{"id":1083774,"date":"2024-09-10T20:36:00","date_gmt":"2024-09-11T03:36:00","guid":{"rendered":"https:\/\/www.microsoft.com\/en-us\/research\/?post_type=msr-blog-post&p=1083774"},"modified":"2024-09-24T23:55:41","modified_gmt":"2024-09-25T06:55:41","slug":"vall-e-2-zh-cn","status":"publish","type":"msr-blog-post","link":"https:\/\/www.microsoft.com\/en-us\/research\/articles\/vall-e-2-zh-cn\/","title":{"rendered":"VALL-E 2\uff0c\u5927\u5e45\u63d0\u5347\u8bed\u97f3\u5927\u6a21\u578b\u7684\u7a33\u5065\u6027\u4e0e\u81ea\u7136\u5ea6"},"content":{"rendered":"\n

\u4f5c\u8005\uff1a\u5218\u6811\u6770<\/em><\/p>\n\n\n\n

\u7f16\u8005\u6309\uff1a\u6587\u672c\u5230\u8bed\u97f3\u5408\u6210\uff08Text-to-Speech\uff0cTTS\uff09\u662f\u4e00\u79cd\u5c06\u4e66\u9762\u6587\u5b57\u8f6c\u5316\u4e3a\u81ea\u7136\u8bed\u97f3\u7684\u6280\u672f\uff0c\u5728\u63d0\u9ad8\u65e0\u969c\u788d\u6027\u3001\u589e\u5f3a\u8de8\u8bed\u8a00\u4ea4\u6d41\u7b49\u65b9\u9762\u53d1\u6325\u7740\u91cd\u8981\u4f5c\u7528\u3002\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u6b64\u524d\u63a8\u51fa\u4e86\u7b2c\u4e00\u4e2a\u79bb\u6563\u7f16\u7801\u7684\u8bed\u97f3\u5927\u6a21\u578b VALL-E\uff0c\u5e76\u5728\u6b64\u57fa\u7840\u4e0a\u901a\u8fc7\u91cd\u590d\u611f\u77e5\u91c7\u6837\u548c\u5206\u7ec4\u7f16\u7801\u5efa\u6a21\u6280\u672f\u5c06\u5176\u5347\u7ea7\u4e3a VALL-E 2 \u7248\u672c\u3002\u65b0\u7248\u672c\u7a81\u7834\u4e86\u8bed\u97f3\u7a33\u5065\u6027\u3001\u81ea\u7136\u5ea6\u548c\u8bf4\u8bdd\u4eba\u76f8\u4f3c\u5ea6\u65b9\u9762\u7684\u754c\u9650\uff0c\u8ba9\u96f6\u6837\u672c TTS \u6027\u80fd\u5728 LibriSpeech \u548c VCTK \u6570\u636e\u96c6\u4e0a\u4e0e\u4eba\u7c7b\u6c34\u5e73\u76f8\u8fd1\u3002<\/p>\n\n\n\n


\n\n\n\n

\u8fd1\u5e74\u6765\u4eba\u5de5\u667a\u80fd\u6280\u672f\u7684\u98de\u901f\u53d1\u5c55\uff0c\u4e0d\u65ad\u63a8\u52a8\u7740\u6587\u672c\u5230\u8bed\u97f3\u5408\u6210\uff08Text-to-Speech\uff0cTTS\uff09\u6280\u672f\u7684\u8fb9\u754c\u3002TTS \u6280\u672f\u7684\u6301\u7eed\u4f18\u5316\u4e0e\u521b\u65b0\uff0c\u4e3a\u4eba\u4eec\u63d0\u4f9b\u4e86\u4e30\u5bcc\u3001\u4fbf\u6377\u7684\u8bed\u97f3\u4ea4\u4e92\u4f53\u9a8c\uff0c\u76f8\u5173\u7684\u7814\u7a76\u6210\u679c\u5728\u6559\u80b2\u3001\u5a31\u4e50\u4ee5\u53ca\u591a\u8bed\u8a00\u4ea4\u6d41\u7b49\u591a\u4e2a\u9886\u57df\u90fd\u6709\u7740\u5e7f\u6cdb\u7684\u5e94\u7528\u524d\u666f\u3002<\/p>\n\n\n\n

