4. Prosody prediction:\u8868\u8fbe\u5bf9\u53e5\u5b50\u7684\u54ea\u4e9b\u90e8\u5206\u91cd\u8bfb\u3001\u6bcf\u4e2a\u90e8\u5206\u7684\u957f\u5ea6\u5982\u4f55\u53d8\u5316\u3001\u8bed\u6c14\u5982\u4f55\u53d8\u5316\u7b49\u7684\u5fae\u5999\u611f\u89c9\u7684\u8bcd\u3002\u5982\u679c\u6ca1\u6709\u8fd9\u4e2a\uff0c\u5b83\u4f1a\u4ea7\u751f\u4e00\u79cd\u771f\u6b63\u611f\u89c9\u50cf\u201c\u673a\u5668\u4eba\u8bf4\u8bdd\u201d\u7684\u58f0\u97f3\u3002\u5c24\u5176\u662f\u82f1\u8bed\uff08stress-based\uff09\u7b49\u8bed\u8a00\u5728\u8fd9\u65b9\u9762\u5dee\u5f02\u5f88\u5927\uff0c\u53ea\u662f\u7a0b\u5ea6\u4e0d\u540c\u800c\u5df2\uff0c\u4f46\u6bcf\u79cd\u8bed\u8a00\u90fd\u6709\u81ea\u5df1\u7684\u97f5\u5f8b\u3002\u5982\u679c\u6211\u4eec\u53ef\u4ee5\u901a\u8fc7\u67e5\u770b\u6587\u672c\u6765\u9884\u6d4b\u8fd9\u4e9b\u97f5\u5f8b\uff0c\u90a3\u80af\u5b9a\u4f1a\u6709\u6240\u5e2e\u52a9\u3002\u4f8b\u5982\uff0c\u6587\u672c\u672b\u5c3e\u7684\u201c?\u201d\u3002\u5982\u679c\u6709\uff0c\u81ea\u7136\u4f1a\u4ea7\u751f\u4e0a\u5347\u7684\u97f3\u8c03\u3002<\/li>
  5. Grapheme-to-phoneme (G2P)<\/strong>\uff1a\u5373\u4f7f\u62fc\u5199\u76f8\u540c\uff0c\u4e5f\u6709\u5f88\u591a\u90e8\u5206\u53d1\u97f3\u4e0d\u540c\u3002\u4f8b\u5982\uff0c\u201cresume\u201d\u8fd9\u4e2a\u8bcd\u6709\u65f6\u4f1a\u8bfb\u4f5c\u201crizju:m\u201d\uff0c\u6709\u65f6\u8bfb\u4f5c\u201crezjumei\u201d\uff0c\u56e0\u6b64\u5fc5\u987b\u67e5\u770b\u6574\u4e2a\u6587\u672c\u7684\u4e0a\u4e0b\u6587<\/strong>\u3002\u6240\u4ee5\uff0c\u5982\u679c\u4f18\u5148\u8003\u8651\u5b57\u7d20\u8f6c\u97f3\u7d20<\/strong>\u7684\u90e8\u5206\uff0c\u4e5f\u5c31\u662f\u5c06\u2018\u8bed\u97f3\u2019\u8f6c\u6362\u6210\u2018spiy ch\u2019\u7b49\u97f3\u6807\u7684\u90e8\u5206\u3002<\/li><\/ol>\n\n\n\n

    \u5728\u8fc7\u53bb\u7684 SPSS \u65f6\u4ee3\uff0c\u6dfb\u52a0\u548c\u5f00\u53d1\u4e86\u8fd9\u4e9b\u4e0d\u540c\u7684\u90e8\u5206\u4ee5\u63d0\u9ad8\u751f\u6210\u97f3\u9891\u7684\u8d28\u91cf\u3002\u5728 neural TTS \u4e2d\uff0c\u8fd9\u4e9b\u90e8\u5206\u5df2\u7ecf\u7b80\u5316\u4e86\u5f88\u591a\uff0c\u4f46\u4ecd\u7136\u6709\u4e00\u4e9b\u90e8\u5206\u662f\u80af\u5b9a\u9700\u8981\u7684\u3002\u4f8b\u59821\u6587\u672c\u89c4\u8303\u5316text normalization \u6216\u80055G2P\u57fa\u672c\u4e0a\u90fd\u662f\u5148\u5904\u7406\u540e\u8f93\u5165\u3002\u5982\u679c\u6709\u7684\u8bba\u6587\u8bf4\u53ef\u4ee5\u63a5\u6536\u5b57\u7b26\u548c\u97f3\u7d20\u4f5c\u4e3a\u8f93\u5165\uff0c\u90a3\u4e48\u5f88\u591a\u60c5\u51b5\u4e0b\u90fd\u4f1a\u5199\u201c\u5b9e\u9645\u4e0a\uff0c\u5f53\u8f93\u5165\u97f3\u7d20\u65f6\u7ed3\u679c\u66f4\u597d\u201d\u3002\u5c3d\u7ba1\u5982\u6b64\uff0c\u5b83\u8fd8\u662f\u6bd4\u4ee5\u524d\u7b80\u5355\u4e86\u5f88\u591a\uff0c\u6240\u4ee5\u5728\u5927\u591a\u6570\u795e\u7ecf TTS \u4e2d\uff0c\u6587\u672c\u5206\u6790\u90e8\u5206\u5e76\u6ca1\u6709\u5355\u72ec\u5904\u7406\uff0c\u5b83\u88ab\u8ba4\u4e3a\u662f\u4e00\u4e2a\u7b80\u5355\u7684\u9884\u5904\u7406\u3002<\/p>\n\n\n\n

    3.\u58f0\u5b66\u6a21\u578b<\/strong><\/h2>\n\n\n\n

    \u58f0\u5b66\u6a21\u578b\u662f\u6307 \u901a\u8fc7\u63a5\u6536\u5b57\u7b26\u6216\u97f3\u7d20\u4f5c\u4e3a\u8f93\u5165\u6216\u901a\u8fc7\u63a5\u6536\u5728\u6587\u672c\u5206\u6790\u90e8\u5206\u521b\u5efa\u7684\u8bed\u8a00\u7279\u5f81\u6765\u751f\u6210\u58f0\u5b66\u7279\u5f81\u7684\u90e8\u5206\u3002\u524d\u9762\u63d0\u5230\uff0c\u5728SPSS\u65f6\u4ee3\uff0cHMM\uff08Hidden Markov Model\uff09\u5728Acoustic Model\u4e2d\u7684\u6bd4\u91cd\u5f88\u5927\uff0c\u540e\u6765\u795e\u7ecf\u7f51\u7edc\u6280\u672f\u9010\u6e10\u53d6\u800c\u4ee3\u4e4b\u3002\u4f8b\u5982\uff0c\u6709\u8bba\u6587\u8868\u660e\u7528 DNN \u66ff\u6362 HMM \u6548\u679c\u66f4\u597d\u3002\u4e0d\u8fc7RNN\u7cfb\u5217\u53ef\u80fd\u66f4\u9002\u5408\u8bed\u97f3\u7b49\u65f6\u95f4\u5e8f\u5217\u3002\u56e0\u6b64\uff0c\u5728\u6709\u4e9b\u8bba\u6587\u4f7f\u7528LSTM\u7b49\u6a21\u578b\u6765\u63d0\u9ad8\u6027\u80fd\u3002\u7136\u800c\uff0c\u5c3d\u7ba1\u4f7f\u7528\u4e86\u795e\u7ecf\u7f51\u7edc\u6a21\u578b\uff0c\u8fd9\u4e9b\u6a21\u578b\u4ecd\u7136\u63a5\u6536\u8bed\u8a00\u7279\u5f81\u4f5c\u4e3a\u8f93\u5165\u548c\u8f93\u51fa\uff0c\u5982 MCC\uff08\u6885\u5c14\u5012\u8c31\u7cfb\u6570\uff09\u3001BAP\uff08\u5e26\u975e\u5468\u671f\u6027\uff09\u3001LSP\uff08\u7ebf\u8c31\u5bf9\uff09\u3001LinS\uff08\u7ebf\u6027\u8c31\u56fe\uff09\u548c F0 .\uff08\u57fa\u9891\uff09\u7b49 \u3002\u56e0\u6b64\uff0c\u8fd9\u4e9b\u6a21\u578b\u53ef\u4ee5\u88ab\u8ba4\u4e3a\u662f\u6539\u8fdb\u7684 SPSS \u6a21\u578b\u3002<\/p>\n\n\n\n

    DeepVoice\u662f\u5434\u6069\u8fbe\u5728\u767e\u5ea6\u7814\u7a76\u9662\u65f6\u5ba3\u5e03\u7684\uff0c\u5176\u5b9e\u66f4\u63a5\u8fd1SPSS\u6a21\u578b\u3002\u5b83\u7531\u51e0\u4e2a\u90e8\u5206\u7ec4\u6210\uff0c\u4f8b\u5982\u4e00\u4e2aG2P\u6a21\u5757\uff0c\u4e00\u4e2a\u5bfb\u627e\u97f3\u7d20\u8fb9\u754c\u7684\u6a21\u5757\uff0c\u4e00\u4e2a\u9884\u6d4b\u97f3\u7d20\u957f\u5ea6\u7684\u6a21\u5757\uff0c\u4e00\u4e2a\u5bfb\u627eF0\u7684\u6a21\u5757\uff0c\u6bcf\u4e2a\u6a21\u5757\u4e2d\u4f7f\u7528\u4e86\u5404\u79cd\u795e\u7ecf\u7f51\u7edc\u6a21\u578b\u3002\u4e4b\u540e\u53d1\u5e03\u7684DeepVoice 2\uff0c\u4e5f\u53ef\u4ee5\u770b\u4f5c\u662f\u7b2c\u4e00\u7248\u7684\u6027\u80fd\u63d0\u5347\u548c\u591a\u626c\u58f0\u5668\u7248\u672c\uff0c\u4f46\u6574\u4f53\u7ed3\u6784\u7c7b\u4f3c\u3002 <\/p>\n\n\n\n

    3.1.\u57fa\u4e8eSeq2seq\u7684\u58f0\u5b66\u6a21\u578b<\/strong><\/h3>\n\n\n\n

    \u57282014-5\u5e74\u7684\u673a\u5668\u7ffb\u8bd1\u9886\u57df\uff0c\u4f7f\u7528attention\u7684seq2seq\u6a21\u578b\u6210\u4e3a\u4e00\u79cd\u8d8b\u52bf\u3002\u7136\u800c\uff0c\u7531\u4e8e\u5b57\u6bcd\u548c\u58f0\u97f3\u4e4b\u95f4\u6709\u5f88\u591a\u76f8\u4f3c\u4e4b\u5904\uff0c\u6240\u4ee5\u53ef\u4ee5\u5e94\u7528\u4e8e\u8bed\u97f3\u3002\u57fa\u4e8e\u8fd9\u4e2a\u60f3\u6cd5\uff0cGoogle \u5f00\u53d1\u4e86 Tacotron[Wang17]\uff08\u56e0\u4e3a\u4f5c\u8005\u559c\u6b22 tacos \u800c\u5f97\u540d\uff09\u3002\u901a\u8fc7\u5c06 CBHG \u6a21\u5757\u6dfb\u52a0\u5230\u4f5c\u4e3a seq2seq \u57fa\u7840\u7684 RNN \u4e2d\uff0c\u7ec8\u4e8e\u5f00\u59cb\u51fa\u73b0\u53ef\u4ee5\u63a5\u6536\u5b57\u7b26\u4f5c\u4e3a\u8f93\u5165\u5e76\u7acb\u5373\u63d0\u53d6\u58f0\u5b66\u7279\u5f81\u7684\u9002\u5f53\u795e\u7ecf TTS\uff0c\u4ece\u800c\u6446\u8131\u4e86\u4ee5\u524d\u7684 SPSS\u3002\u8fd9\u4e2aseq2seq\u6a21\u578b\u4ece\u90a3\u4ee5\u540e\u5f88\u957f\u4e00\u6bb5\u65f6\u95f4\u90fd\u662fTTS\u6a21\u578b\u7684\u57fa\u7840\u3002<\/p>\n\n\n\n

    \u5728\u767e\u5ea6\uff0cDeepVoice 3\u629b\u5f03\u4e86\u4e4b\u524d\u7684\u65e7\u6a21\u578b\uff0c\u52a0\u5165\u4e86\u4f7f\u7528\u6ce8\u610f\u529b\u7684 seq2seq \u3002\u7136\u800c\uff0cDeepVoice \u6301\u7eed\u57fa\u4e8e CNN \u7684\u4f20\u7edf\u4ecd\u7136\u5b58\u5728\u3002DeepVoice \u5728\u7248\u672c 3 \u672b\u5c3e\u505c\u6b62\u4f7f\u7528\u8fd9\u4e2a\u540d\u79f0\uff0c\u4e4b\u540e\u7684 ClariNet\u548c ParaNet\u4e5f\u6cbf\u7528\u4e86\u8be5\u540d\u79f0\u3002\u7279\u522b\u662f\uff0cParaNet \u5f15\u5165\u4e86\u51e0\u79cd\u6280\u672f\u6765\u63d0\u9ad8 seq2seq \u6a21\u578b\u7684\u901f\u5ea6\u3002\u00a0<\/p>\n\n\n\n

