{"id":15743,"date":"2024-07-06T15:29:00","date_gmt":"2024-07-06T07:29:00","guid":{"rendered":"http:\/\/139.9.1.231\/?p=15743"},"modified":"2024-09-20T17:09:17","modified_gmt":"2024-09-20T09:09:17","slug":"gpt-4o-audio-llm","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2024\/07\/06\/gpt-4o-audio-llm\/","title":{"rendered":"GPT-4o\u80cc\u540e\u7684\u8bed\u97f3\u6280\u672f"},"content":{"rendered":"\n

5\u670814\u65e5\u51cc\u6668\uff0cOpenAI\u63a8\u51fa\u4e86\u6700\u65b0\u7684\u751f\u6210\u6a21\u578bGPT-4o\uff0c\u5e26\u6765\u4e86\u4e00\u7cfb\u5217\u9707\u64bc\u7684\u529f\u80fd\uff0c\u7528\u6280\u672f\u5f7b\u5e95\u98a0\u8986\u4e86\u4ea7\u54c1\u5f62\u6001\u3002\u4ea7\u54c1\u6700\u5927\u7684\u4eae\u70b9\u5728\u4e8e\uff1a\u4ee5\u8fd1\u4e4e\u5b8c\u7f8e\u7684\u4ea4\u4e92\u65b9\u5f0f\uff0c\u4e3a\u6bcf\u4f4d\u7528\u6237\u5e26\u6765GPT-4\u7ea7\u522b\u7684\u667a\u80fd\u4f53\u9a8c\u3002\u5728\u8bed\u97f3\u65b9\u9762\uff0cGPT-4o\u505a\u5230\u4e86\u5b9e\u65f6\u4f4e\u5ef6\u8fdf\uff0c\u5e73\u5747\u54cd\u5e94\u65f6\u95f4\u4e0e\u4eba\u7c7b\u53cd\u5e94\u901f\u5ea6\u76f8\u5f53\uff0c\u8f93\u51fa\u7684\u8bed\u97f3\u80fd\u591f\u7406\u89e3\u6781\u5ea6\u8d34\u5408\u5bf9\u8bdd\u4e0a\u4e0b\u6587\uff0c\u80fd\u591f\u7406\u89e3\u4eba\u7c7b\u7684\u60c5\u611f\u60c5\u7eea\uff0c\u542c\u89c9\u8d28\u91cf\u4e0a\u4f73\uff0c\u4e0e\u771f\u4eba\u65e0\u5f02\u3002<\/p><\/blockquote>\n\n\n\n

OpenAI\u7684\u535a\u5ba2\uff1a<\/strong>https:\/\/openai.com\/index\/hello-gpt-4o\/<\/strong><\/a><\/p>\n\n\n\n\n\n

GPT-4o\u662f\u4e00\u4e2aany2any\u7684\u591a\u6a21\u6001\u6a21\u578b\uff0c\u80fd\u591f\u63a5\u53d7\u6587\u672c\u3001\u97f3\u9891\u3001\u56fe\u50cf\u3001\u89c6\u9891\u7b49\u591a\u6a21\u6001\u8f93\u5165\uff0c\u4e5f\u80fd\u591f\u751f\u6210\u5305\u542b\u6587\u672c\u3001\u8bed\u97f3\u3001\u56fe\u50cf\u548c\u89c6\u9891\u7b49\u6df7\u5408\u5185\u5bb9\u7684\u591a\u6a21\u6001\u8f93\u51fa\u3002\u9650\u4e8e\u7bc7\u5e45\uff0c\u672c\u6587\u4e3b\u8981\u8c08\u8c08\u8bed\u97f3\u591a\u6a21\u6001\u7684\u5b9e\u73b0\uff0c\u5e76\u5206\u4eab\u4e00\u4e9b\u5bf9\u4e8e\u8bed\u97f3\u7814\u7a76\u672a\u6765\u53d1\u5c55\u7684\u770b\u6cd5\u3002<\/p>\n\n\n\n

\u5f53\u6211\u4eec\u4e3b\u8981\u5173\u6ce8\u6587\u672c\u548c\u8bed\u97f3\u6a21\u6001\u65f6\uff0cGPT-4o\u5176\u5b9e\u5c31\u662f\u4e00\u4e2a\u8bed\u97f3\u8bed\u8a00\u6a21\u578b(speech language model, SLM)\u3002\u8be5SLM\u540c\u65f6\u5177\u5907\u8bed\u97f3\u7406\u89e3\u80fd\u529b\u548c\u8bed\u97f3\u5408\u6210\u80fd\u529b\uff0c\u8f93\u5165\u7aef\u548c\u8f93\u51fa\u7aef\u5747\u652f\u6301\u6587\u672c\u548c\u8bed\u97f3\u7684\u6df7\u5408\u591a\u6a21\u6001\u3002\u90a3\u4e48\uff0c\u8fd9\u4e00SLM\u5e94\u8be5\u5982\u4f55\u5b9e\u73b0\u5462\uff1f\u5728\u5927\u8bed\u8a00\u6a21\u578b(large language model, LLM)\u6ee5\u89de\u7684\u4eca\u65e5\uff0c\u4e0d\u96be\u60f3\u5230\u8fd9\u6837\u4e00\u79cd\u65b9\u6cd5\uff1a\u5c06\u8fde\u7eed\u7684\u8bed\u97f3\u6570\u636e\u79bb\u6563\u5316\u6210\u5982\u540c\u5355\u8bcd(\u6216\u8005\u79f0token\uff0c\u8bcd\u5143)\u4e00\u6837\u7684\u8868\u793a\uff0c\u5e76\u5165\u5230LLM\u7684\u8bcd\u8868\u4e2d\uff0c\u518d\u8d70\u4e00\u904d\u8bad\u7ec3LLM\u7684\u8001\u8def\u3002<\/strong><\/p>\n\n\n\n

\u57fa\u4e8e\u4e0a\u8ff0\u601d\u60f3\u6765\u6784\u5efaSLM\uff0c\u9700\u8981\u89e3\u51b3\u4ee5\u4e0b\u51e0\u4e2a\u95ee\u9898\uff1a<\/p>\n\n\n\n

  1. \u8bed\u97f3\u5982\u4f55\u79bb\u6563\u5316\uff1f<\/strong><\/li>
  2. \u5982\u4f55\u8ba9LLM\u7406\u89e3\u8bed\u97f3\u7684token\uff1f\u52a0\u5165\u8bed\u97f3token\u4e4b\u540e\uff0cLLM\u5728\u8bed\u97f3\u6570\u636e\u7684\u7406\u89e3\u4e0a\u662f\u5426\u5177\u6709\u6d8c\u73b0\u6027\uff1f<\/strong><\/li>
  3. LLM\u5982\u4f55\u5408\u6210\/\u89e3\u7801\u8bed\u97f3\uff1f<\/strong><\/li><\/ol>\n\n\n\n

    \u63a5\u4e0b\u6765\uff0c\u6211\u4eec\u6309\u56fe\u7d22\u9aa5\uff0c\u5206\u522b\u770b\u770b\u4e0a\u8ff0\u4e09\u4e2a\u95ee\u9898\u5e94\u8be5\u5982\u4f55\u89e3\u51b3\u3002\u770b\u5b8c\u73b0\u6709\u7684\u65b9\u6848\u4e4b\u540e\uff0c\u4e5f\u4f1a\u8c08\u8c08\u4e00\u4e9b\u5173\u4e8e\u5de5\u7a0b\u5b9e\u73b0\u7684\u601d\u8003\u4ee5\u53ca\u65b0\u5174\u8bed\u97f3\u6280\u672f\u5bf9\u4e8e\u6e38\u620f\u4e1a\u52a1\u7684\u5f71\u54cd\u3002\u6700\u540e\uff0c\u6211\u4f1a\u7ed9\u51fa\u4e00\u4e2a\u5b8c\u6574\u7684roadmap\u6765\u6536\u675f\u5168\u6587\u3002<\/p>\n\n\n\n