\u4f20\u7edf\u7684 TTS \u7cfb\u7edf\u4ec5\u4ec5\u4f7f\u7528\u6765\u81ea\u5f55\u97f3\u5ba4\u7684\u9ad8\u8d28\u91cf\u4e14\u5e72\u51c0\u7684\u8bed\u97f3\u6570\u636e\u8fdb\u884c\u8bad\u7ec3\uff0c\u5bfc\u81f4\u5176\u6cdb\u5316\u80fd\u529b\u8f83\u5dee\u3002\u5728\u96f6\u6837\u672c\u60c5\u5883\u4e0b\uff0c\u5bf9\u4e8e\u964c\u751f\u7684\u8bf4\u8bdd\u4eba\uff0c\u8bed\u97f3\u76f8\u4f3c\u5ea6\u548c\u81ea\u7136\u5ea6\u4e5f\u4f1a\u663e\u8457\u4e0b\u964d\u3002\u5bf9\u6b64\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u5c1d\u8bd5\u5c06 LLMs \u6280\u672f\u5f15\u5165\u8bed\u97f3\u5904\u7406\u4efb\u52a1\uff0c\u5e76\u6784\u5efa\u4e86\u7b2c\u4e00\u4e2a\u4f7f\u7528\u79bb\u6563\u7f16\u7801\u7684\u8bed\u97f3\u5927\u6a21\u578b VALL-E\u3002\u8fd9\u4e9b\u79bb\u6563\u7f16\u7801\u6765\u81ea\u73b0\u6210\u7684\u795e\u7ecf\u97f3\u9891\u7f16\u7801\u6a21\u578b\uff0c\u5e76\u5c06 TTS \u89c6\u4e3a\u6761\u4ef6\u8bed\u8a00\u6a21\u578b\u4efb\u52a1\uff0c\u4ece\u800c\u53ef\u4ee5\u5c55\u73b0\u51fa\u5f3a\u5927\u7684\u4e0a\u4e0b\u6587\u5b66\u4e60\u80fd\u529b\u3002\u53ea\u9700\u964c\u751f\u7684\u8bf4\u8bdd\u4eba\u63d0\u4f9b\u4e00\u6bb53\u79d2\u7684\u5f55\u97f3\u4f5c\u4e3a\u63d0\u793a\uff0cVALL-E \u5373\u53ef\u5408\u6210\u9ad8\u8d28\u91cf\u7684\u4e2a\u6027\u5316\u8bed\u97f3\u3002\u4f46\u7531\u4e8e\u81ea\u56de\u5f52\u5efa\u6a21\u548c\u968f\u673a\u91c7\u6837\u63a8\u65ad\uff0cVALL-E \u5728\u9c81\u68d2\u6027\u548c\u6548\u7387\u65b9\u9762\u4ecd\u6709\u5f85\u63d0\u5347\u3002<\/p>\n\n\n\n

\u5728\u6b64\u57fa\u7840\u4e0a\uff0c\u7814\u7a76\u5458\u4eec\u8bbe\u8ba1\u4e86 VALL-E 2<\/a>\uff0c\u5176\u5229\u7528\u91cd\u590d\u611f\u77e5\u91c7\u6837\uff08repetition aware sampling\uff09\u548c\u5206\u7ec4\u7f16\u7801\u5efa\u6a21\u6280\u672f\uff0c\u5b9e\u73b0\u4e86\u96f6\u6837\u672c TTS \u7684\u6027\u80fd\u5728 LibriSpeech \u548c VCTK \u6570\u636e\u96c6\u4e0a\u53ef\u8fbe\u5230\u4e0e\u4eba\u7c7b\u6c34\u5e73\u76f8\u5ab2\u7f8e\u7684\u7a0b\u5ea6\u3002\u5176\u4e2d\uff0c\u91cd\u590d\u611f\u77e5\u91c7\u6837\u901a\u8fc7\u8003\u8651\u89e3\u7801\u5386\u53f2\u4e2d\u7684\u6807\u8bb0\u91cd\u590d\uff0c\u6539\u8fdb\u4e86\u539f\u59cb\u7684\u6838\uff08Nucleus\uff09\u91c7\u6837\u8fc7\u7a0b\u3002\u5b83\u4e0d\u4ec5\u80fd\u7a33\u5b9a\u89e3\u7801\uff0c\u8fd8\u80fd\u907f\u514d\u5728 VALL-E \u4e2d\u9047\u5230\u7684\u65e0\u9650\u5faa\u73af\u95ee\u9898\u3002\u5206\u7ec4\u7f16\u7801\u5efa\u6a21\u5219\u5c06\u7f16\u89e3\u7801\u5668\u7f16\u7801\u7ec4\u7ec7\u6210\u7ec4\uff0c\u6709\u6548\u7f29\u77ed\u4e86\u5e8f\u5217\u957f\u5ea6\uff0c\u5728\u63d0\u9ad8\u63a8\u7406\u901f\u5ea6\u7684\u540c\u65f6\uff0c\u4e5f\u89e3\u51b3\u4e86\u957f\u5e8f\u5217\u5efa\u6a21\u7684\u96be\u9898\u3002<\/p>\n\n\n\n

\"graphical
\u56fe1\uff1aVALL-E 2 \u7684\u6846\u67b6<\/figcaption><\/figure>\n\n\n\n

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

\u901a\u8fc7\u8fd9\u4e24\u9879\u6280\u672f\uff0cVALL-E 2 \u5728\u8bed\u97f3\u7684\u7a33\u5065\u6027\u3001\u81ea\u7136\u5ea6\u548c\u8bf4\u8bdd\u4eba\u76f8\u4f3c\u5ea6\u65b9\u9762\u663e\u8457\u8d85\u8fc7\u4e86\u4ee5\u5f80\u7684\u7cfb\u7edf\u3002\u5373\u4f7f\u662f\u5728\u8f83\u5927\u590d\u6742\u6027\u6216\u91cd\u590d\u77ed\u8bed\u8fd9\u4e9b\u5177\u6709\u6311\u6218\u6027\u7684\u53e5\u5b50\u4e0a\uff0cVALL-E 2 \u4e5f\u80fd\u4e00\u81f4\u5730\u5408\u6210\u9ad8\u8d28\u91cf\u8bed\u97f3\u3002<\/p>\n\n\n\n

VALL-E 2 \u8bba\u6587\uff1ahttps:\/\/arxiv.org\/abs\/2406.05370 (opens in new tab)<\/span><\/a><\/p>\n\n\n\n

VALL-E 2 \u6f14\u793a\u9875\u9762\uff1ahttps:\/\/aka.ms\/valle2 (opens in new tab)<\/span><\/a><\/p>\n\n\n\n

\u91cd\u590d\u611f\u77e5\u91c7\u6837\u548c\u5206\u7ec4\u7f16\u7801\u5efa\u6a21\uff0c\u5b9e\u73b0\u6a21\u578b\u5347\u7ea7<\/h2>\n\n\n\n