    <\/p>\n\n\n\n

    \u00a0\u8c37\u6b4c\u7684 Tacotron \u5728\u4fdd\u6301\u79f0\u4e3a seq2seq \u7684\u57fa\u672c\u5f62\u5f0f\u7684\u540c\u65f6\uff0c\u4e5f\u5411\u5404\u4e2a\u65b9\u5411\u53d1\u5c55\u3002\u7b2c\u4e00\u4e2a\u7248\u672c\u6709\u70b9\u8fc7\u65f6\uff0c\u4f46\u4ece Tacotron 2\u5f00\u59cb\uff0cmel-spectrogram \u88ab\u7528\u4f5c\u9ed8\u8ba4\u7684\u4e2d\u95f4\u8868\u578b\u3002\u5728\u540e\u7eed\u8bba\u6587\u4e2d\uff0c\u5b66\u4e60\u4e86\u5b9a\u4e49\u67d0\u79cd\u8bed\u97f3\u98ce\u683c\u7684\u98ce\u683c\u6807\u8bb0\uff0c\u5e76\u5c06\u5176\u6dfb\u52a0\u5230 Tacotron \u4e2d\uff0c\u4ee5\u521b\u5efa\u4e00\u4e2a\u63a7\u5236\u98ce\u683c\u7684 TTS \u7cfb\u7edf\u3002\u540c\u65f6\u53d1\u8868\u7684\u53e6\u4e00\u7bc7\u8c37\u6b4c\u8bba\u6587 [Skerry-Ryan18] \u4e5f\u63d0\u51fa\u4e86\u4e00\u79cd\u6a21\u578b\uff0c\u53ef\u4ee5\u901a\u8fc7\u6dfb\u52a0\u4e00\u4e2a\u90e8\u5206\u6765\u5b66\u4e60\u97f5\u5f8b\u5d4c\u5165\u5230 Tacotron \u4e2d\u6765\u6539\u53d8\u751f\u6210\u97f3\u9891\u7684\u97f5\u5f8b\u3002\u5728 DCTTS [Tachibana18] \u4e2d\uff0c\u5c06 Tacotron \u7684 RNN \u90e8\u5206\u66ff\u6362\u4e3a Deep CNN \u8868\u660e\u5728\u901f\u5ea6\u65b9\u9762\u6709\u5f88\u5927\u7684\u589e\u76ca\u3002\u4ece\u90a3\u65f6\u8d77\uff0c\u8be5\u6a21\u578b\u5df2\u6539\u8fdb\u4e3a\u5feb\u901f\u6a21\u578b Fast DCTTS\uff0c\u5c3a\u5bf8\u6709\u6548\u51cf\u5c0f\u3002<\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n

    \u5728 DurIAN\u4e2d\uff0cTacotron 2 \u7684\u6ce8\u610f\u529b\u90e8\u5206\u66f4\u6539\u4e3a\u5bf9\u9f50\u6a21\u578b\uff0c\u4ece\u800c\u51cf\u5c11\u4e86\u9519\u8bef\u3002Non-Attentive Tacotron \u4e5f\u505a\u4e86\u7c7b\u4f3c\u7684\u4e8b\u60c5\uff0c\u4f46\u5728\u8fd9\u91cc\uff0cTacotron 2 \u7684\u6ce8\u610f\u529b\u90e8\u5206\u88ab\u66f4\u6539\u4e3a\u6301\u7eed\u65f6\u95f4\u9884\u6d4b\u5668\uff0c\u4ee5\u521b\u5efa\u66f4\u7a33\u5065\u7684\u6a21\u578b\u3002\u5728FCL-TACO2\u4e2d\uff0c\u63d0\u51fa\u4e86\u4e00\u79cd\u534a\u81ea\u56de\u5f52\uff08SAR\uff09\u65b9\u6cd5\uff0c\u6bcf\u4e2a\u97f3\u7d20\u7528AR\u65b9\u6cd5\u5236\u4f5c\uff0c\u6574\u4f53\u7528NAR\u65b9\u6cd5\u5236\u4f5c\uff0c\u4ee5\u63d0\u9ad8\u901f\u5ea6\uff0c\u540c\u65f6\u4fdd\u6301\u8d28\u91cf\u3002\u6b64\u5916\uff0c\u84b8\u998f\u7528\u4e8e\u51cf\u5c0f\u6a21\u578b\u7684\u5927\u5c0f\u3002\u5efa\u8bae\u4f7f\u7528\u57fa\u4e8e Tacotron 2 \u7684\u6a21\u578b\uff0c\u4f46\u901f\u5ea6\u8981\u5feb 17-18 \u500d\u3002\u00a0<\/p>\n\n\n\n

    \"\"<\/figure><\/div>\n\n\n\n

    3.2.\u57fa\u4e8e\u53d8\u538b\u5668\u7684\u58f0\u5b66\u6a21\u578b<\/strong><\/h3>\n\n\n\n

    \u968f\u77402017\u5e74Transformers\u7684\u51fa\u73b0\uff0c\u6ce8\u610f\u529b\u6a21\u578b\u6f14\u53d8\u6210NLP\u9886\u57df\u7684Transformers\uff0c\u4f7f\u7528Transformers\u7684\u6a21\u578b\u4e5f\u5f00\u59cb\u51fa\u73b0\u5728TTS\u9886\u57df\u3002TransformerTTS\u53ef\u4ee5\u770b\u4f5c\u662f\u4e00\u4e2a\u8d77\u70b9\uff0c\u8fd9\u4e2a\u6a21\u578b\u539f\u6837\u6cbf\u7528\u4e86Tacotron 2\u7684\u5927\u90e8\u5206\uff0c\u53ea\u662f\u5c06RNN\u90e8\u5206\u6539\u6210\u4e86Transformer\u3002\u8fd9\u5141\u8bb8\u5e76\u884c\u5904\u7406\u5e76\u5141\u8bb8\u8003\u8651\u66f4\u957f\u7684\u4f9d\u8d56\u6027\u3002<\/p>\n\n\n\n

    FastSpeech\u7cfb\u5217\u53ef\u4ee5\u88ab\u5f15\u7528\u4e3a\u4f7f\u7528 Transformer \u6a21\u578b\u7684 TTS \u7684\u4ee3\u8868\u3002\u5728\u8fd9\u79cd\u60c5\u51b5\u4e0b\uff0c\u53ef\u4ee5\u901a\u8fc7\u4f7f\u7528\u524d\u9988 Transformer \u4ee5\u975e\u5e38\u9ad8\u7684\u901f\u5ea6\u521b\u5efa\u6885\u5c14\u9891\u8c31\u56fe\u3002\u4f5c\u4e3a\u53c2\u8003\uff0cmel-spectrogram\u662f\u4e00\u79cd\u8003\u8651\u4eba\u7684\u542c\u89c9\u7279\u6027\uff0c\u5bf9FFT\u7684\u7ed3\u679c\u8fdb\u884c\u53d8\u6362\u7684\u65b9\u6cd5\uff0c\u867d\u7136\u662f\u6bd4\u8f83\u65e7\u7684\u65b9\u6cd5\uff0c\u4f46\u4ecd\u7136\u88ab\u4f7f\u7528\u3002\u4f18\u70b9\u4e4b\u4e00\u662f\u53ef\u4ee5\u7528\u5c11\u91cf\u7ef4\u5ea6\uff08\u901a\u5e38\u4e3a 80\uff09\u8868\u793a\u3002 <\/p>\n\n\n\n

    \u5728 TTS \u4e2d\uff0c\u5c06\u8f93\u5165\u6587\u672c\u4e0e\u6885\u5c14\u9891\u8c31\u56fe\u7684\u5e27\u76f8\u5339\u914d\u975e\u5e38\u91cd\u8981\u3002\u9700\u8981\u51c6\u786e\u8ba1\u7b97\u51fa\u4e00\u4e2a\u5b57\u7b26\u6216\u97f3\u7d20\u53d8\u5316\u4e86\u591a\u5c11\u5e27\uff0c\u5176\u5b9eattention\u65b9\u6cd5\u8fc7\u4e8e\u7075\u6d3b\uff0c\u5bf9NLP\u53ef\u80fd\u6709\u597d\u5904\uff0c\u4f46\u5728speech\u4e0a\u53cd\u800c\u4e0d\u5229\uff08\u5355\u8bcd\u91cd\u590d\u6216\u8df3\u8fc7\uff09\u3002\u56e0\u6b64\uff0cFastSpeech \u6392\u9664\u4e86\u6ce8\u610f\u529b\u65b9\u6cd5\uff0c\u5e76\u5229\u7528\u4e86\u4e00\u4e2a\u51c6\u786e\u9884\u6d4b\u957f\u5ea6\u7684\u6a21\u5757\uff08\u957f\u5ea6\u8c03\u8282\u5668\uff09\u3002\u540e\u6765\uff0cFastSpeech 2\u8fdb\u4e00\u6b65\u7b80\u5316\u4e86\u7f51\u7edc\u7ed3\u6784\uff0c\u5e76\u989d\u5916\u4f7f\u7528\u4e86\u97f3\u9ad8\u3001\u957f\u5ea6\u548c\u80fd\u91cf\u7b49\u66f4\u591a\u6837\u5316\u7684\u4fe1\u606f\u4f5c\u4e3a\u8f93\u5165\u3002FastPitch\u63d0\u51fa\u4e86\u4e00\u4e2a\u6a21\u578b\uff0c\u901a\u8fc7\u5411 FastSpeech \u6dfb\u52a0\u8be6\u7ec6\u7684\u97f3\u9ad8\u4fe1\u606f\u8fdb\u4e00\u6b65\u6539\u8fdb\u4e86\u7ed3\u679c\u3002LightSpeech\u63d0\u51fa\u4e86\u4e00\u79cd\u7ed3\u6784\uff0c\u901a\u8fc7\u4f7f\u7528 NAS\uff08Neural Architecture Search\uff09\u4f18\u5316\u539f\u672c\u901f\u5ea6\u5f88\u5feb\u7684 FastSpeech \u7684\u7ed3\u6784\uff0c\u5c06\u901f\u5ea6\u63d0\u9ad8\u4e86 6.5 \u500d\u3002<\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n

    MultiSpeech \u8fd8\u4ecb\u7ecd\u4e86\u5404\u79cd\u6280\u672f\u6765\u89e3\u51b3 Transformer \u7684\u7f3a\u70b9\u3002\u5728\u6b64\u57fa\u7840\u4e0a\uff0c\u5bf9 FastSpeech \u8fdb\u884c\u8bad\u7ec3\u4ee5\u521b\u5efa\u4e00\u4e2a\u66f4\u52a0\u6539\u8fdb\u7684 FastSpeech \u6a21\u578b\u3002TransformerTTS \u4f5c\u8005\u968f\u540e\u8fd8\u63d0\u51fa\u4e86\u8fdb\u4e00\u6b65\u6539\u8fdb\u7684 Transformer TTS \u6a21\u578b\uff0c\u5728 RobuTrans\u6a21\u578b\u4e2d\u4f7f\u7528\u57fa\u4e8e\u957f\u5ea6\u7684\u786c\u6ce8\u610f\u529b\u3002AlignTTS \u8fd8\u4ecb\u7ecd\u4e86\u4e00\u79cd\u4f7f\u7528\u5355\u72ec\u7684\u7f51\u7edc\u800c\u4e0d\u662f\u6ce8\u610f\u529b\u6765\u8ba1\u7b97\u5bf9\u9f50\u65b9\u5f0f\u7684\u65b9\u6cd5\u3002\u6765\u81ea Kakao \u7684 JDI-T\u5f15\u5165\u4e86\u4e00\u79cd\u66f4\u7b80\u5355\u7684\u57fa\u4e8e transformer \u7684\u67b6\u6784\uff0c\u8fd8\u4f7f\u7528\u4e86\u6539\u8fdb\u7684\u6ce8\u610f\u529b\u673a\u5236\u3002NCSOFT \u63d0\u51fa\u4e86\u4e00\u79cd\u5728\u6587\u672c\u7f16\u7801\u5668\u548c\u97f3\u9891\u7f16\u7801\u5668\u4e2d\u5206\u5c42\u4f7f\u7528\u8f6c\u6362\u5668\u7684\u65b9\u6cd5\uff0c\u65b9\u6cd5\u662f\u5c06\u5b83\u4eec\u5806\u53e0\u5728\u591a\u4e2a\u5c42\u4e2d\u3002\u9650\u5236\u6ce8\u610f\u529b\u8303\u56f4\u548c\u4f7f\u7528\u591a\u5c42\u6b21\u97f3\u9ad8\u5d4c\u5165\u4e5f\u6709\u52a9\u4e8e\u63d0\u9ad8\u6027\u80fd\u3002<\/p>\n\n\n\n