    \u8bed\u97f3\u7684\u79bb\u6563\u5316\uff1a\u5411LLM\u770b\u9f50\uff01<\/strong><\/h2>\n\n\n\n

    \u5728\u8c08\u53ca\u8bed\u97f3\u79bb\u6563\u5316\u4e4b\u524d\uff0c\u6211\u4eec\u5148\u6765\u770b\u770b\u8bed\u97f3\u548c\u6587\u672c\u4f5c\u4e3a\u4e24\u79cd\u4e0d\u540c\u7684\u6a21\u6001\uff0c\u6709\u4ec0\u4e48\u533a\u522b\uff0c\u6709\u4ec0\u4e48\u8054\u7cfb\u3002\u8fd9\u76f4\u63a5\u5173\u7cfb\u5230\u540e\u6587\u5efa\u6a21\u65b9\u6cd5\u7684\u9009\u62e9\u4ee5\u53ca\u79bb\u6563\u5316\u7279\u5f81\u7684\u5173\u6ce8\u70b9\u3002<\/p>\n\n\n\n

    \u8bed\u97f3\u548c\u6587\u672c\u7684\u5dee\u522b\u4e3b\u8981\u4f53\u73b0\u5728\uff1a\u6587\u672c\u79bb\u6563\u3001\u5e8f\u5217\u77ed\u3001\u4fe1\u606f\u5bc6\u5ea6\u9ad8(\u51e0\u4e4e\u6bcf\u4e2a\u8bcd\u90fd\u5305\u542b\u8bed\u4e49)\uff1b\u8bed\u97f3\u5219\u8fde\u7eed\u3001\u5e8f\u5217\u957f\u3001\u4fe1\u606f\u5bc6\u5ea6\u4f4e\u3002<\/strong>\u8bed\u97f3\u5e8f\u5217\u957f\u3001\u4fe1\u606f\u5bc6\u5ea6\u4f4e\u7684\u7279\u70b9\uff0c\u610f\u5473\u7740\u8bed\u97f3\u6570\u636e\u6709\u5f88\u5927\u7684\u538b\u7f29\u7a7a\u95f4\uff0c\u8fd9\u4e00\u70b9\u548c\u56fe\u50cf\u975e\u5e38\u7c7b\u4f3c\u3002<\/strong>\u56e0\u6b64\uff0c\u4e00\u4e9b\u7528\u4e8e\u56fe\u50cf\u7684\u79bb\u6563\u5316\u538b\u7f29\u65b9\u6cd5\u4e5f\u53ef\u4ee5\u7528\u5728\u8bed\u97f3\u4e0a\u3002<\/p>\n\n\n\n

    \u9664\u4e86\u5dee\u5f02\uff0c\u8bed\u97f3\u548c\u6587\u672c\u4e5f\u6709\u4e00\u5b9a\u7684\u8054\u7cfb\uff1a\u8bed\u97f3\u662f\u6587\u672c\u7684\u8d85\u96c6\uff0c\u65e2\u5305\u542b\u6587\u672c\u5185\u5bb9(\u8bf4\u8bdd\u4eba\u8bf4\u4e86\u4ec0\u4e48\uff0c\u4e5f\u5c31\u662f\u8bed\u4e49\u4fe1\u606f)\uff0c\u4e5f\u5305\u542b\u8bed\u97f3\u7279\u6709\u7684\u97f3\u8272\u3001\u97f5\u5f8b\u3001\u8bed\u901f\u7b49\u58f0\u5b66\u4fe1\u606f(\u4e5f\u53eb\u505a\u526f\u8bed\u8a00)\u3002<\/strong>\u65e2\u7136\u8bed\u97f3\u5305\u542b\u6587\u672c\uff0c\u90a3\u4e48\u5728NLP\u4e2d\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u4e5f\u53ef\u4ee5\u7528\u6765\u5efa\u6a21\u8bed\u97f3\u4e2d\u7684\u4e0a\u4e0b\u6587\u4f9d\u8d56\u5173\u7cfb\uff0c\u4ece\u800c\u5f97\u5230\u8bed\u97f3\u7684\u79bb\u6563\u5316token\u3002\u57fa\u4e8e\u8fd9\u4e9b\u65b9\u6cd5\u5f97\u5230\u7684token\u4e3b\u8981\u5305\u542b\u8bed\u97f3\u7684\u8bed\u4e49\u4fe1\u606f\u3002<\/p>\n\n\n\n

    \u82b1\u5f00\u4e24\u6735\uff0c\u5404\u8868\u4e00\u679d\u3002\u6211\u4eec\u5148\u6765\u770b\u770b\u8bed\u97f3\u7684\u8bed\u4e49token\u5982\u4f55\u83b7\u53d6\u3002<\/p>\n\n\n\n

    \u8bed\u4e49token:  \u7528MLM\u5efa\u6a21\u8bed\u97f3\u7684\u4e0a\u4e0b\u6587\u4f9d\u8d56<\/strong><\/h3>\n\n\n\n

    \u8bed\u97f3\u7684\u8bed\u4e49\u5efa\u6a21\u65b9\u6cd5\uff0c\u6700\u5e38\u7528\u5230\u7684\u5c31\u662fBERT\u7684MLM\u65b9\u6cd5\uff0c\u6bd4\u8f83\u7ecf\u5178\u7684\u5de5\u4f5c\u6709\u4e09\u4e2a\uff1awav2vec 2.0[1]\u3001HuBERT[2]\u548cw2v-BERT[3]\u3002<\/strong><\/p>\n\n\n\n

    \u7c7b\u4f3c\u4e8eBERT\uff0cwav2vec 2.0[1]\u5728\u9690\u7a7a\u95f4(latent space)\u968f\u673amask\u4e86\u4e00\u5b9a\u6bd4\u4f8b\u7684\u8bed\u97f3\u8f93\u5165\uff0c\u7136\u540e\u7528\u57fa\u4e8e\u5bf9\u6bd4\u5b66\u4e60\u7684\u8bad\u7ec3\u76ee\u6807\u5b66\u4e60\u5e27\u7684\u8868\u5f81\u3002\u503c\u5f97\u6ce8\u610f\u7684\u4e00\u70b9\u662f\uff0c\u5bf9\u6bd4\u5b66\u4e60\u4e2d\u76ee\u6807\u5e27\u7684\u79bb\u6563\u5316\u5904\u7406\u662f\u4e00\u4e2a\u975e\u5e38\u5de7\u5999\u7684\u64cd\u4f5c\uff0c\u5b83\u5c06\u65e0\u9650\u7684\u8fde\u7eed\u7279\u5f81\u7a7a\u95f4\u574d\u7f29\u4e3a\u6709\u9650\u7684\u79bb\u6563\u7a7a\u95f4\uff0c\u8ba9\u5e27\u7279\u5f81\u7684\u9c81\u68d2\u6027\u66f4\u5f3a\u4e86\u3002\u8fd9\u5728\u8bed\u97f3\u9886\u57df\u4e0a\u975e\u5e38\u6709\u7528\u7684trick\uff0c\u5141\u8bb8\u6a21\u578b\u63a5\u53d7\u5e26\u6709\u566a\u58f0\u7684\u8bed\u97f3\u4f5c\u4e3a\u8f93\u5165\u3002<\/p>\n\n\n\n