VALL-E 2 \u91c7\u7528\u4e86\u540c VALL-E \u7c7b\u4f3c\u7684\u5c42\u6b21\u7ed3\u6784\uff1a\u4e00\u4e2a\u81ea\u56de\u5f52\uff08AR\uff09\u7f16\u7801\u8bed\u8a00\u6a21\u578b\u548c\u4e00\u4e2a\u975e\u81ea\u56de\u5f52\uff08NAR\uff09\u7f16\u7801\u8bed\u8a00\u6a21\u578b\u3002AR \u6a21\u578b\u4ee5\u81ea\u56de\u5f52\u65b9\u5f0f\u751f\u6210\u6bcf\u5e27\u7684\u7b2c\u4e00\u4e2a\u7f16\u7801\u5e8f\u5217\uff0c\u800c NAR \u6a21\u578b\u5219\u57fa\u4e8e\u4e4b\u524d\u7684\u7f16\u7801\u5e8f\u5217\u4ee5\u975e\u81ea\u56de\u5f52\u65b9\u5f0f\u751f\u6210\u6bcf\u4e2a\u5269\u4f59\u7684\u7f16\u7801\u5e8f\u5217\u3002\u4e24\u4e2a\u6a21\u578b\u90fd\u4f7f\u7528\u76f8\u540c\u7684 Transformer \u67b6\u6784\uff0c\u5305\u542b\u4e00\u4e2a\u6587\u672c\u5d4c\u5165\u5c42\u3001\u4e00\u4e2a\u7f16\u7801\u5d4c\u5165\u5c42\u3001Transformer \u5c42\u548c\u4e00\u4e2a\u7f16\u7801\u9884\u6d4b\u5c42\u3002\u4f46 AR \u6a21\u578b\u548c NAR \u6a21\u578b\u6709\u4e0d\u540c\u7684\u6ce8\u610f\u529b\u63a9\u7801\u7b56\u7565\uff1aAR \u6a21\u578b\u4f7f\u7528\u56e0\u679c\u6ce8\u610f\u529b\u7b56\u7565\uff0c\u800c NAR \u6a21\u578b\u4f7f\u7528\u5168\u6ce8\u610f\u529b\u7b56\u7565\u3002<\/p>\n\n\n\n

\"VALL-E
\u56fe2\uff1aVALL-E 2 \u4e2d\u7684 AR \u548c NAR \u6a21\u578b<\/figcaption><\/figure>\n\n\n\n

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

\u57fa\u4e8e\u8fc7\u5f80\u7ecf\u9a8c\uff0c\u7814\u7a76\u5458\u4eec\u53d1\u73b0 VALL-E \u5728\u63a8\u7406\u4e2d\u4f7f\u7528\u7684\u968f\u673a\u91c7\u6837\u53ef\u80fd\u4f1a\u5bfc\u81f4\u8f93\u51fa\u4e0d\u7a33\u5b9a\u3002\u5c3d\u7ba1\u9519\u8bef\u7f16\u7801\uff08\u56fe3\u4e2d\u7684\u7ea2\u8272\u65b9\u5757\uff09\u7684\u6982\u7387\u5f88\u4f4e\uff0c\u4f46\u7531\u4e8e\u91c7\u6837\u6b65\u9aa4\u592a\u591a\uff0c\u5b83\u4eec\u4ecd\u7136\u4e0d\u53ef\u907f\u514d\u5730\u4f1a\u88ab\u91c7\u6837\u5230\u3002\u4e3a\u4e86\u7a33\u5b9a\u63a8\u7406\u8fc7\u7a0b\uff0c\u901a\u5e38\u4f1a\u5229\u7528 Nucleus \u91c7\u6837\u6765\u4ece\u7d2f\u79ef\u6982\u7387\u4f4e\u4e8e\u9884\u8bbe\u9608\u503c\u7684\u6700\u53ef\u80fd\u6807\u8bb0\u96c6\u5408\u4e2d\u8fdb\u884c\u91c7\u6837\u3002Nucleus \u91c7\u6837\u65b9\u6cd5\u53ef\u4ee5\u51cf\u5c11\u8bf4\u9519\u8bcd\u7684\u9519\u8bef\uff0c\u4f46\u4e5f\u53ef\u80fd\u5bfc\u81f4\u6a21\u578b\u4e3a\u4e86\u51cf\u5c11\u9519\u8bef\u800c\u53ea\u751f\u6210\u9759\u97f3\u3002<\/p>\n\n\n\n