    3.3.\u57fa\u4e8e\u6d41\u7684\u58f0\u5b66\u6a21\u578b<\/strong><\/h3>\n\n\n\n

    2014\u5e74\u5de6\u53f3\u5f00\u59cb\u5e94\u7528\u4e8e\u56fe\u50cf\u9886\u57df\u7684\u65b0\u4e00\u4ee3\u65b9\u6cd5Flow\uff0c\u4e5f\u88ab\u5e94\u7528\u5230\u58f0\u5b66\u6a21\u578b\u4e2d\u3002Flowtron\u53ef\u4ee5\u770b\u4f5c\u662f Tacotron \u7684\u6539\u8fdb\u6a21\u578b\uff0c\u5b83\u662f\u4e00\u4e2a\u901a\u8fc7\u5e94\u7528 IAF\uff08\u9006\u81ea\u56de\u5f52\u6d41\uff09\u751f\u6210\u6885\u5c14\u8c31\u56fe\u7684\u6a21\u578b\u3002\u5728 Flow-TTS\u4e2d\uff0c\u4f7f\u7528\u975e\u81ea\u56de\u5f52\u6d41\u5236\u4f5c\u4e86\u4e00\u4e2a\u66f4\u5feb\u7684\u6a21\u578b\u3002\u5728\u540e\u7eed\u6a21\u578b EfficientTTS\u4e2d\uff0c\u5728\u6a21\u578b\u8fdb\u4e00\u6b65\u6cdb\u5316\u7684\u540c\u65f6\uff0c\u5bf9\u5bf9\u9f50\u90e8\u5206\u8fdb\u884c\u4e86\u8fdb\u4e00\u6b65\u6539\u8fdb\u3002<\/p>\n\n\n\n

    \u6765\u81ea Kakao \u7684 Glow-TTS \u4e5f\u4f7f\u7528\u6d41\u6765\u521b\u5efa\u6885\u5c14\u9891\u8c31\u56fe\u3002Glow-TTS \u4f7f\u7528\u7ecf\u5178\u7684\u52a8\u6001\u89c4\u5212\u6765\u5bfb\u627e\u6587\u672c\u548c\u6885\u5c14\u5e27\u4e4b\u95f4\u7684\u5339\u914d\uff0c\u4f46 TTS \u8868\u660e\u8fd9\u79cd\u65b9\u6cd5\u4e5f\u53ef\u4ee5\u4ea7\u751f\u9ad8\u6548\u51c6\u786e\u7684\u5339\u914d\u3002\u540e\u6765\uff0c\u8fd9\u79cd\u65b9\u6cd5Monotonic Alignment Search<\/strong>\u88ab\u5e7f\u6cdb\u4f7f\u7528\u3002 <\/p>\n\n\n\n

    3.4.\u57fa\u4e8eVAE\u7684\u58f0\u5b66\u6a21\u578b<\/strong><\/h3>\n\n\n\n

    \u53e6\u4e00\u4e2a\u8bde\u751f\u4e8e 2013 \u5e74\u7684\u751f\u6210\u6a21\u578b\u6846\u67b6 Variational autoencoder (VAE) \u4e5f\u88ab\u7528\u5728\u4e86 TTS \u4e2d\u3002\u987e\u540d\u601d\u4e49\uff0c\u8c37\u6b4c\u5ba3\u5e03\u7684 GMVAE-Tacotron\u4f7f\u7528 VAE \u5bf9\u8bed\u97f3\u4e2d\u7684\u5404\u79cd\u6f5c\u5728\u5c5e\u6027\u8fdb\u884c\u5efa\u6a21\u548c\u63a7\u5236\u3002\u540c\u65f6\u95ee\u4e16\u7684VAE-TTS\u4e5f\u53ef\u4ee5\u901a\u8fc7\u5728Tacotron 2\u6a21\u578b\u4e2d\u6dfb\u52a0\u7528VAE\u5efa\u6a21\u7684\u6837\u5f0f\u90e8\u4ef6\u6765\u505a\u7c7b\u4f3c\u7684\u4e8b\u60c5\u3002BVAE-TTS\u4ecb\u7ecd\u4e86\u4e00\u79cd\u4f7f\u7528\u53cc\u5411 VAE \u5feb\u901f\u751f\u6210\u5177\u6709\u5c11\u91cf\u53c2\u6570\u7684 mel \u7684\u6a21\u578b\u3002Parallel Tacotron\u662f Tacotron \u7cfb\u5217\u7684\u6269\u5c55\uff0c\u8fd8\u5f15\u5165\u4e86 VAE \u4ee5\u52a0\u5feb\u8bad\u7ec3\u548c\u521b\u5efa\u901f\u5ea6\u3002 <\/p>\n\n\n\n

    3.5.\u57fa\u4e8eGAN\u7684\u58f0\u5b66\u6a21\u578b<\/strong><\/h3>\n\n\n\n

    \u5728 2014 \u5e74\u63d0\u51fa\u7684 Generative Adversarial Nets (GAN) \u4e2d\uff0cTacotron 2 \u88ab\u7528\u4f5c\u751f\u6210\u5668\uff0cGAN \u88ab\u7528\u4f5c\u751f\u6210\u66f4\u597d\u7684 mels \u7684\u65b9\u6cd5\u3002\u5728\u8bba\u6587\u4e2d\uff0c\u4f7f\u7528 Adversarial training \u65b9\u6cd5\u8ba9 Tacotron Generator \u4e00\u8d77\u5b66\u4e60\u8bed\u97f3\u98ce\u683c\u3002Multi-SpectroGAN\u8fd8\u4ee5\u5bf9\u6297\u65b9\u5f0f\u5b66\u4e60\u4e86\u51e0\u79cd\u6837\u5f0f\u7684\u6f5c\u5728\u8868\u793a\uff0c\u8fd9\u91cc\u4f7f\u7528 FastSpeech2 \u4f5c\u4e3a\u751f\u6210\u5668\u3002GANSpeech\u8fd8\u4f7f\u7528\u5e26\u6709\u751f\u6210\u5668\u7684 GAN \u65b9\u6cd5\u8bad\u7ec3 FastSpeech1\/2\uff0c\u81ea\u9002\u5e94\u8c03\u6574\u7279\u5f81\u5339\u914d\u635f\u5931\u7684\u89c4\u6a21\u6709\u52a9\u4e8e\u63d0\u9ad8\u6027\u80fd\u3002<\/p>\n\n\n\n

    3.6.\u57fa\u4e8e\u6269\u6563\u7684\u58f0\u5b66\u6a21\u578b<\/strong><\/h3>\n\n\n\n

    \u6700\u8fd1\u5907\u53d7\u5173\u6ce8\u7684\u4f7f\u7528\u6269\u6563\u6a21\u578b\u7684TTS\u4e5f\u76f8\u7ee7\u88ab\u63d0\u51fa\u3002Diff-TTS \u901a\u8fc7\u5bf9\u6885\u5c14\u751f\u6210\u90e8\u5206\u4f7f\u7528\u6269\u6563\u6a21\u578b\u8fdb\u4e00\u6b65\u63d0\u9ad8\u4e86\u7ed3\u679c\u7684\u8d28\u91cf\u3002Grad-TTS \u4e5f\u901a\u8fc7\u5c06\u89e3\u7801\u5668\u66f4\u6539\u4e3a\u6269\u6563\u6a21\u578b\u6765\u505a\u7c7b\u4f3c\u7684\u4e8b\u60c5\uff0c\u4f46\u5728\u8fd9\u91cc\uff0cGlow-TTS \u7528\u4e8e\u9664\u89e3\u7801\u5668\u4e4b\u5916\u7684\u5176\u4f59\u7ed3\u6784\u3002\u5728 PriorGrad \u4e2d\uff0c\u4f7f\u7528\u6570\u636e\u7edf\u8ba1\u521b\u5efa\u5148\u9a8c\u5206\u5e03<\/strong>\uff0c\u4ece\u800c\u5b9e\u73b0\u66f4\u9ad8\u6548\u7684\u5efa\u6a21\u3002\u4e5f\u6709TTS\u7cfb\u7edf\u4f7f\u7528\u6bcf\u4e2a\u97f3\u7d20\u7684\u7edf\u8ba1\u4fe1\u606f\u5e94\u7528\u58f0\u5b66\u6a21\u578b\uff0c\u4f8b\u5982\u817e\u8baf\u7684 DiffGAN-TTS\u4e5f\u4f7f\u7528\u6269\u6563\u89e3\u7801\u5668\uff0c\u5b83\u4f7f\u7528\u5bf9\u6297\u8bad\u7ec3\u65b9\u6cd5\u3002\u8fd9\u5927\u5927\u51cf\u5c11\u4e86\u63a8\u7406\u8fc7\u7a0b\u4e2d\u7684\u6b65\u9aa4\u6570\u5e76\u964d\u4f4e\u4e86\u751f\u6210\u901f\u5ea6\u3002 <\/p>\n\n\n\n

    3.7.\u5176\u4ed6\u58f0\u5b66\u6a21\u578b<\/strong><\/h3>\n\n\n\n

    \u5176\u5b9e\u4e0a\u9762\u4ecb\u7ecd\u7684\u8fd9\u4e9b\u6280\u672f\u4e0d\u4e00\u5b9a\u8981\u5355\u72ec\u4f7f\u7528\uff0c\u800c\u662f\u53ef\u4ee5\u76f8\u4e92\u7ed3\u5408\u4f7f\u7528\u7684\u3002FastSpeech \u7684\u4f5c\u8005\u81ea\u5df1\u5206\u6790\u53d1\u73b0\uff0cVAE \u5373\u4f7f\u5728\u5c0f\u5c3a\u5bf8\u4e0b\u4e5f\u80fd\u5f88\u597d\u5730\u6355\u6349\u97f5\u5f8b\u7b49\u957f\u4fe1\u606f\uff0c\u4f46\u8d28\u91cf\u7565\u5dee\uff0c\u800c Flow \u4fdd\u7559\u7ec6\u8282\u5f88\u597d\uff0c\u800c\u6a21\u578b\u9700\u8981\u5f88\u5927\u4e3a\u4e86\u63d0\u9ad8\u8d28\u91cf\uff0c PortaSpeech\u63d0\u51fa\u4e86\u4e00\u79cd\u6a21\u578b\uff0c\u5305\u542bTransformer+VAE+Flow\u7684\u6bcf\u4e00\u4e2a\u5143\u7d20\u3002<\/p>\n\n\n\n

    VoiceLoop\u63d0\u51fa\u4e86\u4e00\u79cd\u6a21\u578b\uff0c\u8be5\u6a21\u578b\u4f7f\u7528\u7c7b\u4f3c\u4e8e\u4eba\u7c7b\u5de5\u4f5c\u8bb0\u5fc6\u6a21\u578b\u7684\u6a21\u578b\u6765\u5b58\u50a8\u548c\u5904\u7406\u8bed\u97f3\u4fe1\u606f\uff0c\u79f0\u4e3a\u8bed\u97f3\u5faa\u73af\u3002\u5b83\u662f\u8003\u8651\u591a\u626c\u58f0\u5668\u7684\u65e9\u671f\u6a21\u578b\uff0c\u4e4b\u540e\uff0c\u5b83\u88ab\u7528\u4f5cFacebook\u5176\u4ed6\u7814\u7a76\u7684\u9aa8\u5e72\u7f51\u7edc\u3002<\/p>\n\n\n\n

    DeviceTTS\u662f\u4e00\u4e2a\u4f7f\u7528\u6df1\u5ea6\u524d\u9988\u987a\u5e8f\u8bb0\u5fc6\u7f51\u7edc\uff08DFSMN\uff09\u4f5c\u4e3a\u57fa\u672c\u5355\u5143\u7684\u6a21\u578b\u3002\u8be5\u7f51\u7edc\u662f\u4e00\u79cd\u5e26\u6709\u8bb0\u5fc6\u5757\u7684\u524d\u9988\u7f51\u7edc\uff0c\u662f\u4e00\u79cd\u5c0f\u578b\u4f46\u9ad8\u6548\u7684\u7f51\u7edc\uff0c\u53ef\u4ee5\u5728\u4e0d\u4f7f\u7528\u9012\u5f52\u65b9\u6848\u7684\u60c5\u51b5\u4e0b\u4fdd\u6301\u957f\u671f\u4f9d\u8d56\u5173\u7cfb\u3002\u7531\u6b64\uff0c\u63d0\u51fa\u4e86\u4e00\u79cd\u53ef\u4ee5\u5728\u4e00\u822c\u79fb\u52a8\u8bbe\u5907\u4e2d\u5145\u5206\u4f7f\u7528\u7684 TTS \u6a21\u578b\u3002\u00a0<\/p>\n\n\n\n