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

    \u56fe1\uff1awav2vec 2.0\u7684\u6a21\u578b\u67b6\u6784<\/strong><\/p>\n\n\n\n

    wav2vec 2.0\u53ea\u662f\u501f\u7528\u4e86BERT\u4e2dmask\u7684\u64cd\u4f5c\uff0c\u8bad\u7ec3\u76ee\u6807\u5927\u4f53\u4e0a\u662f\u57fa\u4e8e\u5bf9\u6bd4\u5b66\u4e60\u7684\u8303\u5f0f\u3002\u90a3\u4e48\uff0c\u80fd\u76f4\u63a5\u7528BERT\u7684MLM\u5efa\u6a21\u76ee\u6807\u6765\u5f97\u5230\u9ad8\u8d28\u91cf\u7684\u8bed\u97f3\u8868\u5f81\u5417\uff1f\u5176\u540e\u7684HuBERT[2]\u505a\u7684\u5c31\u662f\u8fd9\u4e2a\u4e8b\u60c5\u3002HuBERT[2]\u7684\u6838\u5fc3\u70b9\u5728\u4e8e\u4f7f\u7528\u7b80\u5355\u7684KMeans\u805a\u7c7b\u65b9\u6cd5\u4e3a\u8bed\u97f3\u6570\u636e\u62bd\u53d6\u79bb\u6563\u5316\u7684\u5206\u7c7b\u6807\u7b7e\uff0c\u4e5f\u5c31\u662f\u6587\u4e2d\u6240\u8bf4\u7684hidden unit\/acoustic unit\u3002\u6709\u4e86\u5206\u7c7b\u6807\u7b7e\uff0c\u7136\u540e\u5c31\u662f\u7528BERT\u7684MLM loss\u6765\u5b66\u4e60\u8bed\u97f3\u6570\u636e\u4e2d\u5185\u5728\u7684\u4e0a\u4e0b\u6587\u4f9d\u8d56\u5173\u7cfb\u3002\u5bf9\u4e8eKMeans\u805a\u7c7b\u5bf9\u521d\u59cb\u503c\u548cK\u503c\u9ad8\u7075\u654f\u7684\u7279\u70b9\uff0c\u4f5c\u8005\u8bbe\u8ba1\u4e86ensemble\u548citerative refinement\u65b9\u6cd5\u4e88\u4ee5\u89e3\u51b3\u3002\u524d\u8005\u5c31\u662f\u591a\u4e2a\u805a\u7c7b\u6a21\u578bensemble\uff0c\u540e\u8005\u5c31\u662f\u5148\u5728\u57fa\u4e8eMFCC\u7684\u805a\u7c7b\u6807\u7b7e\u4e0a\u8fdb\u884c\u5b66\u4e60\uff0c\u5b66\u4e60\u5230\u4e00\u5b9a\u7a0b\u5ea6\u65f6\uff0c\u5728\u6a21\u578b\u5b66\u4e60\u5230\u7684\u8868\u5f81\u91cd\u65b0\u805a\u7c7b\uff0c\u518d\u505a\u4e00\u6b21BERT\u7684\u5b66\u4e60\u3002<\/p>\n\n\n\n

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

    \u56fe2\uff1aHuBERT\u7684\u6a21\u578b\u67b6\u6784<\/strong><\/p>\n\n\n\n

    \u65e2\u7136\u5bf9\u6bd4\u5b66\u4e60\u53ef\u4ee5\u5b66\u4e60\u8bed\u97f3\u7684\u8bed\u4e49\u8868\u5f81\uff0cBERT\u7684MLM\u4e5f\u53ef\u4ee5\uff0c\u90a3\u5c06\u4e8c\u8005\u7ed3\u5408\u8d77\u6765\uff0c\u4f1a\u4e0d\u4f1a\u6709\u4e92\u8865\u7684\u6548\u679c\u5462\uff1fw2v-BERT[3]\u505a\u7684\u5c31\u662f\u8fd9\u4e2a\u4e8b\u60c5\u3002\u6ce8\u610f\u5230\uff1aHuBERT\u4e2d\u8bed\u97f3\u7684\u79bb\u6563token\u4e0d\u662f\u7aef\u5230\u7aef\u83b7\u5f97\u7684\uff0c\u9700\u8981\u7528KMeans\u7b97\u6cd5\u5bf9\u7279\u5f81\u8fdb\u884c\u79bb\u7ebf\u805a\u7c7b\uff0c\u800cwav2vec 2.0\u53c8\u6b63\u597d\u63d0\u4f9b\u4e86\u97f3\u9891\u5e27\u7684\u91cf\u5316\u79bb\u6563\u8868\u5f81\uff0cHuBERT\u548cwav2vec 2.0\u5f88\u5bb9\u6613\u5c31\u80fd\u7f1d\u5408\u5728\u4e00\u8d77\u3002\u7f1d\u5408\u7684\u65b9\u6cd5\u4e5f\u662f\u663e\u7136\u7684\uff1a\u524d\u9762\u82e5\u5e72\u5c42\u505a\u7c7b\u4f3cwav2vec 2.0\u7684\u5bf9\u6bd4\u5b66\u4e60\uff0c\u5b66\u4e60\u51faHuBERT\u8981\u7528\u7684\u79bb\u6563\u8868\u5f81\uff0c\u7136\u540e\u5728\u540e\u9762\u82e5\u5e72\u5c42\u505a\u7c7b\u4f3cHuBERT\u7684MLM\u8bad\u7ec3\u3002<\/p>\n\n\n\n

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

    \u56fe3\uff1aw2v-BERT\u7684\u6a21\u578b\u67b6\u6784<\/strong><\/p>\n\n\n\n

    \u58f0\u5b66token\uff1a\u538b\u7f29+\u79bb\u6563<\/strong><\/h3>\n\n\n\n

    \u4e0a\u4e00\u90e8\u5206\u4ecb\u7ecd\u7684\u9884\u8bad\u7ec3\u6a21\u578b\u505a\u7684\u662f\u4e0a\u4e0b\u6587\u5173\u7cfb\u7684\u9884\u8bad\u7ec3\uff0c\u5b66\u4e60\u5230\u7684\u8868\u5f81\u4e3b\u8981\u5305\u542b\u4e0e\u4e0a\u4e0b\u6587\u76f8\u5173\u7684\u8bed\u4e49\u4fe1\u606f\u3002\u8981\u60f3\u5c06\u8bed\u97f3\u7684token\u8fd8\u539f\u6210\u4e3a\u771f\u6b63\u5177\u6709\u771f\u4eba\u8868\u73b0\u529b\u7684\u4fe1\u53f7\uff0c\u8fd8\u9700\u8981\u6709\u5305\u542b\u97f3\u8272\u3001\u97f5\u5f8b\u3001\u8bed\u901f<\/strong>\u7b49\u526f\u8bed\u8a00\u4fe1\u606f\u7684\u58f0\u5b66\u7279\u5f81\u3002\u58f0\u5b66\u7279\u5f81\u7684\u5b66\u4e60\u5728\u5f88\u5927\u7a0b\u5ea6\u4e0a\u53c2\u8003\u4e86\u56fe\u50cf\u9886\u57df\u7684\u5de5\u4f5c\uff0c\u7528\u5230\u7684\u4e3b\u8981\u662f\u7c7b\u4f3c\u4e8eVQVAE[4]\u3001VQGAN\u7b49\u7684\u79bb\u6563\u5316\u538b\u7f29\u65b9\u6cd5\uff0c\u5e76\u9488\u5bf9\u8bed\u97f3\u6570\u636e\u7684\u7279\u6027\u505a\u4e86\u4f18\u5316\u3002\u8fd9\u4e00\u90e8\u5206\u6bd4\u8f83\u7ecf\u5178\u7684\u5de5\u4f5c\u5c31\u662fSoundStream[5]\u548cEncodec[6]\uff0c\u4e8c\u8005\u7684\u5de5\u4f5c\u9ad8\u5ea6\u7c7b\u4f3c\uff0c\u6211\u4eec\u653e\u5728\u4e00\u8d77\u6765\u770b\u3002<\/p>\n\n\n\n