\u56e0\u6b64\uff0c\u4e3a\u4e86\u5e73\u8861\u968f\u673a\u91c7\u6837\u548c Nucleus \u91c7\u6837\uff0c\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86\u91cd\u590d\u611f\u77e5\u91c7\u6837\u7684\u65b9\u6cd5\u3002\u5728\u7ed9\u5b9a AR \u6a21\u578b\u9884\u6d4b\u7684\u6982\u7387\u5206\u5e03\u57fa\u7840\u4e0a\uff0c\u7814\u7a76\u5458\u4eec\u9996\u5148\u4f7f\u7528\u9884\u5b9a\u4e49\u7684 top-p \u503c\u901a\u8fc7 Nucleus \u91c7\u6837\u751f\u6210\u76ee\u6807\u7f16\u7801\u3002\u7136\u540e\u4f7f\u7528\u56fa\u5b9a\u7684\u7a97\u53e3\u5927\u5c0f\u6765\u8ba1\u7b97\u9884\u6d4b\u7f16\u7801\u5728\u524d\u9762\u7684\u7f16\u7801\u5e8f\u5217\u4e2d\u7684\u91cd\u590d\u6bd4\u4f8b\u3002\u5982\u679c\u91cd\u590d\u6bd4\u4f8b\u8d85\u8fc7\u9884\u5b9a\u4e49\u7684\u91cd\u590d\u9608\u503c\uff0c\u7814\u7a76\u5458\u4eec\u5c31\u4f1a\u4f7f\u7528\u968f\u673a\u91c7\u6837\u4ece\u539f\u59cb\u6982\u7387\u5206\u5e03\u4e2d\u83b7\u5f97\u65b0\u7684\u9884\u6d4b\u7ed3\u679c\u6765\u66ff\u6362\u539f\u6765\u7684\u76ee\u6807\u7f16\u7801\u3002\u901a\u8fc7\u8fd9\u79cd\u91cd\u590d\u611f\u77e5\u91c7\u6837\u7684\u65b9\u6cd5\uff0c\u89e3\u7801\u8fc7\u7a0b\u65e2\u53ef\u4ee5\u53d7\u76ca\u4e8e Nucleus \u91c7\u6837\u7684\u7a33\u5b9a\u6027\uff0c\u8fd8\u53ef\u4ee5\u501f\u52a9\u968f\u673a\u91c7\u6837\u907f\u514d\u9677\u5165\u9759\u97f3\u7684\u65e0\u9650\u5faa\u73af\u3002<\/p>\n\n\n\n

\"VALL-E
\u56fe3\uff1aVALL-E 2 \u4e2d\u7684\u91cd\u590d\u611f\u77e5\u91c7\u6837\u548c\u5206\u7ec4\u7f16\u7801\u5efa\u6a21<\/figcaption><\/figure>\n\n\n\n

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

\u540c\u65f6\uff0c\u7814\u7a76\u5458\u4eec\u53d1\u73b0\uff0cVALL-E \u7684\u81ea\u56de\u5f52\u67b6\u6784\u53d7\u9650\u4e8e\u73b0\u6210\u97f3\u9891\u7f16\u7801\u6a21\u578b\u7684\u9ad8\u5e27\u7387\uff0c\u4ece\u800c\u5bfc\u81f4\u63a8\u7406\u901f\u5ea6\u7f13\u6162\uff0c\u7279\u522b\u662f\u5728\u81ea\u56de\u5f52\u6a21\u578b\u7684\u63a8\u7406\u8fc7\u7a0b\u4e2d\u5c24\u4e3a\u663e\u8457\u3002\u4e3a\u4e86\u52a0\u901f\u63a8\u7406\u8fc7\u7a0b\uff0c\u7814\u7a76\u5458\u4eec\u5728 VALL-E 2 \u4e2d\u91c7\u7528\u4e86\u5206\u7ec4\u7f16\u7801\u5efa\u6a21\u65b9\u6cd5\uff0c\u5c06\u7f16\u7801\u5e8f\u5217\u5212\u5206\u4e3a\u4e00\u5b9a\u5927\u5c0f\u7684\u5206\u7ec4\uff0c\u5e76\u5c06\u6bcf\u7ec4\u7f16\u7801\u5efa\u6a21\u4e3a\u4e00\u6b65\u3002\u5728\u81ea\u56de\u5f52\u6a21\u578b\u4e2d\uff0c\u7814\u7a76\u5458\u4eec\u5229\u7528\u5206\u7ec4\u5d4c\u5165\u5c42\u5c06\u7f16\u7801\u5d4c\u5165\u7ed3\u679c\u6620\u5c04\u5230\u5206\u7ec4\u5d4c\u5165\u5e76\u4f5c\u4e3a\u7f51\u7edc\u7684\u8f93\u5165\uff0c\u5e76\u5728\u8f93\u51fa\u5c42\u4f7f\u7528\u5206\u7ec4\u9884\u6d4b\u5c42\u5bf9\u4e00\u7ec4\u4e2d\u7684\u7f16\u7801\u8fdb\u884c\u9884\u6d4b\u3002\u901a\u8fc7\u8fd9\u79cd\u65b9\u5f0f\uff0c\u73b0\u6709\u795e\u7ecf\u97f3\u9891\u7f16\u7801\u6a21\u578b\u7684\u5e27\u7387\u9650\u5236\u5f97\u4ee5\u6446\u8131\uff0c\u5e27\u7387\u6210\u500d\u6570\u5730\u964d\u4f4e\u5f97\u5230\u5b9e\u73b0\u3002\u8fd9\u4e0d\u4ec5\u6709\u5229\u4e8e\u63a8\u7406\u6548\u7387\u7684\u63d0\u9ad8\uff0c\u8fd8\u53ef\u4ee5\u7f13\u89e3\u957f\u4e0a\u4e0b\u6587\u5efa\u6a21\u7684\u95ee\u9898\u5e76\u63d0\u5347\u6574\u4f53\u8bed\u97f3\u8d28\u91cf\u3002<\/p>\n\n\n\n

\u7a33\u5065\u6027\u3001\u81ea\u7136\u5ea6\u548c\u76f8\u4f3c\u5ea6\u7684\u5927\u5e45\u63d0\u5347<\/h2>\n\n\n\n