    4.\u58f0\u7801\u5668<\/strong><\/h2>\n\n\n\n

    \u58f0\u7801\u5668\u662f\u4f7f\u7528\u58f0\u5b66\u6a21\u578b\u751f\u6210\u7684\u58f0\u5b66\u7279\u5f81\u5e76\u5c06\u5176\u8f6c\u6362\u4e3a\u6ce2\u5f62\u7684\u90e8\u4ef6\u3002\u5373\u4f7f\u5728 SPSS \u65f6\u4ee3\uff0c\u5f53\u7136\u4e5f\u9700\u8981\u58f0\u7801\u5668\uff0c\u6b64\u65f6\u4f7f\u7528\u7684\u58f0\u7801\u5668\u5305\u62ec STRAIGHT \u548c WORLD\u3002<\/p>\n\n\n\n

    4.1.\u81ea\u56de\u5f52\u58f0\u7801\u5668<\/strong><\/h3>\n\n\n\n

    Neural Vocoder \u4ece WaveNet\u5f15\u5165\u6269\u5f20\u5377\u79ef\u5c42\u6765\u521b\u5efa\u957f\u97f3\u9891\u6837\u672c\u5f88\u91cd\u8981\uff0c\u5e76\u4e14\u53ef\u4ee5\u4f7f\u7528\u81ea\u56de\u5f52\u65b9\u6cd5\u751f\u6210\u9ad8\u7ea7\u97f3\u9891\uff0c\u8be5\u65b9\u6cd5\u4f7f\u7528\u5148\u524d\u521b\u5efa\u7684\u6837\u672c\u751f\u6210\u4e0b\u4e00\u4e2a\u97f3\u9891\u6837\u672c\uff08\u4e00\u4e2a\u63a5\u4e00\u4e2a\uff09\u3002\u5b9e\u9645\u4e0a\uff0cWaveNet\u672c\u8eab\u53ef\u4ee5\u4f5c\u4e3a\u4e00\u4e2aAcoustic Model+Vocoder\uff0c\u5c06\u8bed\u8a00\u7279\u5f81\u4f5c\u4e3a\u8f93\u5165\uff0c\u751f\u6210\u97f3\u9891\u3002\u7136\u800c\uff0c\u4ece\u90a3\u65f6\u8d77\uff0c\u901a\u8fc7\u66f4\u590d\u6742\u7684\u58f0\u5b66\u6a21\u578b\u521b\u5efa\u6885\u5c14\u9891\u8c31\u56fe\uff0c\u5e76\u57fa\u4e8e WaveNet \u751f\u6210\u97f3\u9891\u5c31\u53d8\u5f97\u5f88\u666e\u904d\u3002<\/p>\n\n\n\n

    \u5728 Tacotron \u4e2d\uff0c\u521b\u5efa\u4e86\u4e00\u4e2a\u7ebf\u6027\u9891\u8c31\u56fe\uff0c\u5e76\u4f7f\u7528 Griffin-Lim \u7b97\u6cd5 \u5c06\u5176\u8f6c\u6362\u4e3a\u6ce2\u5f62\u3002\u7531\u4e8e\u8be5\u7b97\u6cd5\u662f40\u5e74\u524d\u4f7f\u7528\u7684\uff0c\u5c3d\u7ba1\u7f51\u7edc\u7684\u6574\u4f53\u7ed3\u6784\u975e\u5e38\u597d\uff0c\u4f46\u5f97\u5230\u7684\u97f3\u9891\u5e76\u4e0d\u662f\u5f88\u4ee4\u4eba\u6ee1\u610f\u3002\u5728 DeepVoice\u4e2d\uff0c\u4ece\u4e00\u5f00\u59cb\u5c31\u4f7f\u7528\u4e86 WaveNet \u58f0\u7801\u5668\uff0c\u7279\u522b\u662f\u5728\u8bba\u6587 DeepVoice2\u4e2d\uff0c\u9664\u4e86\u4ed6\u4eec\u81ea\u5df1\u7684\u6a21\u578b\u5916\uff0c\u8fd8\u901a\u8fc7\u5c06 WaveNet \u58f0\u7801\u5668\u6dfb\u52a0\u5230\u53e6\u4e00\u5bb6\u516c\u53f8\u7684\u6a21\u578b Tacotron \u6765\u63d0\u9ad8\u6027\u80fd\uff08\u8fd9\u4e48\u8bf4\u6765\uff0c\u5728\u5355\u4e2aspeaker\u4e0a\u6bd4DeepVoice2\u597d\uff09\u7ed9\u51fa\u4e86\u66f4\u597d\u7684\u6027\u80fd\u3002\u81ea\u7248\u672c2\u4ee5\u6765\uff0cTacotron \u4f7f\u7528 WaveNet \u4f5c\u4e3a\u9ed8\u8ba4\u58f0\u7801\u5668\u3002<\/p>\n\n\n\n

    SampleRNN\u662f\u53e6\u4e00\u79cd\u81ea\u56de\u5f52\u6a21\u578b\uff0c\u5728 RNN \u65b9\u6cd5\u4e2d\u4e00\u4e2a\u4e00\u4e2a\u5730\u521b\u5efa\u6837\u672c\u3002\u8fd9\u4e9b\u81ea\u56de\u5f52\u6a21\u578b\u751f\u6210\u97f3\u9891\u7684\u901f\u5ea6\u975e\u5e38\u6162\uff0c\u56e0\u4e3a\u5b83\u4eec\u901a\u8fc7\u4e0a\u4e00\u4e2a\u6837\u672c\u4e00\u4e2a\u4e00\u4e2a\u5730\u6784\u5efa\u4e0b\u4e00\u4e2a\u6837\u672c\u3002\u56e0\u6b64\uff0c\u8bb8\u591a\u540e\u6765\u7684\u7814\u7a76\u5efa\u8bae\u91c7\u7528\u66f4\u5feb\u751f\u4ea7\u7387\u7684\u6a21\u578b\u3002<\/p>\n\n\n\n

    FFTNet\u7740\u773c\u4e8eWaveNet\u7684dilated convolution\u7684\u5f62\u72b6\u4e0eFFT\u7684\u5f62\u72b6\u76f8\u4f3c\uff0c\u63d0\u51fa\u4e86\u4e00\u79cd\u53ef\u4ee5\u52a0\u5feb\u751f\u6210\u901f\u5ea6\u7684\u6280\u672f\u3002\u5728 WaveRNN\u4e2d\uff0c\u4f7f\u7528\u4e86\u5404\u79cd\u6280\u672f\uff08GPU \u5185\u6838\u7f16\u7801\u3001\u526a\u679d\u3001\u7f29\u653e\u7b49\uff09\u6765\u52a0\u901f WaveNet \u3002WaveRNN \u4ece\u6b64\u6f14\u53d8\u6210\u901a\u7528\u795e\u7ecf\u58f0\u7801\u5668\u548c\u5404\u79cd\u5f62\u5f0f\u3002\u5728Towards achieving robust universal neural vocoding(<\/strong>Interspeech 2019)\u4e2d\uff0c\u4f7f\u7528 74 \u4f4d\u8bf4\u8bdd\u4eba\u548c 17 \u79cd\u8bed\u8a00\u7684\u6570\u636e\u5bf9 WaveRNN \u8fdb\u884c\u4e86\u8bad\u7ec3\uff0c\u4ee5\u521b\u5efa RNN_MS\uff08\u591a\u8bf4\u8bdd\u4eba\uff09\u6a21\u578b\uff0c\u8bc1\u660e\u5b83\u662f\u4e00\u79cd\u5373\u4f7f\u5728\u8bf4\u8bdd\u4eba\u548c\u73af\u5883\u4e2d\u4e5f\u80fd\u4ea7\u751f\u826f\u597d\u8d28\u91cf\u7684\u58f0\u7801\u5668\u3002\u6570\u636e\u3002Speaker Conditional WaveRNN: Towards universal neural vocoder for unseen speaker and recording conditions.(<\/strong>Interspeech 2020)\u63d0\u51fa\u4e86Speaker Conditional)_WaveRNN \u6a21\u578b\uff0c\u5373\u901a\u8fc7\u989d\u5916\u4f7f\u7528 speaker embedding \u6765\u5b66\u4e60\u7684\u6a21\u578b\u3002\u8be5\u6a21\u578b\u8fd8\u8868\u660e\u5b83\u9002\u7528\u4e8e\u4e0d\u5728\u6570\u636e\u4e2d\u7684\u8bf4\u8bdd\u4eba\u548c\u73af\u5883\u3002<\/p>\n\n\n\n

    \u82f9\u679c\u7684TTS\u4e5f\u4f7f\u7528\u4e86WaveRNN\u4f5c\u4e3a\u58f0\u7801\u5668\uff0c\u5e76\u4e14\u5728server\u7aef\u548cmobile\u7aef\u505a\u4e86\u5404\u79cd\u4f18\u5316\u7f16\u7801\u548c\u53c2\u6570\u8bbe\u7f6e\uff0c\u4f7f\u5176\u53ef\u4ee5\u5728\u79fb\u52a8\u8bbe\u5907\u4e0a\u4f7f\u7528\u3002<\/p>\n\n\n\n

    \u901a\u8fc7\u5c06\u97f3\u9891\u4fe1\u53f7\u5206\u6210\u51e0\u4e2a\u5b50\u5e26\u6765\u5904\u7406\u97f3\u9891\u4fe1\u53f7\u7684\u65b9\u6cd5\uff0c\u5373\u8f83\u77ed\u7684\u4e0b\u91c7\u6837\u7248\u672c\uff0c\u5df2\u5e94\u7528\u4e8e\u591a\u4e2a\u6a21\u578b\uff0c\u56e0\u4e3a\u5b83\u5177\u6709\u53ef\u4ee5\u5feb\u901f\u5e76\u884c\u8ba1\u7b97\u7684\u4f18\u70b9\uff0c\u5e76\u4e14\u53ef\u4ee5\u5bf9\u6bcf\u4e2a\u5b50\u5e26\u6267\u884c\u4e0d\u540c\u7684\u5904\u7406\u3002<\/p>\n\n\n\n

    \u73b0\u5728\uff0c\u5f88\u591a\u540e\u6765\u63a8\u51fa\u7684\u58f0\u7801\u5668\u90fd\u4f7f\u7528\u975e\u81ea\u56de\u5f52\u65b9\u6cd5\u6765\u6539\u5584\u81ea\u56de\u5f52\u65b9\u6cd5\u751f\u6210\u901f\u5ea6\u6162\u7684\u95ee\u9898\u3002\u6362\u53e5\u8bdd\u8bf4\uff0c\u4e00\u79cd\u65e0\u9700\u67e5\u770b\u5148\u524d\u6837\u672c\uff08\u901a\u5e38\u8868\u793a\u4e3a\u5e73\u884c\uff09\u5373\u53ef\u751f\u6210\u540e\u7eed\u6837\u672c\u7684\u65b9\u6cd5\u3002\u5df2\u7ecf\u63d0\u51fa\u4e86\u5404\u79cd\u5404\u6837\u7684\u975e\u81ea\u56de\u5f52\u65b9\u6cd5\uff0c\u4f46\u6700\u8fd1\u4e00\u7bc7\u8868\u660e\u81ea\u56de\u5f52\u65b9\u6cd5\u4f9d\u65e7\u6297\u6253\u7684\u8bba\u6587\u662f Chunked Autoregressive GAN (CARGAN)\uff0c\u5b83\u8868\u660e\u8bb8\u591a\u975e\u81ea\u56de\u5f52\u58f0\u7801\u5668\u5b58\u5728\u97f3\u9ad8\u9519\u8bef\uff0c\u8fd9\u4e2a\u95ee\u9898\u53ef\u4ee5\u901a\u8fc7\u4f7f\u7528\u81ea\u56de\u5f52\u65b9\u6cd5\u6765\u89e3\u51b3\u3002\u5f53\u7136\uff0c\u901f\u5ea6\u662f\u4e2a\u95ee\u9898\uff0c\u4f46\u662f\u901a\u8fc7\u63d0\u793a\u53ef\u4ee5\u5206\u6210chunked\u5355\u5143\u8ba1\u7b97\uff0c\u7ecd\u4e00\u79cd\u53ef\u4ee5\u663e\u7740\u964d\u4f4e\u901f\u5ea6\u548c\u5185\u5b58\u7684\u65b9\u6cd5\u3002<\/p>\n\n\n\n