    \u8bf4\u5230\u538b\u7f29\uff0c\u6700\u5148\u60f3\u5230\u7684\u6a21\u578b\u5f53\u7136\u5c31\u662fAutoEncoder(\u81ea\u7f16\u7801\u5668)\u3002\u4e3a\u63d0\u5347\u538b\u7f29\u6548\u7387\uff0c\u6709\u5229\u4e8e\u6570\u5b57\u4f20\u8f93\u548c\u5b58\u50a8\uff0c\u4ee5\u53ca\u79bb\u6563\u5316\u5efa\u6a21\u7684\u8981\u6c42\uff0c\u538b\u7f29\u6a21\u578b\u4e2d\u8fd8\u9700\u8981\u5305\u542b\u91cf\u5316(quantization)\uff0c\u5c06\u8fde\u7eed\u7684\u97f3\u9891\u4fe1\u53f7\u8f6c\u6362\u4e3a\u79bb\u6563\u7684\u6570\u503c\u3002\u57fa\u4e8e\u4e0a\u8ff0\u8003\u8651\uff0c\u6a21\u578b\u5927\u4f53\u4e0a\u5e94\u8be5\u662fVQVAE[4]\u7684\u7ed3\u6784\u3002\u4e3a\u4e86\u5e73\u8861VQ(Vector Quantization\uff0c\u5411\u91cf\u91cf\u5316)\u4e0e\u97f3\u9891\u5b9e\u65f6\u9ad8\u4fdd\u771f\u4f20\u8f93\u7684\u77db\u76fe\uff0c\u901a\u5e38\u91c7\u7528\u591a\u4e2a\u6b8b\u5dee\u8fde\u63a5\u7684codebook\u6765\u8fdb\u884c\u91cf\u5316\uff0c\u8fd9\u4e2a\u5c31\u662f\u6240\u8c13\u7684RVQ(\u5177\u4f53\u5206\u6790\u8fc7\u7a0b\u53ef\u4ee5\u53c2\u89c1\u77e5\u4e4e\u6587\u7ae0)\u3002\u91c7\u7528RVQ\u7684\u597d\u5904\u4e3b\u8981\u6709\u4e24\u4e2a\uff1a\u5176\u4e00\uff0c\u533a\u5206\u4e0d\u540cquantization block\u7684\u5206\u5de5\uff0c\u7b2c\u4e00\u4e2ablock\u5305\u542b\u6700\u91cd\u8981\u7684\u8bed\u4e49\u4fe1\u606f\uff0c\u540e\u7eed\u7684block\u5305\u542b\u8fd8\u539f\u8bed\u97f3\u7684\u526f\u8bed\u8a00\u4fe1\u606f\uff1b\u7b2c\u4e8c\uff0c\u6a21\u578b\u8bad\u7ec3\u65f6\u53ef\u968f\u673a\u91c7\u6837\u524d\u9762\u82e5\u5e72\u4e2ablock\u6765\u8bad\u7ec3\uff0c\u4fdd\u6301\u4e00\u5b9a\u7cbe\u5ea6\uff0c\u5b9e\u73b0\u5bf9\u6bd4\u7279\u7387\u7684\u52a8\u6001\u9002\u5e94\u3002<\/p>\n\n\n\n

    \u603b\u800c\u8a00\u4e4b\uff0cSoundStream<\/strong>[5]\/Encodec<\/strong>[6]\u5176\u5b9e\u5c31\u662f\u4e00\u4e2aRVQ-VAE\uff0c\u5b83\u4eec\u6240\u5efa\u6a21\u7684\u8bed\u97f3\u79bb\u6563\u5316token\u5305\u542b\u4e86\u5c42\u6b21\u5316\u7684\u8bed\u4e49\u4fe1\u606f\u548c\u58f0\u5b66\u4fe1\u606f\u3002<\/p>\n\n\n\n

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

    \u56fe4\uff1aEncodec\u7684\u6a21\u578b\u67b6\u6784<\/strong><\/p>\n\n\n\n

    \u8bed\u97f3\u7684\u7edf\u4e00\u8868\u5f81\uff1f<\/strong><\/h3>\n\n\n\n

    \u4e0d\u96be\u53d1\u73b0\uff0c\u867d\u7136\u8bf4SoundStream[5]\u548cEncodec[6]\u8fd9\u6837\u7684\u57fa\u4e8eRVQ-VAE\u7684\u538b\u7f29\u5efa\u6a21\u65b9\u6cd5\u5305\u542b\u4e86\u8bed\u97f3\u7684\u58f0\u5b66\u7279\u5f81\uff0c\u4f46\u5176\u4e2d\u4e5f\u4e0d\u53ef\u907f\u514d\u5730\u5e26\u5165\u4e86\u8bed\u4e49\u7279\u5f81\u3002\u4e8c\u8005\u63d0\u53d6\u7684\u5b9e\u9645\u4e0a\u66f4\u50cf\u662f\u4e00\u79cd\u8bed\u4e49\u7279\u5f81\u548c\u58f0\u5b66\u7279\u5f81\u7684\u6df7\u5408\u4f53\u3002<\/strong>\u57fa\u4e8e\u6b64\uff0cSpeechTokenizer<\/strong>[7]\u5728\u4e8c\u8005\u7684\u57fa\u7840\u4e0a\uff0c\u5f15\u5165\u4e86\u8bed\u4e49\u5f15\u5bfc\u4fe1\u606f\u6765\u89e3\u8026\u8bed\u4e49\u7279\u5f81\u548c\u58f0\u5b66\u7279\u5f81\u3002\u8bed\u4e49\u7279\u5f81\u548c\u58f0\u5b66\u7279\u5f81\u7684\u89e3\u8026\u5bf9\u4e8e\u6700\u7ec8\u7684\u8bed\u97f3\u5408\u6210\u6709\u7740\u76f8\u5f53\u7684\u91cd\u8981\u6027\u3002SpeechTokenizer\u7684\u5177\u4f53\u505a\u6cd5\u662f\uff1a\u4f7f\u7528HuBERT[2]\u7684\u7279\u5f81\u5bf9RVQ1\u7684\u7279\u5f81\u505a\u8bed\u4e49\u84b8\u998f\uff0c\u5176\u4f59\u90e8\u5206\u4fdd\u7559\u58f0\u5b66\u4fe1\u606f\u3002<\/p>\n\n\n\n

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

    \u56fe5\uff1aSpeechTokenizer\u7684\u6a21\u578b\u67b6\u6784<\/strong><\/p>\n\n\n\n


    \u8bed\u97f3\u7684\u5176\u4ed6\u8868\u5f81\uff1aMEL\u4f9d\u65e7\u6709\u7528\uff01<\/strong><\/h3>\n\n\n\n

    \u4e0a\u8ff0\u7684\u8bed\u97f3\u79bb\u6563\u8868\u5f81\uff0c\u4e0d\u7ba1\u662f\u57fa\u4e8eHuBERT[2]\u7684\u8bed\u4e49token\uff0c\u8fd8\u662f\u57fa\u4e8eEncodec[6]\u7684\u58f0\u5b66token\uff0c\u5b83\u4eec\u90fd\u662f\u76f4\u63a5\u57fa\u4e8e\u539f\u59cb\u7684\u97f3\u9891\u6ce2\u5f62\u62bd\u53d6\u7684\u3002\u9664\u6b64\u4e4b\u5916\uff0c\u4e5f\u53ef\u4ee5\u57fa\u4e8e\u8bed\u97f3\u7684\u4e2d\u95f4\u8868\u5f81\u6765\u62bd\u53d6\u3002\u6700\u5178\u578b\u7684\u8bed\u97f3\u4e2d\u95f4\u8868\u5f81\u5c31\u662f\u6885\u5c14\u8c31(MEL spectrogram\uff0c\u4e0b\u6587\u7b80\u79f0MEL)\u3002\u6885\u5c14\u8c31\u672c\u8eab\u5c31\u5bf9\u8bed\u97f3\u8fdb\u884c\u4e86\u538b\u7f29\uff0c\u5c06\u6885\u5c14\u8c31\u7c7b\u6bd4\u4e8e\u56fe\u50cf\uff0c\u4f7f\u7528\u5355\u7801\u672c\u7684VQ\u4e5f\u53ef\u4ee5\u8fbe\u5230\u4e0eSoundStream\u548cEncodec\u90a3\u6837\u7c7b\u4f3c\u7684\u538b\u7f29\u7a0b\u5ea6\u3002\u8fd9\u79cdMEL+VQ\u7684\u505a\u6cd5\u5728\u5404\u79cd\u8bed\u97f3\u5408\u6210\u6a21\u578b\u4e2d\u4e5f\u76f8\u5f53\u5e38\u89c1\u3002\u6211\u4eec\u5728\u8bed\u97f3\u5408\u6210\u90e8\u5206\u4f1a\u8be6\u7ec6\u4ecb\u7ecd\u3002<\/p>\n\n\n\n