\u4e3a\u4e86\u5c55\u793a VALL-E 2 \u7684\u6027\u80fd\uff0c\u7814\u7a76\u5458\u4eec\u5728 LibriSpeech \u548c VCTK \u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u4e86\u5b9e\u9a8c\uff0c\u5e76\u5728\u7a33\u5065\u6027\u3001\u81ea\u7136\u5ea6\u548c\u97f3\u8272\u76f8\u4f3c\u5ea6\u8bc4\u5206\u65b9\u9762\u4e0e\u591a\u4e2a\u57fa\u7ebf\u6a21\u578b\u8fdb\u884c\u4e86\u6bd4\u8f83\u3002\u8fd9\u4e9b\u8bc4\u5206\u662f\u6839\u636e\u539f\u59cb\u8bba\u6587\u4e2d\u62a5\u544a\u7684\u7ed3\u679c\u8ba1\u7b97\u5f97\u5230\u7684\u76f8\u5bf9\u6570\u503c\uff0c\u4e0d\u8003\u8651\u6a21\u578b\u67b6\u6784\u548c\u8bad\u7ec3\u6570\u636e\u7684\u5dee\u5f02\u3002<\/p>\n\n\n\n

\u4ece\u56fe4\u4e2d\u53ef\u4ee5\u770b\u5230\uff0cVALL-E 2 \u76f8\u6bd4\u4ee5\u524d\u7684\u65b9\u6cd5\u663e\u8457\u63d0\u5347\u4e86\u5404\u4e2a\u6307\u6807\u7684\u6027\u80fd\uff0c\u5e76\u4e14\u9996\u6b21\u5b9e\u73b0\u4e86\u4e0e\u4eba\u7c7b\u540c\u7b49\u6c34\u5e73\u76f8\u8fd1\u7684\u96f6\u6837\u672c TTS\u3002\u4e0e\u4eba\u7c7b\u6c34\u5e73\u76f8\u8fd1\u6307\u7684\u662f VALL-E 2 \u7684\u7a33\u5065\u6027\u3001\u81ea\u7136\u5ea6\u548c\u76f8\u4f3c\u5ea6\u6307\u6807\u8d85\u8d8a\u4e86\u771f\u5b9e\u6837\u672c\uff0c\u5373 WER(GroundTruth)-WER(VALL-E 2)>0\uff0cCMOS(VALL-E 2)-CMOS(GroundTruth)>0 \u548c SMOS(VALL-E 2)-SMOS(GroundTruth)>0\uff0c\u8fd9\u610f\u5473\u7740 VALL-E 2 \u80fd\u591f\u751f\u6210\u4e0e\u539f\u59cb\u8bf4\u8bdd\u4eba\u58f0\u97f3\u8f83\u4e3a\u4e00\u81f4\u7684\u51c6\u786e\u81ea\u7136\u7684\u8bed\u97f3\u3002\u4f46\u9700\u8981\u6ce8\u610f\uff0c\u8fd9\u4e00\u7ed3\u8bba\u4ec5\u4ec5\u57fa\u4e8e LibriSpeech \u548c VCTK \u6570\u636e\u96c6\u7684\u5b9e\u9a8c\u7ed3\u679c\u3002<\/p>\n\n\n\n

\"VALL-E
\u56fe4\uff1aVALL-E 2 \u540c\u4ee5\u5f80\u7cfb\u7edf\u7684\u6bd4\u8f83\u7ed3\u679c<\/figcaption><\/figure>\n\n\n\n

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

\u91cd\u590d\u611f\u77e5\u91c7\u6837\u548c\u5206\u7ec4\u7f16\u7801\u5efa\u6a21\u4e24\u79cd\u65b9\u6cd5\u7684\u5f15\u5165\uff0c\u4f7f\u5f97 VALL-E 2 \u80fd\u591f\u53ef\u9760\u5730\u5408\u6210\u590d\u6742\u53e5\u5b50\u7684\u8bed\u97f3\uff0c\u5305\u62ec\u90a3\u4e9b\u96be\u4ee5\u9605\u8bfb\u6216\u5305\u542b\u5927\u91cf\u91cd\u590d\u77ed\u8bed\u7684\u53e5\u5b50\u3002VALL-E 2 \u7684\u6280\u672f\u4f18\u52bf\uff0c\u53ef\u4ee5\u4e3a\u4eba\u5de5\u667a\u80fd\u5411\u5584\u7684\u573a\u666f\u505a\u51fa\u8d21\u732e\uff0c\u4f8b\u5982\u4e3a\u5931\u8bed\u75c7\u60a3\u8005\u6216\u808c\u840e\u7f29\u4fa7\u7d22\u786c\u5316\u75c7\u7684\u60a3\u8005\u751f\u6210\u8bed\u97f3\u7b49\u3002<\/p>\n\n\n\n