    4.2.\u57fa\u4e8e\u6d41\u7684\u58f0\u7801\u5668<\/strong><\/h3>\n\n\n\n

    \u5f52\u4e00\u5316\u57fa\u4e8e\u6d41\u7684\u6280\u672f\u53ef\u4ee5\u5206\u4e3a\u4e24\u5927\u7c7b\u3002\u9996\u5148\u662f\u81ea\u56de\u5f52\u53d8\u6362\uff0c\u5728\u6709\u4ee3\u8868\u6027\u7684IAF\uff08inverse autoregressive flow\uff09\u7684\u60c5\u51b5\u4e0b\uff0c\u751f\u6210\u901f\u5ea6\u975e\u5e38\u5feb\uff0c\u800c\u4e0d\u662f\u9700\u8981\u5f88\u957f\u7684\u8bad\u7ec3\u65f6\u95f4\u3002\u56e0\u6b64\uff0c\u5b83\u53ef\u4ee5\u7528\u6765\u5feb\u901f\u751f\u6210\u97f3\u9891\u3002\u7136\u800c\uff0c\u8bad\u7ec3\u901f\u5ea6\u6162\u662f\u4e00\u4e2a\u95ee\u9898\uff0c\u5728Parallel WaveNet\u4e2d\uff0c\u9996\u5148\u521b\u5efa\u4e00\u4e2a\u81ea\u56de\u5f52WaveNet\u6a21\u578b\uff0c\u7136\u540e\u8bad\u7ec3\u4e00\u4e2a\u7c7b\u4f3c\u7684\u975e\u81ea\u56de\u5f52IAF\u6a21\u578b\u3002\u8fd9\u79f0\u4e3a\u6559\u5e08-\u5b66\u751f\u6a21\u578b\uff0c\u6216\u84b8\u998f\u3002\u4e4b\u540e\uff0cClariNet\u4f7f\u7528\u7c7b\u4f3c\u7684\u65b9\u6cd5\u63d0\u51fa\u4e86\u4e00\u79cd\u66f4\u7b80\u5355\u3001\u66f4\u7a33\u5b9a\u7684\u8bad\u7ec3\u65b9\u6cd5\u3002\u5728\u6210\u529f\u8bad\u7ec3 IAF \u6a21\u578b\u540e\uff0c\u73b0\u5728\u53ef\u4ee5\u5feb\u901f\u751f\u6210\u97f3\u9891\u3002\u4f46\u8bad\u7ec3\u65b9\u6cd5\u590d\u6742\uff0c\u8ba1\u7b97\u91cf\u5927\u3002<\/p>\n\n\n\n

    \u53e6\u4e00\u79cd\u6d41\u6280\u672f\u79f0\u4e3a\u4e8c\u5206\u53d8\u6362\uff0c\u4e00\u79cd\u4f7f\u7528\u79f0\u4e3a\u4eff\u5c04\u8026\u5408\u5c42\u7684\u5c42\u6765\u52a0\u901f\u8bad\u7ec3\u548c\u751f\u6210\u7684\u65b9\u6cd5\u3002\u5927\u7ea6\u5728\u540c\u4e00\u65f6\u95f4\uff0c\u63d0\u51fa\u4e86\u4e24\u4e2a\u4f7f\u7528\u8fd9\u79cd\u65b9\u6cd5\u7684\u58f0\u7801\u5668\uff0cWaveGlow\u548c FloWaveNet\u3002\u8fd9\u4e24\u7bc7\u8bba\u6587\u6765\u81ea\u51e0\u4e4e\u76f8\u4f3c\u7684\u60f3\u6cd5\uff0c\u53ea\u6709\u7ec6\u5fae\u7684\u7ed3\u6784\u5dee\u5f02\uff0c\u5305\u62ec\u6df7\u5408\u901a\u9053\u7684\u65b9\u6cd5\u3002Bipartite transform\u7684\u4f18\u70b9\u662f\u7b80\u5355\uff0c\u4f46\u4e5f\u6709\u7f3a\u70b9\uff0c\u8981\u521b\u5efa\u4e00\u4e2a\u7b49\u4ef7\u4e8eIAF\u7684\u6a21\u578b\uff0c\u9700\u8981\u5806\u53e0\u597d\u51e0\u5c42\uff0c\u6240\u4ee5\u53c2\u6570\u91cf\u6bd4\u8f83\u5927\u3002<\/p>\n\n\n\n

    \u4ece\u90a3\u65f6\u8d77\uff0cWaveFlow\u63d0\u4f9b\u4e86\u51e0\u79cd\u97f3\u9891\u751f\u6210\u65b9\u6cd5\u7684\u7efc\u5408\u89c6\u56fe\u3002\u4e0d\u4ec5\u89e3\u91ca\u4e86 WaveGlow \u548c FloWaveNet \u7b49\u6d41\u65b9\u6cd5\uff0c\u8fd8\u89e3\u91ca\u4e86WaveNet \u4f5c\u4e3a\u5e7f\u4e49\u6a21\u578b\u7684\u751f\u6210\u65b9\u6cd5\uff0c\u6211\u4eec\u63d0\u51fa\u4e86\u4e00\u4e2a\u8ba1\u7b97\u901f\u5ea6\u6bd4\u8fd9\u4e9b\u66f4\u5feb\u7684\u6a21\u578b\u3002\u6b64\u5916\uff0cSqueezeWave\u63d0\u51fa\u4e86\u4e00\u4e2a\u6a21\u578b\uff0c\u8be5\u6a21\u578b\u901a\u8fc7\u6d88\u9664 WaveGlow \u6a21\u578b\u7684\u4f4e\u6548\u7387\u5e76\u4f7f\u7528\u6df1\u5ea6\u53ef\u5206\u79bb\u5377\u79ef\uff0c\u901f\u5ea6\u63d0\u9ad8\u4e86\u51e0\u4e2a\u6570\u91cf\u7ea7\uff08\u6027\u80fd\u7565\u6709\u4e0b\u964d\uff09\u3002WG-WaveNet\u8fd8\u63d0\u51fa\u901a\u8fc7\u5728 WaveGlow \u4e2d\u4f7f\u7528\u6743\u91cd\u5171\u4eab\u663e\u7740\u51cf\u5c0f\u6a21\u578b\u5927\u5c0f\u5e76\u6dfb\u52a0\u4e00\u4e2a\u5c0f\u7684 WaveNet \u6ee4\u6ce2\u5668\u6765\u63d0\u9ad8\u97f3\u9891\u8d28\u91cf\u6765\u521b\u5efa\u6a21\u578b\uff0c\u4ece\u800c\u4f7f 44.1kHz \u97f3\u9891\u5728 CPU \u4e0a\u6bd4\u5b9e\u65f6\u97f3\u9891\u66f4\u5feb\u97f3\u9891…<\/p>\n\n\n\n

    4.3.\u57fa\u4e8e GAN \u7684\u58f0\u7801\u5668<\/strong><\/h3>\n\n\n\n

    \u5e7f\u6cdb\u5e94\u7528\u4e8e\u56fe\u50cf\u9886\u57df\u7684\u751f\u6210\u5bf9\u6297\u7f51\u7edc\uff08GANs\uff09\u7ecf\u8fc7\u5f88\u957f\u4e00\u6bb5\u65f6\u95f4\uff084-5\u5e74\uff09\u540e\u6210\u529f\u5e94\u7528\u4e8e\u97f3\u9891\u751f\u6210\u9886\u57df\u3002WaveGAN\u53ef\u4ee5\u4f5c\u4e3a\u7b2c\u4e00\u4e2a\u4e3b\u8981\u7814\u7a76\u6210\u679c\u88ab\u5f15\u7528\u3002\u5728\u56fe\u50cf\u9886\u57df\u53d1\u5c55\u8d77\u6765\u7684\u7ed3\u6784\u5728\u97f3\u9891\u9886\u57df\u88ab\u6cbf\u7528\uff0c\u6240\u4ee5\u867d\u7136\u521b\u9020\u4e86\u4e00\u5b9a\u8d28\u91cf\u7684\u97f3\u9891\uff0c\u4f46\u4f3c\u4e4e\u4ecd\u7136\u6709\u6240\u6b20\u7f3a\u3002<\/p>\n\n\n\n

    \u4eceGAN-TTS\u5f00\u59cb\uff0c\u4e3a\u4e86\u8ba9\u6a21\u578b\u66f4\u9002\u5408\u97f3\u9891\uff0cvits\u4f5c\u8005\u601d\u8003\u5982\u4f55\u505a\u4e00\u4e2a\u80fd\u591f\u5f88\u597d\u6355\u6349\u6ce2\u5f62\u7279\u5f81\u7684\u5224\u522b\u5668\u3002\u5728 GAN-TTS \u4e2d\uff0c\u4f7f\u7528\u591a\u4e2a\u968f\u673a\u7a97\u53e3\uff08Random window discriminators\uff09\u6765\u8003\u8651\u66f4\u591a\u6837\u5316\u7684\u7279\u5f81\uff0c\u800c\u5728 MelGAN\u4e2d\uff0c\u4f7f\u7528\u4e86\u4e00\u79cd\u5728\u591a\u4e2a\u5c3a\u5ea6\uff08Multi-scale discriminator\uff09\u4e2d\u67e5\u770b\u97f3\u9891\u7684\u65b9\u6cd5\u3002\u6765\u81eaKakao\u7684HiFi-GAN\u63d0\u51fa\u4e86\u4e00\u79cd\u8003\u8651\u66f4\u591a\u97f3\u9891\u7279\u5f81\u7684\u65b9\u6cd5\uff0c\u5373\u4e00\u4e2a\u5468\u671f\uff08Multi-period discriminator\uff09\u3002\u5728 VocGAN\u7684\u60c5\u51b5\u4e0b\uff0c\u8fd8\u4f7f\u7528\u4e86\u5177\u6709\u591a\u79cd\u5206\u8fa8\u7387\u7684\u9274\u522b\u5668\u3002\u5728A spectral energy distance for parallel speech synthesis. <\/strong>NeurIPS 2020.\u4e2d\uff0c\u751f\u6210\u7684\u5206\u5e03\u4e0e\u5b9e\u9645\u5206\u5e03\u4e4b\u95f4\u7684\u5dee\u5f02\u4ee5\u5e7f\u4e49\u80fd\u91cf\u8ddd\u79bb (GED) \u7684\u5f62\u5f0f\u5b9a\u4e49\uff0c\u5e76\u5728\u6700\u5c0f\u5316\u5b83\u7684\u65b9\u5411\u4e0a\u5b66\u4e60\u3002\u590d\u6742\u7684\u9274\u522b\u5668\u4ee5\u5404\u79cd\u65b9\u5f0f\u6781\u5927\u5730\u63d0\u9ad8\u4e86\u751f\u6210\u97f3\u9891\u7684\u6027\u80fd\u3002GAN Vocoder: Multi-resolution discriminator is all you need<\/strong>(Interspeech 2021)\u8fdb\u4e00\u6b65\u5206\u6790\u4e86\u8fd9\u4e00\u70b9\uff0c\u5e76\u63d0\u5230\u4e86\u591a\u5206\u8fa8\u7387\u9274\u522b\u5668\u7684\u91cd\u8981\u6027\u3002\u5728 Fre-GAN\u4e2d\uff0c\u751f\u6210\u5668\u548c\u9274\u522b\u5668\u90fd\u4f7f\u7528\u591a\u5206\u8fa8\u7387\u65b9\u6cd5\u8fde\u63a5\u3002\u4f7f\u7528\u79bb\u6563\u6ce2\u5f62\u53d8\u6362 (DWT) \u4e5f\u6709\u5e2e\u52a9\u3002 <\/p>\n\n\n\n

    \u5728generator\u7684\u60c5\u51b5\u4e0b\uff0c\u5f88\u591a\u6a21\u578b\u4f7f\u7528\u4e86MelGAN\u63d0\u51fa\u7684dilated + transposed convolution\u7ec4\u5408\u3002\u5982\u679c\u7a0d\u6709\u4e0d\u540c\uff0cParallel WaveGAN \u00a0\u4e5f\u63a5\u6536\u9ad8\u65af\u566a\u58f0\u4f5c\u4e3a\u8f93\u5165\uff0c\u800c VocGAN \u751f\u6210\u5404\u79cd\u5c3a\u5ea6\u7684\u6ce2\u5f62\u3002\u5728 HiFi-GAN \u4e2d\uff0c\u4f7f\u7528\u4e86\u5177\u6709\u591a\u4e2a\u611f\u53d7\u91ce\u7684\u751f\u6210\u5668\u3002<\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n