    \u8ba9LLM\u7406\u89e3\u8bed\u97f3token\uff01<\/strong><\/h2>\n\n\n\n

    \u6709\u4e86\u4e0a\u9762\u6240\u8bf4\u7684\u8bed\u4e49token\u548c\u58f0\u5b66token\u4e4b\u540e\uff0c\u5176\u5b9e\u5c31\u53ef\u4ee5\u5229\u7528\u5b83\u4eec\u6765\u6784\u5efa\u8bed\u97f3\u5c42\u9762\u7684\u8bed\u8a00\u6a21\u578b\u4e86\u3002\u6bd4\u8f83\u7ecf\u5178\u7684\u5de5\u4f5c\u6709\uff1a\u8c37\u6b4c\u7684AudioLM[8]\u548cAudioPaLM[9]\u3001\u5b57\u8282\u7684SALMONN[10]\u3001\u590d\u65e6\u7684SpeechGPT[11]\/SpeechGPT-Gen[12]\/SpeechAlign[13]\u3001\u963f\u91cc\u7684LauraGPT[14]\u548c\u65b0\u52a0\u5761\u56fd\u7acb\u5927\u5b66\u7684NextGPT[15]\u3002\u5b83\u4eec\u7684\u505a\u6cd5\u5176\u5b9e\u90fd\u5927\u5dee\u4e0d\u5dee\uff0c\u6211\u4eec\u770b\u51e0\u4e2a\u5c31\u77e5\u9053\u662f\u600e\u4e48\u56de\u4e8b\u4e86\u3002<\/p>\n\n\n\n

    AudioLM\uff1a\u6700\u521d\u7684SLM<\/strong><\/h3>\n\n\n\n

    \u89c1\u540d\u77e5\u4e49\uff0cAudioLM[8]\u6784\u5efa\u7684\u662f\u8bed\u97f3\u5c42\u9762\u7684\u8bed\u8a00\u6a21\u578b\u2014\u2014\u7ed9\u5b9a\u4e00\u6bb5\u8bed\u97f3\uff0c\u6a21\u578b\u9884\u6d4b\u540e\u7eed\u7684\u8bed\u97f3\u3002\u8f93\u5165\u4fa7\u548c\u8f93\u51fa\u4fa7\u90fd\u53ea\u6709\u8bed\u97f3\u6a21\u6001\u3002\u8fd9\u4e2a\u4efb\u52a1\u5f62\u5f0f\u548cGPT-4o\u975e\u5e38\u7c7b\u4f3c\uff0c\u4e0d\u4f1a\u7ecf\u5386ASR->LM->TTS\u7684\u8fc7\u7a0b\uff0c\u800c\u662f\u76f4\u63a5\u4ece\u8bed\u97f3\u4e0a\u4e0b\u6587\u4e2d\u63a8\u7406\u8bed\u4e49\u4fe1\u606f\uff0c\u518d\u7ed3\u5408\u58f0\u5b66\u4fe1\u606f\u5408\u6210\u8d34\u5408\u4e0a\u4e0b\u6587\u7684\u9ad8\u8868\u73b0\u529b\u8bed\u97f3\u3002\u800c\u4e0a\u6587\u6240\u8ff0\u7684\u8bed\u4e49token\u548c\u58f0\u5b66token\u6b63\u597d\u5c31\u80fd\u6ee1\u8db3\u8fd9\u4e2a\u4efb\u52a1\u7684\u8981\u6c42\u3002<\/p>\n\n\n\n

    AudioLM\u7684\u5177\u4f53\u505a\u6cd5\u662f\uff1a\u7528SoundStream[5]\u63d0\u53d6\u58f0\u5b66token\uff0c\u7528w2v-BERT[3]\u63d0\u53d6\u8bed\u4e49token\uff0c\u6a21\u578b\u4e3b\u4f53\u5c31\u662f\u4e00\u4e2a\u5e38\u89c4\u7684GPT\uff0c\u8bcd\u8868\u5305\u542b\u6240\u6709\u7684\u58f0\u5b66token\u548c\u8bed\u4e49token\u3002\u5b83\u7684\u5efa\u6a21\u8fc7\u7a0b\u4e5f\u76f8\u5f53\u6709\u610f\u601d\uff0c\u6709\u5f88\u5927\u7684\u53c2\u8003\u610f\u4e49\uff1a\u5148\u505a\u6700\u91cd\u8981\u7684\u8bed\u4e49\u5efa\u6a21\uff0c\u7136\u540e\u5148\u9884\u6d4bSoundStream\u7684\u524d\u82e5\u5e72\u5c42\u7279\u5f81\uff0c\u5efa\u6a21\u7c97\u7cd9\u7684\u58f0\u5b66\u7279\u5f81\uff0c\u5728\u9884\u6d4bSoundStream\u7684\u5269\u4f59\u5c42\u7279\u5f81\uff0c\u5efa\u6a21\u58f0\u97f3\u7684\u7ec6\u8282\u4fe1\u606f\uff0c\u6700\u540e\u57fa\u4e8e\u6240\u6709\u7684\u58f0\u5b66token\u8fd8\u539f\u4e3a\u8bed\u97f3\u3002\u8fd9\u79cd\u5c42\u6b21\u5316\u7684\u5efa\u6a21\u5728\u8bf8\u5982VALL-E[16]\u8fd9\u6837\u7684\u8bed\u97f3\u5408\u6210\u6a21\u578b\u4e2d\u4e5f\u975e\u5e38\u5e38\u89c1\u3002<\/p>\n\n\n\n

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

    \u56fe6\uff1aAudioLM\u7684tokenizer<\/strong><\/p>\n\n\n\n

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

    \u56fe7\uff1aAudioLM\u7684\u5efa\u6a21\u6d41\u7a0b<\/strong><\/p>\n\n\n\n

    \u5f53\u7136\uff0cAudioLM[8]\u4ec5\u4ec5\u5173\u6ce8\u8bed\u97f3\u6a21\u6001\uff0cLM\u4e5f\u5f88\u5e38\u89c4\uff0c\u4e0d\u5177\u5907\u5982\u540cGPT-4o\u4e00\u6837\u5f3a\u608d\u7684\u6307\u4ee4\u9075\u5faa\u80fd\u529b\u548c\u5bf9\u8bdd\u80fd\u529b\uff0c\u8bed\u97f3\u5bf9\u8bdd\u7684\u8fde\u8d2f\u6027\u548c\u8868\u73b0\u529b\u90fd\u76f8\u5f53\u5f31\u3002\u4f46\u8fd9\u4e00\u5de5\u4f5c\u4ecd\u7136\u5177\u6709\u76f8\u5f53\u7684\u542f\u53d1\u6027\u548c\u5f00\u62d3\u6027\uff0c\u8bc1\u660e\u4e86\uff1a\u5373\u4f7f\u662f\u5e38\u89c4\u7684LM\uff0c\u7167\u6837\u4e5f\u80fd\u7406\u89e3\u8bed\u97f3token\u3002<\/p>\n\n\n\n

    AudioPaLM[9]\uff1a\u6574\u5408LLM<\/strong><\/h3>\n\n\n\n

    \u8fd9\u4e2a\u5c31\u662fAudioLM\u7684\u540e\u7eed\u4e86\uff0c\u8c37\u6b4c\u5c06\u5e38\u89c4\u7684LM\u66ff\u6362\u6210\u5df2\u7ecf\u8bad\u7ec3\u597d\u7684\u3001\u5177\u6709\u5f3a\u5927\u6587\u672c\u7406\u89e3\u80fd\u529b\u548c\u751f\u6210\u80fd\u529b\u7684\u5927\u8bed\u8a00\u6a21\u578b\u2014\u2014PaLM-2[17]\uff0c\u65e2\u7ee7\u627f\u4e86AudioLM\u4fdd\u7559\u526f\u8bed\u8a00\u7684\u80fd\u529b\uff0c\u53c8\u878d\u5408\u4e86PaLM-2\u5f3a\u5927\u7684\u8bed\u4e49\u7406\u89e3\u80fd\u529b\u548c\u63a8\u7406\u80fd\u529b\u3002\u800c\u4e14\uff0c\u8be5\u6a21\u578b\u7684\u8bcd\u8868\u540c\u65f6\u5305\u542b\u5927\u8bed\u8a00\u6a21\u578b\u7684token\u548c\u8bed\u97f3token\uff0c\u53ef\u4ee5\u540c\u65f6\u505a\u8bed\u97f3\u7406\u89e3\u4efb\u52a1\u548c\u5408\u6210\u751f\u6210\u4efb\u52a1\uff0c\u7b2c\u4e00\u5c06\u8fd9\u4e9b\u4efb\u52a1\u6574\u5408\u5728\u4e00\u4e2a\u6a21\u578b\u4e2d\u8fdb\u884c\u89e3\u51b3\u3002<\/p>\n\n\n\n