\u6ce8\uff1aVALL-E 2 \u662f\u4e00\u4e2a\u7eaf\u7cb9\u7684\u7814\u7a76\u9879\u76ee\uff0c\u76ee\u524d\u6ca1\u6709\u8ba1\u5212\u7eb3\u5165\u4ea7\u54c1\u6216\u5411\u516c\u4f17\u5f00\u653e\u3002VALL-E 2 \u53ef\u4ee5\u5408\u6210\u4fdd\u6301\u8bf4\u8bdd\u8005\u97f3\u8272\u7684\u8bed\u97f3\uff0c\u53ef\u7528\u4e8e\u6559\u80b2\u5b66\u4e60\u3001\u5a31\u4e50\u3001\u65b0\u95fb\u3001\u81ea\u521b\u5185\u5bb9\u3001\u65e0\u969c\u788d\u529f\u80fd\u3001\u4e92\u52a8\u8bed\u97f3\u5e94\u7b54\u7cfb\u7edf\u3001\u7ffb\u8bd1\u3001\u804a\u5929\u673a\u5668\u4eba\u7b49\u3002\u867d\u7136 VALL-E 2 \u53ef\u4ee5\u751f\u6210\u5ab2\u7f8e\u771f\u4eba\u58f0\u97f3\u7684\u8bed\u97f3\uff0c\u4f46\u5176\u76f8\u4f3c\u5ea6\u548c\u81ea\u7136\u5ea6\u4ecd\u53d6\u51b3\u4e8e\u8bed\u97f3\u63d0\u793a\u7684\u957f\u5ea6\u548c\u8d28\u91cf\u3001\u80cc\u666f\u566a\u97f3\u4ee5\u53ca\u5176\u4ed6\u56e0\u7d20\u3002\u8be5\u6a21\u578b\u53ef\u80fd\u5728\u8bef\u7528\u65b9\u9762\u5b58\u5728\u6f5c\u5728\u98ce\u9669\uff0c\u4f8b\u5982\u4f2a\u9020\u8bed\u97f3\u8bc6\u522b\u6216\u5192\u5145\u7279\u5b9a\u8bf4\u8bdd\u8005\u3002\u5728\u8bed\u97f3\u5408\u6210\u7814\u7a76\u4e2d\uff0c\u6211\u4eec\u5df2\u7ecf\u83b7\u5f97\u7528\u6237\u7684\u6388\u6743\uff0c\u5c06\u5176\u4f5c\u4e3a\u5b9e\u9a8c\u7684\u76ee\u6807\u8bf4\u8bdd\u8005\uff0c\u5982\u679c\u8be5\u6a21\u578b\u9700\u8981\u63a8\u5e7f\u5230\u73b0\u5b9e\u4e16\u754c\u4e2d\u7684\u672a\u89c1\u8fc7\u7684\u8bf4\u8bdd\u8005\uff0c\u5e94\u786e\u4fdd\u8bf4\u8bdd\u8005\u540c\u610f\u4f7f\u7528\u5176\u58f0\u97f3\u7684\u534f\u8bae\u548c\u5408\u6210\u8bed\u97f3\u68c0\u6d4b\u6a21\u578b\u3002\u5982\u679c\u60a8\u53d1\u73b0 VALL-E 2 \u88ab\u6ee5\u7528\u3001\u975e\u6cd5\u4f7f\u7528\u6216\u4fb5\u72af\u4e86\u60a8\u7684\u6743\u5229\u6216\u4ed6\u4eba\u7684\u6743\u5229\uff0c\u53ef\u4ee5\u5728\u5fae\u8f6f\u7684\u6ee5\u7528\u62a5\u544a\u95e8\u6237\u7f51\u7ad9\uff08https:\/\/msrc.microsoft.com\/report\/\uff09\u8fdb\u884c\u4e3e\u62a5\u3002<\/em><\/p>\n\n\n\n

\u968f\u7740\u4eba\u5de5\u667a\u80fd\u6280\u672f\u7684\u5feb\u901f\u53d1\u5c55\uff0c\u786e\u4fdd\u76f8\u5173\u6280\u672f\u80fd\u88ab\u4eba\u4eec\u4fe1\u8d56\u662f\u4e00\u4e2a\u9700\u8981\u653b\u575a\u7684\u95ee\u9898\u3002\u5fae\u8f6f\u4e3b\u52a8\u91c7\u53d6\u4e86\u4e00\u7cfb\u5217\u63aa\u65bd\u6765\u9884\u5224\u548c\u964d\u4f4e\u4eba\u5de5\u667a\u80fd\u6280\u672f\u6240\u5e26\u6765\u7684\u98ce\u9669\u3002\u5fae\u8f6f\u81f4\u529b\u4e8e\u4f9d\u7167\u4ee5\u4eba\u4e3a\u672c\u7684\u4f26\u7406\u539f\u5219\u63a8\u8fdb\u4eba\u5de5\u667a\u80fd\u7684\u53d1\u5c55\uff0c\u65e9\u57282018\u5e74\u5c31\u53d1\u5e03\u4e86\u201c\u516c\u5e73\u3001\u5305\u5bb9\u3001\u53ef\u9760\u4e0e\u5b89\u5168\u3001\u900f\u660e\u3001\u9690\u79c1\u4e0e\u4fdd\u969c\u3001\u8d1f\u8d23\u201d\u516d\u4e2a\u8d1f\u8d23\u4efb\u7684\u4eba\u5de5\u667a\u80fd\u539f\u5219\uff08Responsible AI Principles\uff09\uff0c\u968f\u540e\u53c8\u53d1\u5e03\u4e86\u8d1f\u8d23\u4efb\u7684\u4eba\u5de5\u667a\u80fd\u6807\u51c6\uff08Responsible AI Standards\uff09\u5c06\u5404\u9879\u539f\u5219\u5b9e\u65bd\u843d\u5730\uff0c\u5e76\u8bbe\u7f6e\u4e86\u6cbb\u7406\u67b6\u6784\u786e\u4fdd\u5404\u56e2\u961f\u628a\u5404\u9879\u539f\u5219\u548c\u6807\u51c6\u843d\u5b9e\u5230\u65e5\u5e38\u5de5\u4f5c\u4e2d\u3002\u5fae\u8f6f\u4e5f\u6301\u7eed\u4e0e\u5168\u7403\u7684\u7814\u7a76\u4eba\u5458\u548c\u5b66\u672f\u673a\u6784\u5408\u4f5c\uff0c\u4e0d\u65ad\u63a8\u8fdb\u8d1f\u8d23\u4efb\u7684\u4eba\u5de5\u667a\u80fd\u7684\u5b9e\u8df5\u548c\u6280\u672f\u3002<\/em><\/p>\n","protected":false},"excerpt":{"rendered":"