    \u524d\u9762\u63d0\u5230\u7684 Parallel WaveGAN\u662f Naver\/Line \u63d0\u51fa\u7684\u4e00\u79cd\u6a21\u578b\uff0c\u5b83\u53ef\u4ee5\u901a\u8fc7\u63d0\u51fa\u975e\u81ea\u56de\u5f52 WaveNet \u751f\u6210\u5668\u6765\u4ee5\u975e\u5e38\u9ad8\u7684\u901f\u5ea6\u751f\u6210\u97f3\u9891\u3002\u4e4b\u540e\uff0c\u63d0\u51fa\u4e86\u4e00\u79cd\u8fdb\u4e00\u6b65\u6539\u8fdb\u7684 Parallel WaveGAN\uff0c\u901a\u8fc7\u5e94\u7528\u611f\u77e5\u63a9\u853d\u6ee4\u6ce2\u5668\u6765\u51cf\u5c11\u542c\u89c9\u654f\u611f\u9519\u8bef\u3002\u6b64\u5916\uff0c[Wang21] \u63d0\u51fa\u4e86\u4e00\u79cd\u901a\u8fc7\u5c06 Pointwise Improved Parallel WaveGAN vocoder with perceptually weighted spectrogram loss.Relativistic LSGAN\uff08\u4e00\u79cd\u6539\u8fdb\u7684\u6700\u5c0f\u4e8c\u4e58 GAN\uff09\u5e94\u7528\u4e8e\u97f3\u9891\u6765\u521b\u5efa\u5177\u6709\u8f83\u5c11\u5c40\u90e8\u4f2a\u5f71\u7684 Parallel WaveGAN\uff08\u548c MelGAN\uff09\u7684\u65b9\u6cd5\u3002\u5728 LVCNet\u4e2d\uff0c\u4f7f\u7528\u6839\u636e\u6761\u4ef6\u53d8\u5316\u7684\u5377\u79ef\u5c42\u7684\u751f\u6210\u5668\uff0c\u79f0\u4e3a\u4f4d\u7f6e\u53ef\u53d8\u5377\u79ef\uff0c\u88ab\u653e\u5165 Parallel WaveGAN \u5e76\u8bad\u7ec3\u4ee5\u521b\u5efa\u66f4\u5feb\uff084x\uff09\u7684\u751f\u6210\u6a21\u578b\uff0c\u8d28\u91cf\u5dee\u5f02\u5f88\u5c0f\u3002 <\/p>\n\n\n\n

    \u6b64\u540e\uff0cMelGAN \u4e5f\u5f97\u5230\u4e86\u591a\u79cd\u5f62\u5f0f\u7684\u6539\u8fdb\u3002\u5728Multi-Band MelGAN \u4e2d\uff0c\u589e\u52a0\u4e86\u539f\u6709MelGAN\u7684\u611f\u53d7\u91ce\uff0c\u589e\u52a0\u4e86\u591a\u5206\u8fa8\u7387STFT loss\uff08Parallel WaveGAN\u5efa\u8bae\uff09\uff0c\u8ba1\u7b97\u4e86\u591a\u6ce2\u6bb5\u5212\u5206\uff08DurIAN\u5efa\u8bae\uff09\uff0c\u4f7f\u5f97\u901f\u5ea6\u66f4\u5feb\uff0c\u66f4\u7a33\u5b9a\u7684\u6a21\u578b\u3002\u8fd8\u63d0\u51fa\u4e86 Universal MelGAN \u7684\u591a\u626c\u58f0\u5668\u7248\u672c\uff0c\u5b83\u4e5f\u4f7f\u7528\u591a\u5206\u8fa8\u7387\u9274\u522b\u5668\u6765\u751f\u6210\u5177\u6709\u66f4\u591a\u7ec6\u8282\u7684\u97f3\u9891\u3002\u8fd9\u4e2a\u60f3\u6cd5\u5728\u540e\u7eed\u7684\u7814\u7a76 UnivNet\u4e2d\u5f97\u5230\u5ef6\u7eed\uff0c\u5e76\u8fdb\u4e00\u6b65\u6539\u8fdb\uff0c\u6bd4\u5982\u4e00\u8d77\u4f7f\u7528\u591a\u5468\u671f\u5224\u522b\u5668\u3002\u5728\u8fd9\u4e9b\u7814\u7a76\u4e2d\uff0c\u97f3\u9891\u8d28\u91cf\u4e5f\u901a\u8fc7\u4f7f\u7528\u66f4\u5bbd\u7684\u9891\u5e26 (80->100) mel \u5f97\u5230\u6539\u5584\u3002<\/p>\n\n\n\n

    \u9996\u5c14\u56fd\u7acb\u5927\u5b66\/NVIDIA \u63a8\u51fa\u4e86\u4e00\u79cd\u540d\u4e3a BigVGAN\u7684\u65b0\u578b\u58f0\u7801\u5668\u3002\u4f5c\u4e3a\u8003\u8651\u5404\u79cd\u5f55\u97f3\u73af\u5883\u548c\u672a\u89c1\u8bed\u8a00\u7b49\u7684\u901a\u7528Vocoder\uff0c\u4f5c\u4e3a\u6280\u672f\u6539\u8fdb\uff0c\u4f7f\u7528snake\u51fd\u6570\u4e3aHiFi-GAN\u751f\u6210\u5668\u63d0\u4f9b\u5468\u671f\u6027\u7684\u5f52\u7eb3\u504f\u7f6e\uff0c\u5e76\u52a0\u5165\u4f4e\u901a\u6ee4\u6ce2\u5668\u4ee5\u51cf\u5c11\u8fb9\u7531\u6b64\u9020\u6210\u7684\u5f71\u54cd\u3002\u53e6\u5916\uff0c\u6a21\u578b\u7684\u5927\u5c0f\u4e5f\u5927\u5927\u589e\u52a0\u4e86\uff08~112M\uff09\uff0c\u8bad\u7ec3\u4e5f\u6210\u529f\u4e86\u3002<\/p>\n\n\n\n

    4.4.\u57fa\u4e8e\u6269\u6563\u7684\u58f0\u7801\u5668<\/strong><\/h3>\n\n\n\n

    \u6269\u6563\u6a21\u578b\u53ef\u4ee5\u79f0\u4e3a\u6700\u65b0\u4e00\u4ee3\u6a21\u578b\uff0c\u8f83\u65e9\u5730\u5e94\u7528\u4e8e\u58f0\u7801\u5668\u3002ICLR21\u540c\u65f6\u4ecb\u7ecd\u4e86\u601d\u8def\u76f8\u4f3c\u7684DiffWave\u548cWaveGrad\u3002Diffusion Model\u7528\u4e8e\u97f3\u9891\u751f\u6210\u90e8\u5206\u662f\u4e00\u6837\u7684\uff0c\u4f46DiffWave\u7c7b\u4f3c\u4e8eWaveNet\uff0cWaveGrad\u57fa\u4e8eGAN-TTS\u3002\u5904\u7406\u8fed\u4ee3\u7684\u65b9\u5f0f\u4e5f\u6709\u6240\u4e0d\u540c\u3002\u4e4b\u524d\u58f0\u5b66\u6a21\u578b\u90e8\u5206\u4ecb\u7ecd\u7684PriorGrad \u4e5f\u4ee5\u521b\u5efa\u58f0\u7801\u5668\u4e3a\u4f8b\u8fdb\u884c\u4e86\u4ecb\u7ecd\u3002\u5728\u8fd9\u91cc\uff0c\u5148<\/strong>\u9a8c\u662f\u4f7f\u7528\u6885\u5c14\u8c31\u56fe\u7684\u80fd\u91cf\u8ba1\u7b97\u7684<\/strong>\u3002 <\/p>\n\n\n\n

    \u6269\u6563\u6cd5\u7684\u4f18\u70b9\u662f\u53ef\u4ee5\u5b66\u4e60\u590d\u6742\u7684\u6570\u636e\u5206\u5e03\u5e76\u4ea7\u751f\u9ad8\u8d28\u91cf\u7684\u7ed3\u679c\uff0c\u4f46\u6700\u5927\u7684\u7f3a\u70b9\u662f\u751f\u6210\u65f6\u95f4\u76f8\u5bf9\u8f83\u957f\u3002\u53e6\u5916\uff0c\u7531\u4e8e\u8fd9\u79cd\u65b9\u6cd5\u672c\u8eab\u662f\u4ee5\u53bb\u9664\u566a\u58f0\u7684\u65b9\u5f0f\u8fdb\u884c\u7684\uff0c\u56e0\u6b64\u5982\u679c\u8fdb\u884c\u65f6\u95f4\u8fc7\u957f\uff0c\u5b58\u5728\u539f\u59cb\u97f3\u9891\u4e2d\u5b58\u5728\u7684\u8bb8\u591a\u566a\u58f0\uff08\u6e05\u97f3\u7b49\uff09\u4e5f\u4f1a\u6d88\u5931\u7684\u7f3a\u70b9\u3002FastDiff\u901a\u8fc7\u5c06 LVCNet\u7684\u601d\u60f3\u5e94\u7528\u5230\u6269\u6563\u6a21\u578b\u4e2d\uff0c\u63d0\u51fa\u4e86\u65f6\u95f4\u611f\u77e5\u7684\u4f4d\u7f6e-\u53d8\u5316\u5377\u79ef\u3002\u901a\u8fc7\u8fd9\u79cd\u65b9\u5f0f\uff0c\u53ef\u4ee5\u66f4\u7a33\u5065\u5730\u5e94\u7528\u6269\u6563\uff0c\u5e76\u4e14\u53ef\u4ee5\u901a\u8fc7\u4f7f\u7528\u566a\u58f0\u8c03\u5ea6\u9884\u6d4b\u5668\u8fdb\u4e00\u6b65\u51cf\u5c11\u751f\u6210\u65f6\u95f4\u3002 <\/p>\n\n\n\n

    \u6765\u81ea\u817e\u8baf\u7684 BDDM\u4e5f\u63d0\u51fa\u4e86\u4e00\u79cd\u5927\u5927\u51cf\u5c11\u521b\u5efa\u65f6\u95f4\u7684\u65b9\u6cd5\u3002\u6362\u53e5\u8bdd\u8bf4\uff0c\u6269\u6563\u8fc7\u7a0b\u7684\u6b63\u5411\u548c\u53cd\u5411\u8fc7\u7a0b\u4f7f\u7528\u4e0d\u540c\u7684\u7f51\u7edc\uff08\u6b63\u5411\uff1a\u8c03\u5ea6\u7f51\u7edc\uff0c\u53cd\u5411\uff1a\u5206\u6570\u7f51\u7edc\uff09\uff0c\u5e76\u4e3a\u6b64\u63d0\u51fa\u4e86\u4e00\u4e2a\u65b0\u7684\u7406\u8bba\u76ee\u6807\u3002\u5728\u8fd9\u91cc\uff0c\u6211\u4eec\u5c55\u793a\u4e86\u81f3\u5c11\u53ef\u4ee5\u901a\u8fc7\u4e09\u4e2a\u6b65\u9aa4\u751f\u6210\u97f3\u9891\u3002\u5728\u8fd9\u4e2a\u901f\u5ea6\u4e0b\uff0c\u6269\u6563\u6cd5\u4e5f\u53ef\u4ee5\u7528\u4e8e\u5b9e\u9645\u76ee\u7684\u3002\u867d\u7136\u4ee5\u524d\u7684\u5927\u591a\u6570\u7814\u7a76\u4f7f\u7528 DDPM \u578b\u5efa\u6a21\uff0c\u4f46\u6269\u6563\u6a21\u578b\u4e5f\u53ef\u4ee5\u7528\u968f\u673a\u5fae\u5206\u65b9\u7a0b (SDE) \u7684\u5f62\u5f0f\u8868\u793a\u3002ItoWave\u5c55\u793a\u4e86\u4f7f\u7528 SDE \u7c7b\u578b\u5efa\u6a21\u751f\u6210\u97f3\u9891\u7684\u793a\u4f8b\u3002<\/p>\n\n\n\n

    4.5.\u57fa\u4e8e\u6e90\u6ee4\u6ce2\u5668\u7684\u58f0\u7801\u5668<\/strong><\/h3>\n\n\n\n

    \u5728\u8fd9\u7bc7\u6587\u7ae0\u7684\u5f00\u5934\uff0c\u5728\u5904\u7406 TTS \u7684\u5386\u53f2\u65f6\uff0c\u6211\u4eec\u7b80\u5355\u5730\u4e86\u89e3\u4e86 Formant Synthesis\u3002\u4eba\u58f0\u662f\u4e00\u79cd\u5efa\u6a21\u65b9\u6cd5\uff0c\u8ba4\u4e3a\u57fa\u672c\u58f0\u6e90\uff08\u6b63\u5f26\u97f3\u7b49\uff09\u7ecf\u8fc7\u53e3\u90e8\u7ed3\u6784\u8fc7\u6ee4\uff0c\u8f6c\u5316\u4e3a\u6211\u4eec\u542c\u5230\u7684\u58f0\u97f3\u3002\u8fd9\u79cd\u65b9\u6cd5\u6700\u91cd\u8981\u7684\u90e8\u5206\u662f\u5982\u4f55\u5236\u4f5c\u8fc7\u6ee4\u5668\u3002\u5728 DL \u65f6\u4ee3\uff0c\u5982\u679c\u8fd9\u4e2a\u8fc7\u6ee4\u5668\u7528\u795e\u7ecf\u7f51\u7edc\u5efa\u6a21\uff0c\u6027\u80fd\u4f1a\u4e0d\u4f1a\u66f4\u597d\u3002\u5728\u795e\u7ecf\u6e90\u6ee4\u6ce2\u5668\u65b9\u6cd5 [Wang19a] \u4e2d\uff0c\u4f7f\u7528 f0\uff08\u97f3\u9ad8\uff09\u4fe1\u606f\u521b\u5efa\u57fa\u672c\u6b63\u5f26\u58f0\u97f3\uff0c\u5e76\u8bad\u7ec3\u4f7f\u7528\u6269\u5f20\u5377\u79ef\u7684\u6ee4\u6ce2\u5668\u4ee5\u4ea7\u751f\u4f18\u8d28\u58f0\u97f3\u3002\u4e0d\u662f\u81ea\u56de\u5f52\u7684\u65b9\u6cd5\uff0c\u6240\u4ee5\u901f\u5ea6\u5f88\u5feb\u3002\u4e4b\u540e\uff0c\u5728Neural harmonic-plus-noise waveform model with trainable maximum voice frequency for text-to-speech synthesis.\u4e2d\uff0c\u5c06\u5176\u6269\u5c55\u91cd\u6784\u4e3a\u8c10\u6ce2+\u566a\u58f0\u6a21\u578b\u4ee5\u63d0\u9ad8\u6027\u80fd\u3002DDSP \u63d0\u51fa\u4e86\u4e00\u79cd\u4f7f\u7528\u795e\u7ecf\u7f51\u7edc\u548c\u591a\u4e2a DSP \u7ec4\u4ef6\u521b\u5efa\u5404\u79cd\u58f0\u97f3\u7684\u65b9\u6cd5\uff0c\u5176\u4e2d\u8c10\u6ce2\u4f7f\u7528\u52a0\u6cd5\u5408\u6210\u65b9\u6cd5\uff0c\u566a\u58f0\u4f7f\u7528\u7ebf\u6027\u65f6\u53d8\u6ee4\u6ce2\u5668\u3002 <\/p>\n\n\n\n