    \u4e0d\u8fc7\uff0c\u9700\u8981\u6307\u51fa\u5730\u662f\uff0c\u6587\u4e2d\u7684\u8bed\u97f3token embedding\u662f\u76f4\u63a5\u8f93\u5165\u5230Transformer\u4e2d\u7684\uff0c\u5e76\u6ca1\u6709\u4f7f\u7528\u97f3\u9891\u7f16\u7801\u5668\u505a\u4e00\u6b21\u8f6c\u6362\u3002\u800c\u4e14\uff0cAudioPaLM\u7684\u8bad\u7ec3\u66f4\u52a0\u63a5\u8fd1\u6587\u672c\u591a\u4efb\u52a1\u7684T5\uff0c\u5e76\u672a\u7528\u5230\u590d\u6742\u7684\u3001\u4e30\u5bcc\u591a\u6837\u7684\u6307\u4ee4\u6765\u8868\u8fbe\u4efb\u52a1\u7684\u610f\u56fe\uff0c\u8fd8\u4e0d\u80fd\u7b97\u662f\u771f\u6b63\u4e25\u683c\u7684instruction fine-tuning\u3002<\/p>\n\n\n\n

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

    \u56fe8\uff1aAudioPaLM\u7684\u6a21\u578b\u67b6\u6784<\/strong><\/p>\n\n\n\n

    SALMONN[10]\uff1a\u8ba9LLM\u7406\u89e3\u8bed\u97f3<\/strong><\/h3>\n\n\n\n

    \u8fd9\u662f\u5b57\u8282\u8df3\u52a8\u548c\u6e05\u534e\u5927\u5b66\u7535\u5b50\u7cfb(\u4e5f\u662f\u6211\u4eec\u5b9e\u9a8c\u5ba4)\u7684\u5408\u4f5c\u6210\u679c\u3002\u867d\u7136\u8fd9\u4e2a\u5de5\u4f5c\u7684\u76ee\u7684\u662f\u8ba9LLM\u80fd\u591f\u7406\u89e3\u8bed\u97f3\uff0c\u8fd8\u4e0d\u80fd\u751f\u6210\u8bed\u97f3\uff0c\u4f46\u5b83\u7684\u8bad\u7ec3\u65b9\u6cd5\u548cLLM\u6bd4\u8f83\u63a5\u8fd1\uff0c\u800c\u4e14\u5728\u8bf8\u591a\u8bed\u97f3\u76f8\u5173\u7684\u4efb\u52a1\u4e0a\u90fd\u663e\u793a\u51fa\u4e86\u6d8c\u73b0\u6027\uff0c\u53ef\u4ee5\u7528\u4f5cuniversal\u7684\u7279\u5f81\u63d0\u53d6\u5668\uff0c\u8fd9\u5bf9\u4e8e\u6784\u5efa\u9ad8\u8d28\u91cf\u7684\u3001\u5305\u542b\u8bed\u97f3-\u6587\u672c\u591a\u6a21\u6001\u7684\u6307\u4ee4\u5fae\u8c03\u6570\u636e\u96c6\u5177\u6709\u76f8\u5f53\u5927\u7684\u610f\u4e49\u3002<\/p>\n\n\n\n

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

    \u56fe9\uff1aSALMONN\u7684\u6a21\u578b\u67b6\u6784<\/strong><\/p>\n\n\n\n

    SpeechGPT\/SpeechGPT-Gen\/SpeechAlign\uff1a\u5411LLM\u7684\u8bad\u7ec3\u65b9\u6cd5\u770b\u9f50<\/strong><\/h3>\n\n\n\n

    \u8fd9\u7b97\u662f\u590d\u65e6\u5927\u5b66\u90b1\u9521\u9e4f\u7ec4\u5728\u8fd9\u4e2a\u9886\u57df\u4e00\u4e2a\u6210\u7cfb\u5217\u7684\u5de5\u4f5c\uff0c\u6211\u4eec\u4e00\u4e2a\u4e00\u4e2a\u6765\u770b\u3002<\/p>\n\n\n\n

    SpeechGPT[11]\u505a\u7684\u4e5f\u662f\u517c\u5177\u8bed\u97f3\u7406\u89e3\u80fd\u529b\u548c\u8bed\u97f3\u751f\u6210\u80fd\u529b\u7684\u591a\u6a21\u6001\u6a21\u578b\u3002\u5728\u6a21\u578b\u7684\u8bad\u7ec3\u4e0a\uff0cSpeechGPT\u5927\u5e45\u5ea6\u5411LLM\u770b\u9f50\uff0c\u4f7f\u7528\u4e86\u4e09\u6bb5\u5f0f\u7684\u8bad\u7ec3\u65b9\u6cd5\uff1a\u7b2c\u4e00\u9636\u6bb5\u5148\u505a\u6a21\u6001\u9002\u5e94\u7684\u9884\u8bad\u7ec3\uff0c\u5176\u5b9e\u5c31\u662f\u62ffASR\u7684\u8bed\u97f3\u6570\u636e\u6765\u505a\u9884\u8bad\u7ec3\uff1b\u7b2c\u4e8c\u9636\u6bb5\u548c\u7b2c\u4e09\u9636\u6bb5\u90fd\u662f\u6307\u4ee4\u5fae\u8c03\uff0c\u4e0d\u8fc7\u6839\u636e\u6307\u4ee4\u6a21\u6001\u7684\u4e0d\u540c\uff0c\u7ec6\u5206\u4e3a\u4e86\u8de8\u6a21\u6001\u7684\u6307\u4ee4\u5fae\u8c03\u548c\u6a21\u6001\u94fe\u6307\u4ee4\u5fae\u8c03\u3002\u6307\u4ee4\u5fae\u8c03\u7684\u6570\u636e\u96c6\u90fd\u662f\u6765\u81eaASR\u6570\u636e\u96c6\u3002\u63cf\u8ff0\u4efb\u52a1\u9700\u6c42\u7684\u6307\u4ee4\u7531GPT-4\u751f\u6210\u3002<\/p>\n\n\n\n