\u4f5c\u8005\uff1a\u5218\u6811\u6770 \u7f16\u8005\u6309\uff1a\u6587\u672c\u5230\u8bed\u97f3\u5408\u6210\uff08Text-to-Speech\uff0cTTS\uff09\u662f\u4e00\u79cd\u5c06\u4e66\u9762\u6587\u5b57\u8f6c\u5316\u4e3a\u81ea\u7136\u8bed\u97f3\u7684\u6280\u672f\uff0c\u5728\u63d0\u9ad8\u65e0\u969c\u788d\u6027\u3001\u589e\u5f3a\u8de8\u8bed\u8a00\u4ea4\u6d41\u7b49\u65b9\u9762\u53d1\u6325\u7740\u91cd\u8981\u4f5c\u7528\u3002\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u6b64\u524d\u63a8\u51fa\u4e86\u7b2c\u4e00\u4e2a\u79bb\u6563\u7f16\u7801\u7684\u8bed\u97f3\u5927\u6a21\u578b VALL-E\uff0c\u5e76\u5728\u6b64\u57fa\u7840\u4e0a\u901a\u8fc7\u91cd\u590d\u611f\u77e5\u91c7\u6837\u548c\u5206\u7ec4\u7f16\u7801\u5efa\u6a21\u6280\u672f\u5c06\u5176\u5347\u7ea7\u4e3a VALL-E 2 \u7248\u672c\u3002\u65b0\u7248\u672c\u7a81\u7834\u4e86\u8bed\u97f3\u7a33\u5065\u6027\u3001\u81ea\u7136\u5ea6\u548c\u8bf4\u8bdd\u4eba\u76f8\u4f3c\u5ea6\u65b9\u9762\u7684\u754c\u9650\uff0c\u8ba9\u96f6\u6837\u672c TTS \u6027\u80fd\u5728 LibriSpeech \u548c VCTK \u6570\u636e\u96c6\u4e0a\u4e0e\u4eba\u7c7b\u6c34\u5e73\u76f8\u8fd1\u3002 \u8fd1\u5e74\u6765\u4eba\u5de5\u667a\u80fd\u6280\u672f\u7684\u98de\u901f\u53d1\u5c55\uff0c\u4e0d\u65ad\u63a8\u52a8\u7740\u6587\u672c\u5230\u8bed\u97f3\u5408\u6210\uff08Text-to-Speech\uff0cTTS\uff09\u6280\u672f\u7684\u8fb9\u754c\u3002TTS \u6280\u672f\u7684\u6301\u7eed\u4f18\u5316\u4e0e\u521b\u65b0\uff0c\u4e3a\u4eba\u4eec\u63d0\u4f9b\u4e86\u4e30\u5bcc\u3001\u4fbf\u6377\u7684\u8bed\u97f3\u4ea4\u4e92\u4f53\u9a8c\uff0c\u76f8\u5173\u7684\u7814\u7a76\u6210\u679c\u5728\u6559\u80b2\u3001\u5a31\u4e50\u4ee5\u53ca\u591a\u8bed\u8a00\u4ea4\u6d41\u7b49\u591a\u4e2a\u9886\u57df\u90fd\u6709\u7740\u5e7f\u6cdb\u7684\u5e94\u7528\u524d\u666f\u3002 \u4f20\u7edf\u7684 TTS \u7cfb\u7edf\u4ec5\u4ec5\u4f7f\u7528\u6765\u81ea\u5f55\u97f3\u5ba4\u7684\u9ad8\u8d28\u91cf\u4e14\u5e72\u51c0\u7684\u8bed\u97f3\u6570\u636e\u8fdb\u884c\u8bad\u7ec3\uff0c\u5bfc\u81f4\u5176\u6cdb\u5316\u80fd\u529b\u8f83\u5dee\u3002\u5728\u96f6\u6837\u672c\u60c5\u5883\u4e0b\uff0c\u5bf9\u4e8e\u964c\u751f\u7684\u8bf4\u8bdd\u4eba\uff0c\u8bed\u97f3\u76f8\u4f3c\u5ea6\u548c\u81ea\u7136\u5ea6\u4e5f\u4f1a\u663e\u8457\u4e0b\u964d\u3002\u5bf9\u6b64\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u5c1d\u8bd5\u5c06 LLMs \u6280\u672f\u5f15\u5165\u8bed\u97f3\u5904\u7406\u4efb\u52a1\uff0c\u5e76\u6784\u5efa\u4e86\u7b2c\u4e00\u4e2a\u4f7f\u7528\u79bb\u6563\u7f16\u7801\u7684\u8bed\u97f3\u5927\u6a21\u578b VALL-E\u3002\u8fd9\u4e9b\u79bb\u6563\u7f16\u7801\u6765\u81ea\u73b0\u6210\u7684\u795e\u7ecf\u97f3\u9891\u7f16\u7801\u6a21\u578b\uff0c\u5e76\u5c06 TTS \u89c6\u4e3a\u6761\u4ef6\u8bed\u8a00\u6a21\u578b\u4efb\u52a1\uff0c\u4ece\u800c\u53ef\u4ee5\u5c55\u73b0\u51fa\u5f3a\u5927\u7684\u4e0a\u4e0b\u6587\u5b66\u4e60\u80fd\u529b\u3002\u53ea\u9700\u964c\u751f\u7684\u8bf4\u8bdd\u4eba\u63d0\u4f9b\u4e00\u6bb53\u79d2\u7684\u5f55\u97f3\u4f5c\u4e3a\u63d0\u793a\uff0cVALL-E \u5373\u53ef\u5408\u6210\u9ad8\u8d28\u91cf\u7684\u4e2a\u6027\u5316\u8bed\u97f3\u3002\u4f46\u7531\u4e8e\u81ea\u56de\u5f52\u5efa\u6a21\u548c\u968f\u673a\u91c7\u6837\u63a8\u65ad\uff0cVALL-E \u5728\u9c81\u68d2\u6027\u548c\u6548\u7387\u65b9\u9762\u4ecd\u6709\u5f85\u63d0\u5347\u3002 \u5728\u6b64\u57fa\u7840\u4e0a\uff0c\u7814\u7a76\u5458\u4eec\u8bbe\u8ba1\u4e86 VALL-E 2\uff0c\u5176\u5229\u7528\u91cd\u590d\u611f\u77e5\u91c7\u6837\uff08repetition