    \u53e6\u4e00\u79cd\u65b9\u6cd5\u662f\u5c06\u4e0e\u8bed\u97f3\u97f3\u9ad8\u76f8\u5173\u7684\u90e8\u5206\uff08\u5171\u632f\u5cf0\uff09\u548c\u5176\u4ed6\u90e8\u5206\uff08\u79f0\u4e3a\u6b8b\u5dee\u3001\u6fc0\u52b1\u7b49\uff09\u8fdb\u884c\u5212\u5206\u548c\u5904\u7406\u7684\u65b9\u6cd5\u3002\u8fd9\u4e5f\u662f\u4e00\u79cd\u5386\u53f2\u60a0\u4e45\u7684\u65b9\u6cd5\u3002\u5171\u632f\u5cf0\u4e3b\u8981\u4f7f\u7528\u4e86LP\uff08\u7ebf\u6027\u9884\u6d4b\uff09\uff0c\u6fc0\u52b1\u4f7f\u7528\u4e86\u5404\u79cd\u6a21\u578b\u3002GlotNet\u5728\u795e\u7ecf\u7f51\u7edc\u65f6\u4ee3\u63d0\u51fa\uff0c\u5c06\uff08\u58f0\u95e8\uff09\u6fc0\u52b1\u5efa\u6a21\u4e3a WaveNet\u3002\u4e4b\u540e\uff0cGELP \u7528 GAN \u8bad\u7ec3\u65b9\u6cd5\u5c06\u5176\u6269\u5c55\u4e3a\u5e76\u884c\u683c\u5f0f\u3002<\/p>\n\n\n\n

    Naver\/Yonsei University \u7684 ExcitNet\u4e5f\u53ef\u4ee5\u770b\u4f5c\u662f\u5177\u6709\u7c7b\u4f3c\u601d\u60f3\u7684\u6a21\u578b\uff0c\u7136\u540e\uff0c\u5728\u6269\u5c55\u6a21\u578b LP-WaveNet\u4e2d\uff0csource \u548c filter \u4e00\u8d77\u8bad\u7ec3\uff0c\u5e76\u4f7f\u7528\u66f4\u590d\u6742\u7684\u6a21\u578b\u3002\u5728 Neural text-to-speech with a modeling-by-generation excitation vocoder(Interspeech 2020)<\/strong>\u4e2d\uff0c\u5f15\u5165\u4e86\u9010\u4ee3\u5efa\u6a21 (MbG) \u6982\u5ff5\uff0c\u4ece\u58f0\u5b66\u6a21\u578b\u751f\u6210\u7684\u4fe1\u606f\u53ef\u7528\u4e8e\u58f0\u7801\u5668\u4ee5\u63d0\u9ad8\u6027\u80fd\u3002\u5728\u795e\u7ecf\u540c\u6001\u58f0\u7801\u5668\u4e2d\uff0c\u8c10\u6ce2\u4f7f\u7528\u7ebf\u6027\u65f6\u53d8 (LTV) \u8109\u51b2\u5e8f\u5217\uff0c\u566a\u58f0\u4f7f\u7528 LTV \u566a\u58f0\u3002Unified source-filter GAN: Unified source-filter network based on factorization of quasi-periodic Parallel WaveGAN(Interspeech 2021)<\/strong>\u63d0\u51fa\u4e86\u4e00\u79cd\u6a21\u578b\uff0c\u5b83\u4f7f\u7528 Parallel WaveGAN \u4f5c\u4e3a\u58f0\u7801\u5668\uff0c\u5e76\u96c6\u6210\u4e86\u4e0a\u8ff0\u51e0\u79cd\u6e90\u6ee4\u6ce2\u5668\u6a21\u578b\u3002Parallel WaveGAN\u672c\u8eab\u4e5f\u88abNaver\u4e0d\u65ad\u6269\u5145\uff0c\u9996\u5148\u5728High-fidelity Parallel WaveGAN with multi-band harmonic-plus-noise model(Interspeech 2021)<\/strong>\u4e2d\uff0cGenerator\u88ab\u6269\u5145\u4e3aHarmonic + Noise\u6a21\u578b\uff0c\u540c\u65f6\u4e5f\u52a0\u5165\u4e86subband\u7248\u672c\u3002<\/p>\n\n\n\n

    LPCNet\u53ef\u4ee5\u88ab\u8ba4\u4e3a\u662f\u7ee7\u8fd9\u79cd\u6e90\u8fc7\u6ee4\u5668\u65b9\u6cd5\u4e4b\u540e\u4f7f\u7528\u6700\u5e7f\u6cdb\u7684\u6a21\u578b\u3002\u4f5c\u4e3a\u5728 WaveRNN \u4e2d\u52a0\u5165\u7ebf\u6027\u9884\u6d4b\u7684\u6a21\u578b\uff0c  LPCNet \u6b64\u540e\u4e5f\u8fdb\u884c\u4e86\u591a\u65b9\u9762\u7684\u6539\u8fdb\u3002\u5728 Bunched LPCNet \u4e2d\uff0c\u901a\u8fc7\u5229\u7528\u539f\u59cb WaveRNN \u4e2d\u5f15\u5165\u7684\u6280\u672f\uff0cLPCNet \u53d8\u5f97\u66f4\u52a0\u9ad8\u6548\u3002Gaussian LPCNet\u8fd8\u901a\u8fc7\u5141\u8bb8\u540c\u65f6\u9884\u6d4b\u591a\u4e2a\u6837\u672c\u6765\u63d0\u9ad8\u6548\u7387\u3002Lightweight LPCNet-based neural vocoder with tensor decomposition(Interspeech 2020)\u901a\u8fc7\u4f7f\u7528\u5f20\u91cf\u5206\u89e3\u8fdb\u4e00\u6b65\u51cf\u5c0f WaveRNN \u5185\u90e8\u7ec4\u4ef6\u7684\u5927\u5c0f\u6765\u63d0\u9ad8\u53e6\u4e00\u4e2a\u65b9\u5411\u7684\u6548\u7387\u3002iLPCNet\u8be5\u6a21\u578b\u901a\u8fc7\u5229\u7528\u8fde\u7eed\u5f62\u5f0f\u7684\u6df7\u5408\u5bc6\u5ea6\u7f51\u7edc\u663e\u793a\u51fa\u6bd4\u73b0\u6709 LPCNet \u66f4\u9ad8\u7684\u6027\u80fd\u3002Fast and lightweight on-device tts with Tacotron2 and LPCNet(Interspeech 2020)\u63d0\u51fa\u4e86\u4e00\u79cd\u6a21\u578b\uff0c\u5728LPCNet\u4e2d\u7684\u8bed\u97f3\u4e2d\u627e\u5230\u53ef\u4ee5\u5207\u65ad\u7684\u90e8\u5206\uff08\u4f8b\u5982\uff0c\u505c\u987f\u6216\u6e05\u97f3\uff09\uff0c\u5c06\u5b83\u4eec\u5212\u5206\uff0c\u5e76\u884c\u5904\u7406\uff0c\u5e76\u901a\u8fc7\u4ea4\u53c9\u6de1\u5165\u6de1\u51fa\u6765\u52a0\u5feb\u751f\u6210\u901f\u5ea6. LPCNet \u4e5f\u6269\u5c55\u5230\u4e86\u5b50\u5e26\u7248\u672c\uff0c\u9996\u5148\u5728 FeatherWave\u4e2d\u5f15\u5165\u5b50\u5e26 LPCNet\u3002\u5728An efficient subband linear prediction for lpcnet-based neural synthesis(Interspeech 2020)\u4e2d\uff0c\u63d0\u51fa\u4e86\u8003\u8651\u5b50\u5e26\u4e4b\u95f4\u76f8\u5173\u6027\u7684\u5b50\u5e26 LPCNet \u7684\u6539\u8fdb\u7248\u672c.<\/p>\n\n\n\n

    \u58f0\u7801\u5668\u7684\u53d1\u5c55\u6b63\u671d\u7740\u4ece\u9ad8\u8d28\u91cf\u3001\u6162\u901f\u7684AR\uff08Autoregressive\uff09\u65b9\u6cd5\u5411\u5feb\u901f\u7684NAR\uff08Non-autoregressive\uff09\u65b9\u6cd5\u8f6c\u53d8\u7684\u65b9\u5411\u53d1\u5c55\u3002\u7531\u4e8e\u51e0\u79cd\u5148\u8fdb\u7684\u751f\u6210\u6280\u672f\uff0cNAR \u4e5f\u9010\u6e10\u8fbe\u5230 AR \u7684\u6c34\u5e73\u3002\u4f8b\u5982\u5728TTS-BY-TTS [Hwang21a]\u4e2d\uff0c\u4f7f\u7528AR\u65b9\u6cd5\u521b\u5efa\u4e86\u5927\u91cf\u6570\u636e\u5e76\u7528\u4e8eNAR\u6a21\u578b\u7684\u8bad\u7ec3\uff0c\u6548\u679c\u4e0d\u9519\u3002\u4f46\u662f\uff0c\u4f7f\u7528\u6240\u6709\u6570\u636e\u53ef\u80fd\u4f1a\u5f88\u7cdf\u7cd5\u3002\u56e0\u6b64\uff0cTTS-BY-TTS2\u63d0\u51fa\u4e86\u4e00\u79cd\u4ec5\u4f7f\u7528\u6b64\u6570\u636e\u8fdb\u884c\u8bad\u7ec3\u7684\u65b9\u6cd5\uff0c\u65b9\u6cd5\u662f\u4f7f\u7528 RankSVM \u83b7\u5f97\u4e0e\u539f\u59cb\u97f3\u9891\u66f4\u76f8\u4f3c\u7684\u5408\u6210\u97f3\u9891\u3002 <\/p>\n\n\n\n

    DelightfulTTS\uff0c\u5fae\u8f6f\u4f7f\u7528\u7684 TTS \u7cfb\u7edf\uff0c\u6709\u4e00\u4e9b\u81ea\u5df1\u7684\u7ed3\u6784\u4fee\u6539\uff0c\u4f8b\u5982\u4f7f\u7528 conformers\uff0c\u5e76\u4e14\u7279\u522b\u4ee5\u751f\u6210 48 kHz \u7684\u6700\u7ec8\u97f3\u9891\u4e3a\u7279\u5f81\uff08\u5927\u591a\u6570 TTS \u7cfb\u7edf\u901a\u5e38\u751f\u6210 16 kHz \u97f3\u9891\uff09\u3002\u4e3a\u6b64\uff0c\u6885\u5c14\u9891\u8c31\u56fe\u4ee5 16kHz \u7684\u9891\u7387\u751f\u6210\uff0c\u4f46\u6700\u7ec8\u97f3\u9891\u662f\u4f7f\u7528\u5185\u90e8\u5236\u4f5c\u7684 HiFiNet \u4ee5 48kHz \u7684\u9891\u7387\u751f\u6210\u7684\u3002<\/p>\n\n\n\n