    \u5728\u6211\u770b\u6765\uff0c\u8fd9\u4e2a\u5de5\u4f5c\u8fd8\u662f\u76f8\u5f53\u504f\u5b66\u672f\u5316\u7684\u4f5c\u54c1\uff0c\u6587\u4e2d\u6709\u4e0d\u5c11\u70b9\u90fd\u6709\u503c\u5f97\u5546\u69b7\u7684\u5730\u65b9\uff1a\u7b2c\u4e00\uff0c\u8bed\u97f3\u7684\u79bb\u6563\u5316\u4ec5\u4ec5\u7528\u4e86HuBERT[2]\uff0c\u6a21\u578b\u53ea\u80fd\u770b\u5230\u8bed\u97f3\u7684\u8bed\u4e49\u7279\u5f81\uff0c\u8fd9\u5bf9\u6a21\u578b\u5408\u6210\u8bed\u97f3\u7684\u97f3\u8d28\u548c\u8868\u73b0\u529b\u6709\u975e\u5e38\u5927\u7684\u5f71\u54cd\uff0cdemo\u7684\u8bed\u97f3\u4e5f\u9a8c\u8bc1\u4e86\u6211\u7684\u5224\u65ad\uff1b\u7b2c\u4e8c\uff0c\u6307\u4ee4\u5fae\u8c03\u6570\u636e\u96c6\u7684\u6784\u9020\u4e0a\u6709\u95ee\u9898\u3002\u4ed6\u4eec\u7528\u7684\u662fASR\u6570\u636e\u96c6\uff0c\u5176\u5b9e\u66f4\u597d\u7684\u9009\u62e9\u5e94\u8be5\u662fTTS\u6570\u636e\u96c6\uff0c\u53ef\u60dc\u9ad8\u8d28\u91cf\u7684TTS\u6570\u636e\u96c6\u5b9e\u5728\u662f\u592a\u5c11\u4e86\u3002ASR\u6570\u636e\u96c6\u4e2d\u7684\u6587\u672c\u548c\u8bed\u97f3\u53ef\u80fd\u5e76\u4e0d\u662f\u4e25\u683c\u5bf9\u9f50\u7684\uff0cGPT-4\u4ea7\u751f\u7684meta-prompt\u548c\u8bed\u97f3\u672c\u8eab\u7684\u7279\u5f81\u4e5f\u6709\u53ef\u80fd\u662f\u5bf9\u4e0d\u4e0a\u7684\uff0c\u6bd4\u5982prompt\u8981\u6c42\u5927\u58f0\u6717\u8bfb\uff0c\u4f46\u8bed\u97f3\u672c\u8eab\u53ef\u80fd\u662f\u7279\u5b9a\u4f4e\u6c89\u7684\u3002meta-prompt\u672c\u8eab\u5c31\u65e0\u6cd5\u505a\u5230\u8db3\u591f\u590d\u6742\u4e30\u5bcc\uff0c\u4e0d\u80fd\u63cf\u8ff0\u5230\u8bed\u97f3\u7684\u4e00\u4e9b\u7ec6\u7c92\u5ea6\u4fe1\u606f\u3002<\/p>\n\n\n\n

    \u8fd9\u4e00\u90e8\u5206\uff0c\u6700\u597d\u8981\u6709\u50cf\u8bf8\u5982SALMONN[10]\u8fd9\u6837\u7684\u591a\u6a21\u6001\u8bed\u97f3\u7406\u89e3\u6a21\u578b\u7684\u4ecb\u5165\uff0c\u50cfDALLE3\u4e00\u6837\u4e30\u5bcc\u6307\u4ee4\u7684\u591a\u6837\u6027\u3002\u81f3\u4e8e\u8bed\u97f3\u65b9\u9762\uff0c\u53ef\u4ee5\u8003\u8651\u5f15\u5165zero-shot\u7684\u8bed\u97f3\u5408\u6210\u6a21\u578b\u6216\u8005\u53d8\u58f0\u6a21\u578b\u6765\u505a\u5408\u6210\u6570\u636e\u3002\u7b2c\u4e09\uff0c\u6587\u4e2d\u7684\u8bad\u7ec3\u65b9\u6cd5\u4e5f\u6ca1\u6709\u4e0e\u4eba\u7c7b\u504f\u597d\u505a\u5bf9\u9f50\u3002<\/p>\n\n\n\n

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

    \u56fe10\uff1aSpeechGPT\u7684\u6a21\u578b\u67b6\u6784<\/strong><\/p>\n\n\n\n

    \u5bf9\u4e8e\u4e0a\u9762\u7684\u7b2c\u4e00\u4e2a\u95ee\u9898\uff0c\u4f5c\u8005\u5728\u5176\u540e\u7684SpeechGPT-Gen[12]\u4e2d\u505a\u4e86\u89e3\u51b3\u3002\u89e3\u51b3\u601d\u8def\u7684\u6838\u5fc3\u70b9\u5c31\u662f\uff1a\u8ba9\u6a21\u578b\u4e0d\u4ec5\u770b\u5230\u8bed\u97f3\u7684\u8bed\u4e49token\uff0c\u4e5f\u8981\u770b\u5230\u8bed\u97f3\u7684\u58f0\u5b66token\u3002\u5177\u4f53\u505a\u6cd5\u662f\uff1aSpeechGPT\u7684HuBERT\u7279\u5f81\u66ff\u6362\u6210\u4e86SpeechTokenizer[7]\u4e2d\u7684\u8bed\u4e49\u7279\u5f81\uff0c\u7528SpeechGPT\u8fd9\u4e00LLM\u6765\u81ea\u56de\u5f52\u5730\u5efa\u6a21\u8bed\u4e49\u7279\u5f81\uff0c\u6709\u4e86\u8bed\u4e49\u7279\u5f81\u4e4b\u540e\uff0c\u518d\u4f7f\u7528Flow-Matching\u8fd9\u6837\u7684\u6269\u6563\u6a21\u578b\u6765\u5efa\u6a21\u58f0\u5b66\u7279\u5f81\u3002\u8fd9\u91cc\u9009\u7528Flow-Matching\u6269\u6563\u6a21\u578b\uff0c\u53ef\u80fd\u662f\u53d7\u4e86SD3\u548cVoicebox\/Audiobox\u7684\u5f71\u54cd\u3002\u4e3a\u4e86\u589e\u5f3a\u4e24\u9636\u6bb5\u5efa\u6a21\u7684\u4f9d\u8d56\u5173\u7cfb\uff0c\u4f5c\u8005\u5c06\u8bed\u4e49\u7279\u5f81\u7684\u5148\u9a8c\u4fe1\u606f\u6ce8\u5165\u5230\u7b2c\u4e8c\u9636\u6bb5\u6269\u6563\u6a21\u578b\u7684\u5148\u9a8c\u5206\u5e03\u4e2d\u3002\u53ef\u4ee5\u770b\u5230\uff0c\u8fd9\u91cc\u8bed\u97f3\u7684\u89e3\u7801\u5176\u5b9e\u4e5f\u662f\u4e00\u79cd\u5c42\u6b21\u5316\u6e10\u8fdb\u5f0f\u89e3\u7801\u3002<\/p>\n\n\n\n

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

    \u56fe11\uff1aSpeechGPT-Gen\u7684\u6a21\u578b\u67b6\u6784<\/strong><\/p>\n\n\n\n

    SpeechAlign[13]\u505a\u7684\u5219\u662fSLM\u4e0e\u4eba\u7c7b\u504f\u597d\u7684\u5bf9\u9f50\uff0c\u5f7b\u5e95\u5730\u5411LLM\u7684\u8bad\u7ec3\u65b9\u6cd5\u770b\u9f50\u3002\u8be5\u5de5\u4f5c\u6784\u5efa\u4e86\u5bf9\u6bd4gold token\u548c\u5408\u6210token\u7684encodec\u6570\u636e\u96c6\uff0c\u7136\u540e\u8fdb\u884c\u504f\u597d\u4f18\u5316\u6765\u8fdb\u884c\u6539\u8fdb\u3002\u4f7f\u7528\u7684\u504f\u597d\u4f18\u5316\u65b9\u6cd5\u5305\u62ecRLHF\u548cChain of Hindsight\u3002<\/p>\n\n\n\n

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

    \u56fe12\uff1aSpeechAlign\u7684\u6d41\u7a0b\u56fe<\/strong><\/p>\n\n\n\n

    \u7b80\u5355\u603b\u7ed3\u4e00\u4e0b\u4e0a\u9762\u8fd9\u4e9b\u5de5\u4f5c\u4e2d\u503c\u5f97\u5173\u6ce8\u7684\u70b9\uff1a<\/strong><\/p>\n\n\n\n