aware sampling\uff09\u548c\u5206\u7ec4\u7f16\u7801\u5efa\u6a21\u6280\u672f\uff0c\u5b9e\u73b0\u4e86\u96f6\u6837\u672c TTS \u7684\u6027\u80fd\u5728 LibriSpeech \u548c VCTK \u6570\u636e\u96c6\u4e0a\u53ef\u8fbe\u5230\u4e0e\u4eba\u7c7b\u6c34\u5e73\u76f8\u5ab2\u7f8e\u7684\u7a0b\u5ea6\u3002\u5176\u4e2d\uff0c\u91cd\u590d\u611f\u77e5\u91c7\u6837\u901a\u8fc7\u8003\u8651\u89e3\u7801\u5386\u53f2\u4e2d\u7684\u6807\u8bb0\u91cd\u590d\uff0c\u6539\u8fdb\u4e86\u539f\u59cb\u7684\u6838\uff08Nucleus\uff09\u91c7\u6837\u8fc7\u7a0b\u3002\u5b83\u4e0d\u4ec5\u80fd\u7a33\u5b9a\u89e3\u7801\uff0c\u8fd8\u80fd\u907f\u514d\u5728 VALL-E \u4e2d\u9047\u5230\u7684\u65e0\u9650\u5faa\u73af\u95ee\u9898\u3002\u5206\u7ec4\u7f16\u7801\u5efa\u6a21\u5219\u5c06\u7f16\u89e3\u7801\u5668\u7f16\u7801\u7ec4\u7ec7\u6210\u7ec4\uff0c\u6709\u6548\u7f29\u77ed\u4e86\u5e8f\u5217\u957f\u5ea6\uff0c\u5728\u63d0\u9ad8\u63a8\u7406\u901f\u5ea6\u7684\u540c\u65f6\uff0c\u4e5f\u89e3\u51b3\u4e86\u957f\u5e8f\u5217\u5efa\u6a21\u7684\u96be\u9898\u3002 \u901a\u8fc7\u8fd9\u4e24\u9879\u6280\u672f\uff0cVALL-E 2 \u5728\u8bed\u97f3\u7684\u7a33\u5065\u6027\u3001\u81ea\u7136\u5ea6\u548c\u8bf4\u8bdd\u4eba\u76f8\u4f3c\u5ea6\u65b9\u9762\u663e\u8457\u8d85\u8fc7\u4e86\u4ee5\u5f80\u7684\u7cfb\u7edf\u3002\u5373\u4f7f\u662f\u5728\u8f83\u5927\u590d\u6742\u6027\u6216\u91cd\u590d\u77ed\u8bed\u8fd9\u4e9b\u5177\u6709\u6311\u6218\u6027\u7684\u53e5\u5b50\u4e0a\uff0cVALL-E 2 \u4e5f\u80fd\u4e00\u81f4\u5730\u5408\u6210\u9ad8\u8d28\u91cf\u8bed\u97f3\u3002 VALL-E 2 \u8bba\u6587\uff1ahttps:\/\/arxiv.org\/abs\/2406.05370 (opens in new tab) VALL-E 2 \u6f14\u793a\u9875\u9762\uff1ahttps:\/\/aka.ms\/valle2 (opens in new […]<\/p>\n","protected":false},"author":34512,"featured_media":1083768,"template":"","meta":{"msr-url-field":"","msr-podcast-episode":"","msrModifiedDate":"","msrModifiedDateEnabled":false,"ep_exclude_from_search":false,"_classifai_error":"","msr-content-parent":1012650,"footnotes":""},"research-area":[],"msr-locale":[268881],"msr-post-option":[],"class_list":["post-1083774","msr-blog-post","type-msr-blog-post","status-publish","has-post-thumbnail","hentry","msr-locale-zh_cn"],"msr_assoc_parent":{"id":1012650,"type":"lab"},"_links":{"self":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1083774"}],"collection":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post"}],"about":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/types\/msr-blog-post"}],"author":[{"embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/users\/34512"}],"version-history":[{"count":6,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1083774\/revisions"}],"predecessor-version":[{"id":1084164,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1083774\/revisions\/1084164"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/media\/1083768"}],"wp:attachment":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/media?parent=1083774"}],"wp:term":[{"taxonomy":"msr-research-area","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/research-area?post=1083774"},{"taxonomy":"msr-locale","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-locale?post=1083774"},{"taxonomy":"msr-post-option","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-post-option?post=1083774"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}