    5.\u5b8c\u5168\u7aef\u5230\u7aef\u7684TTS<\/strong><\/h2>\n\n\n\n

    \u901a\u8fc7\u4e00\u8d77\u5b66\u4e60\u58f0\u5b66\u6a21\u578b\u548c\u58f0\u7801\u5668\uff0c\u4ecb\u7ecd\u5728\u8f93\u5165\u6587\u672c\u6216\u97f3\u7d20\u65f6\u7acb\u5373\u521b\u5efa\u6ce2\u5f62\u97f3\u9891\u7684\u6a21\u578b\u3002\u5b9e\u9645\u4e0a\uff0c\u6700\u597d\u4e00\u6b21\u5b8c\u6210\u6240\u6709\u64cd\u4f5c\uff0c\u65e0\u9700\u5212\u5206\u8bad\u7ec3\u6b65\u9aa4\uff0c\u66f4\u5c11\u7684\u6b65\u9aa4\u51cf\u5c11\u9519\u8bef\u3002\u65e0\u9700\u4f7f\u7528 Mel Spectrum \u7b49\u58f0\u5b66\u529f\u80fd\u3002\u5176\u5b9eMel\u662f\u597d\u7684\uff0c\u4f46\u662f\u88ab\u4eba\u4efb\u610f\u8bbe\u5b9a\u4e86\uff08\u6b21\u4f18\uff09\uff0c\u76f8\u4f4d\u4fe1\u606f\u4e5f\u4e22\u5931\u4e86\u3002\u7136\u800c\uff0c\u8fd9\u4e9b\u6a21\u578b\u4e4b\u6240\u4ee5\u4e0d\u5bb9\u6613\u4ece\u4e00\u5f00\u59cb\u5c31\u5f00\u53d1\u51fa\u6765\uff0c\u662f\u56e0\u4e3a\u5f88\u96be\u4e00\u6b21\u5168\u90e8\u5b8c\u6210\u3002<\/p>\n\n\n\n

    \u4f8b\u5982\uff0c\u4f5c\u4e3a\u8f93\u5165\u7684\u6587\u672c\u5728 5 \u79d2\u5185\u5927\u7ea6\u4e3a 20\uff0c\u5bf9\u4e8e\u97f3\u7d20\u5927\u7ea6\u4e3a 100\u3002\u4f46\u6ce2\u5f62\u662f 80,000 \u4e2a\u6837\u672c\uff08\u91c7\u6837\u7387\u4e3a 16 kHz\uff09\u3002\u56e0\u6b64\uff0c\u4e00\u65e6\u6210\u4e3a\u95ee\u9898\uff0c\u4e0d\u597d\u5b8c\u5168\u4e0e\u5176\u5339\u914d\uff08\u6587\u672c->\u97f3\u9891\u6837\u672c\uff09\uff0c\u4e0d\u5982\u4f7f\u7528\u4e2d\u7b49\u5206\u8fa8\u7387\u7684\u8868\u8fbe\u65b9\u5f0f\uff08\u5982Mel\uff09\u5206\u4e24\u6b65\u8fdb\u884c\u6bd4\u8f83\u7b80\u5355\u3002\u4f46\u662f\uff0c\u968f\u7740\u6280\u672f\u7684\u9010\u6e10\u53d1\u5c55\uff0c\u53ef\u4ee5\u627e\u5230\u4e00\u4e9b\u7528\u8fd9\u79cd Fully End-to-End \u65b9\u6cd5\u8bad\u7ec3\u7684\u6a21\u578b\u3002\u4f5c\u4e3a\u53c2\u8003\uff0c\u5728\u8bb8\u591a\u5904\u7406\u58f0\u5b66\u6a21\u578b\u7684\u8bba\u6587\u4e2d\uff0c\u4ed6\u4eec\u7ecf\u5e38\u4f7f\u7528\u672f\u8bed\u7aef\u5230\u7aef\u6a21\u578b\uff0c\u8fd9\u610f\u5473\u7740\u6587\u672c\u5206\u6790\u90e8\u5206\u5df2\u88ab\u4e00\u8d77\u5438\u6536\u5230\u4ed6\u4eec\u7684\u6a21\u578b\u4e2d\uff0c\u6216\u8005\u4ed6\u4eec\u53ef\u4ee5\u901a\u8fc7\u5c06\u58f0\u7801\u5668\u9644\u52a0\u5230\u4ed6\u4eec\u7684\u6a21\u578b\u6765\u751f\u6210\u97f3\u9891. \u5b83\u901a\u5e38\u7528\u4e8e\u8868\u793a\u80fd\u591f\u3002  <\/p>\n\n\n\n

    \u4e5f\u8bb8\u8fd9\u4e2a\u9886\u57df\u7684\u7b2c\u4e00\u4e2a\u662f Char2Wav ,\u8fd9\u662f\u8499\u7279\u5229\u5c14\u5927\u5b66\u540d\u4ebaYoshua Bengio\u6559\u6388\u56e2\u961f\u7684\u8bba\u6587\uff0c\u901a\u8fc7\u5c06\u5176\u56e2\u961f\u5236\u4f5c\u7684SampleRNN vocoder\u6dfb\u52a0\u5230Acoustic Model using seq2seq\u4e2d\u4e00\u6b21\u6027\u8bad\u7ec3\u800c\u6210\u3002ClariNet\u7684\u4e3b\u8981\u5185\u5bb9\u5176\u5b9e\u5c31\u662f\u8ba9WaveNet->IAF<\/strong>\u65b9\u6cd5\u7684Vocoder\u66f4\u52a0\u9ad8\u6548\u3002<\/p>\n\n\n\n

    FastSpeech 2\u4e5f\u662f\u5173\u4e8e\u4e00\u4e2a\u597d\u7684 Acoustic Model\uff0c\u8fd9\u7bc7\u8bba\u6587\u4e5f\u4ecb\u7ecd\u4e86\u4e00\u4e2a Fully End-to-End \u6a21\u578b\uff0c\u53eb\u505a FastSpeech 2s\u3002FastSpeech 2\u6a21\u578b\u9644\u52a0\u4e86\u4e00\u4e2aWaveNet\u58f0\u7801\u5668\uff0c\u4e3a\u4e86\u514b\u670d\u8bad\u7ec3\u7684\u56f0\u96be\uff0c\u91c7\u53d6\u4e86\u4f7f\u7528\u9884\u5148\u5236\u4f5c\u7684mel\u7f16\u7801\u5668\u7684\u65b9\u6cd5\u3002\u540d\u4e3aEATS\u7684\u6a21\u578b\u4f7f\u7528\u4ed6\u4eec\u56e2\u961f\uff08\u8c37\u6b4c\uff09\u521b\u5efa\u7684GAN-TTS\u4f5c\u4e3a\u58f0\u7801\u5668\uff0c\u521b\u5efa\u4e00\u4e2a\u65b0\u7684Acoustic Model\uff0c\u5e76\u4e00\u8d77\u8bad\u7ec3\u3002\u4f46\u662f\uff0c\u4e00\u6b21\u8bad\u7ec3\u5f88\u56f0\u96be\uff0c\u56e0\u6b64\u521b\u5efa\u5e76\u4f7f\u7528\u4e86\u4e2d\u7b49\u5206\u8fa8\u7387\u7684\u8868\u793a\u3002Wave-Tacotron\uff0c\u662f\u4e00\u79cd\u901a\u8fc7\u5c06\u58f0\u7801\u5668\u8fde\u63a5\u5230 Tacotron \u6765\u7acb\u5373\u8bad\u7ec3\u7684\u6a21\u578b\u3002\u8fd9\u91cc\u4f7f\u7528\u4e86\u6d41\u5f0f\u58f0\u7801\u5668\uff0c\u4f5c\u8005\u4f7f\u7528 Kingma\uff0c\u56e0\u6b64\u53ef\u4ee5\u5728\u4e0d\u663e\u7740\u964d\u4f4e\u6027\u80fd\u7684\u60c5\u51b5\u4e0b\u521b\u5efa\u66f4\u5feb\u7684\u6a21\u578b\u3002 <\/p>\n\n\n\n

    \u4e4b\u524dAcoustic Model\u90e8\u5206\u4ecb\u7ecd\u7684EfficientTTS\u4e5f\u4ecb\u7ecd\u4e86\u4e00\u79cd\u6a21\u578b\uff08EFTS-Wav\uff09\uff0c\u901a\u8fc7\u5c06decoder\u6362\u6210MelGAN\uff0c\u4ee5\u7aef\u5230\u7aef\u7684\u65b9\u5f0f\u8fdb\u884c\u8bad\u7ec3\u3002\u8be5\u6a21\u578b\u8fd8\u8868\u660e\uff0c\u5b83\u53ef\u4ee5\u663e\u7740\u52a0\u5feb\u97f3\u9891\u751f\u6210\u901f\u5ea6\uff0c\u540c\u65f6\u4ecd\u7136\u8868\u73b0\u826f\u597d\u3002Kakao \u56e2\u961f\u5f00\u53d1\u4e86\u4e00\u79cd\u540d\u4e3a Glow-TTS\u7684\u58f0\u5b66\u6a21\u578b\u548c\u4e00\u79cd\u540d\u4e3a HiFi-GAN\u7684\u58f0\u7801\u5668\u3002\u7136\u540e\u53ef\u4ee5\u5c06\u4e24\u8005\u653e\u5728\u4e00\u8d77\u4ee5\u521b\u5efa\u7aef\u5230\u7aef\u6a21\u578b,\u8fd9\u5c31\u662f VITS \uff0c\u5b83\u4f7f\u7528 VAE \u8fde\u63a5\u4e24\u4e2a\u90e8\u5206\uff0c\u5e76\u4f7f\u7528\u5bf9\u6297\u6027\u65b9\u6cd5\u8fdb\u884c\u6574\u4e2a\u8bad\u7ec3\uff0c\u63d0\u51fa\u4e86\u5177\u6709\u826f\u597d\u901f\u5ea6\u548c\u8d28\u91cf\u7684\u6a21\u578b\u3002<\/p>\n\n\n\n

    \u5ef6\u4e16\u5927\u5b66\/Naver \u8fd8\u5728 2021 \u5e74\u63a8\u51fa\u4e86 LiteTTS\uff0c\u8fd9\u662f\u4e00\u79cd\u9ad8\u6548\u7684\u5b8c\u5168\u7aef\u5230\u7aef TTS\u3002\u4f7f\u7528\u4e86\u524d\u9988\u53d8\u6362\u5668\u548c HiFi-GAN \u7ed3\u6784\u7684\u8f7b\u91cf\u7ea7\u7248\u672c\u3002\u7279\u522b\u662f\uff0c\u57df\u4f20\u8f93\u7f16\u7801\u5668\u7528\u4e8e\u5b66\u4e60\u4e0e\u97f5\u5f8b\u5d4c\u5165\u76f8\u5173\u7684\u6587\u672c\u4fe1\u606f\u3002\u817e\u8baf\u548c\u6d59\u6c5f\u5927\u5b66\u63d0\u51fa\u4e86\u4e00\u79cd\u540d\u4e3a FastDiff\u7684\u58f0\u7801\u5668\uff0c\u8fd8\u5f15\u5165\u4e86 FastDiff-TTS\uff0c\u8fd9\u662f\u4e00\u79cd\u7ed3\u5408 FastSpeech 2\u7684\u5b8c\u5168\u7aef\u5230\u7aef\u6a21\u578b\u3002Kakao \u8fd8\u5f15\u5165\u4e86 JETS\uff0c\u5b83\u53ef\u4ee5\u4e00\u8d77\u8bad\u7ec3 FastSpeech2 \u548c HiFi-GAN\u3002\u5fae\u8f6f\u5728\u5c06\u73b0\u6709\u7684 DelightfulTTS \u5347\u7ea7\u5230\u7248\u672c 2 \u7684\u540c\u65f6\uff0c\u4e5f\u5f15\u5165\u4e86 Fully End-to-End \u65b9\u6cd5\u3002\u8fd9\u91cc\uff0cVQ\u97f3\u9891\u7f16\u7801\u5668\u88ab\u7528\u4f5c\u4e2d\u95f4\u8868\u8fbe\u65b9\u6cd5\u3002 <\/p>\n\n\n\n

    \u622a\u6b62\u5230\u73b0\u5728\uff0c\u4e3b\u6d41\u7684\u8bed\u97f3\u5408\u6210\u6846\u67b6\u4ee5\u4ee5\u4e0a\u65b9\u6cd5\u4e3a\u4e3b\u6d41\u8fdb\u884c\u7814\u7a76\u53d1\u5c55\uff0c\u672a\u6765\u4f1a\u518d\u6b21\u7edf\u8ba1\u5e76\u6982\u8ff0\u6700\u65b0\u8bba\u6587\u4ee5\u53ca\u65b9\u6cd5\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

    \u539f\u521b\u00a0AI Pulse Text-to-Speech\uff08\u901a\u5e38\u7f29\u5199\u4e3aTTS\uff09\u662f\u6307\u4e00\u79cd\u5c06\u6587\u672c\u8f6c\u4e3a\u97f3\u9891\u7684\u6280\u672f\u3002 1.\u5386 … \u7ee7\u7eed\u9605\u8bfbTTS\u8c03\u7814 | \u8bed\u97f3\u5408\u6210\u7cfb\u5217\u57fa\u7840\u77e5\u8bc6\u53ca\u8bba\u6587\u603b\u7ed3<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[40,4,9,38,34],"tags":[],"_links":{"self":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/23204"}],"collection":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/comments?post=23204"}],"version-history":[{"count":14,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/23204\/revisions"}],"predecessor-version":[{"id":23222,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/23204\/revisions\/23222"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=23204"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=23204"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=23204"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}