    1. \u8981\u60f3\u8ba9LLM\u8f93\u51fa\u4e0a\u4e0b\u6587\u8fde\u8d2f\u7684\u9ad8\u8868\u73b0\u529b\u8bed\u97f3\uff0c\u5fc5\u987b\u8981\u8ba9LLM\u770b\u5230\u8bed\u4e49token\u548c\u58f0\u5b66token\uff0c\u53ea\u6709\u8bed\u4e49token\uff0c\u90a3\u8bed\u97f3\u5c31\u4f1a\u663e\u5f97\u5446\u677f\u673a\u68b0\uff0c\u53ea\u6709\u58f0\u5b66token\uff0c\u90a3\u8bed\u97f3\u5c31\u4e0d\u77e5\u6240\u4e91\uff1b<\/li>
    2. LLM\u7684\u6307\u4ee4\u5fae\u8c03\u540c\u6837\u53ef\u4ee5\u8fc1\u79fb\u5230\u8bed\u97f3-\u6587\u672c\u591a\u6a21\u6001\u9886\u57df\u4e2d\uff0cLLM\u7684\u6307\u4ee4\u5fae\u8c03\u540c\u6837\u53ef\u4ee5\u5e26\u6765\u5982\u540cNLP\u4e00\u6837\u7684\u6d8c\u73b0\u6027\uff1b<\/li>
    3. \u9ad8\u8d28\u91cf\u6307\u4ee4\u5fae\u8c03\u6570\u636e\u96c6\u7684\u6784\u5efa\u5e94\u8be5\u662f\u6700\u5927\u7684\u74f6\u9888\uff01\u4e00\u4e0b\u5b50\u8ba9LLM\u540c\u65f6\u505a\u8bed\u97f3\u7406\u89e3\u548c\u8bed\u97f3\u751f\u6210\uff0c\u96be\u5ea6\u975e\u5e38\u5927\u3002\u4e0d\u5982\u5206\u6b65\u8fdb\u884c\u3002<\/li>
    4. \u5982\u679c\u8981\u5206\u6b65\u8fdb\u884c\u7684\u8bdd\uff0c\u8981\u5148\u5b9e\u73b0\u4e00\u4e2a\u7c7b\u4f3c\u4e8eSALMONN[10]\u90a3\u6837\u7684\u591a\u6a21\u6001\u7406\u89e3\u6a21\u578b\u548c\u4e00\u4e2a\u5f3a\u5927\u7684Zero-shot TTS\u6a21\u578b\u3002\u524d\u8005\u7528\u4e8e\u7ed9\u8bed\u97f3\u6570\u636e\u6253\u4e0a\u4e30\u5bcc\u7684\u6807\u7b7e\uff0c\u53ef\u4ee5\u662f\u60c5\u611f\u60c5\u7eea\u3001\u97f5\u5f8b\u3001\u97f3\u9ad8\u3001\u8bed\u901f\uff0c\u4e5f\u53ef\u4ee5\u662f\u53e3\u97f3\u3001\u610f\u56fe\u548c\u8bf4\u8bdd\u73af\u5883\uff1b\u540e\u8005\u5219\u7528\u4e8e\u751f\u6210\u9ad8\u8d28\u91cf\u7684\u8bed\u97f3\u6570\u636e\u3002\u6bd5\u7adf\uff0c\u9ad8\u8d28\u91cf\u7684\u3001\u6587\u672c\u548c\u8bed\u97f3\u4e25\u683c\u5bf9\u9f50\u7684TTS\u6570\u636e\u5b9e\u5728\u662f\u592a\u5c11\u4e86\uff0c\u5c24\u5176\u662f\u4e2d\u6587\u9886\u57df\u3002\u6709\u4e86\u8fd9\u4e24\u4e2a\u6a21\u578b\u7684\u52a0\u6301\uff0c\u6211\u4eec\u5176\u5b9e\u5c31\u80fd\u591f\u6784\u9020\u51fa\u9ad8\u8d28\u91cf\u7684\u6307\u4ee4\u5fae\u8c03\u6570\u636e\u96c6\u3002\u6211\u4e0d\u77e5\u9053OpenAI\u662f\u5426\u6709SALMONN\u8fd9\u6837\u7684\u6a21\u578b\uff0c\u4f46OpenAI\u7684OpenVoice\u6a21\u578b\u5e94\u8be5\u8db3\u591f\u4e3a\u5176\u63d0\u4f9b\u9ad8\u8d28\u91cf\u7684\u8bed\u97f3\u6570\u636e\u4e86\u3002<\/li><\/ol>\n\n\n\n

      \u65e2\u7136\u6211\u4eec\u5728\u4e0a\u9762\u7684\u7bc7\u5e45\u4e2d\u8bba\u8ff0\u4e86\u8bed\u97f3\u7406\u89e3\u591a\u6a21\u6001\u6a21\u578b\u7684\u6784\u5efa\uff0c\u90a3\u6211\u4eec\u5728\u4e0b\u4e00\u90e8\u5206\u5c31\u91cd\u70b9\u5173\u6ce8zero-shot TTS\u6a21\u578b\uff0c\u5b83\u5bf9\u9ad8\u8d28\u91cf\u6307\u4ee4\u5fae\u8c03\u6570\u636e\u96c6\u7684\u6784\u5efa\u540c\u6837\u81f3\u5173\u91cd\u8981\u3002\u540c\u65f6\uff0cLLM\u89e3\u7801\u8bed\u97f3\u7684\u65b9\u6cd5\u4e5f\u80fd\u4ecezero-shot TTS\u65b9\u6848\u4e2d\u5f97\u5230\u4e0d\u5c11\u7684\u542f\u53d1\u3002<\/p>\n\n\n\n

      LLM\u5982\u4f55\u5408\u6210\u8bed\u97f3\uff1aZero-shot TTS<\/strong><\/h2>\n\n\n\n

      \u524d\u9762\u8bf4\u5230\uff0cSLM\u8bcd\u8868\u4e2d\u5305\u542b\u4e86\u8bed\u97f3\u7684\u8bed\u4e49token\u548c\u58f0\u5b66token\u3002\u8bed\u4e49token\u4fdd\u8bc1\u751f\u6210\u8bed\u97f3\u4e0e\u5bf9\u8bdd\u4e0a\u4e0b\u6587\u7684\u8fde\u8d2f\u6027\uff0c\u58f0\u5b66token\u4fdd\u8bc1\u4e86\u5408\u6210\u8bed\u97f3\u7684\u8d28\u91cf\u548c\u8868\u73b0\u529b\u3002\u8981\u60f3\u505a\u5230\u5408\u6210\u4e0a\u4e0b\u6587\u8fde\u8d2f\u7684\u9ad8\u81ea\u7136\u5ea6\u8bed\u97f3\uff0c\u6709\u4e24\u4e2a\u95ee\u9898\u5fc5\u987b\u8981\u89e3\u51b3\uff1a<\/p>\n\n\n\n

      1. \u8bed\u97f3\u65e2\u6709\u8bed\u4e49token\uff0c\u53c8\u6709\u58f0\u5b66token\uff0c\u5e94\u8be5\u8981\u5982\u4f55\u89e3\u7801\u6210\u8bed\u97f3\uff1f<\/strong><\/li>
      2. SLM\u5728\u5408\u6210\u8bed\u97f3\u7684\u8fc7\u7a0b\u4e2d\u662f\u5426\u80fd\u591f\u9075\u5faa\u591a\u8f6e\u5bf9\u8bdd\u4e2d\u7684\u6587\u672c\u6307\u4ee4\u548c\u8bed\u97f3\u6307\u4ee4\uff1f\u8fd9\u4e2a\u5f88\u91cd\u8981\uff01\u8fd9\u5141\u8bb8\u6a21\u578b\u6839\u636e\u7528\u6237\u7684\u5373\u65f6\u8981\u6c42\u6765\u751f\u6210\u8bed\u97f3\u56de\u590d\u3002\u6bd4\u5982\u8bf4\uff0cOpenAI\u6f14\u793a\u89c6\u9891\u4e2d\u51fa\u73b0\u7684\uff1a\u201c\u5c06\u8bed\u901f\u63d0\u9ad8\u4e24\u500d\u201d\u3001\u201c\u91c7\u7528\u66f4\u52a0\u673a\u68b0\u5316\u7684\u8bed\u6c14\u201d\u8fd9\u6837\u7684\u8981\u6c42\u3002<\/strong><\/li><\/ol>\n\n\n\n

        \u5bf9\u4e8e\u7b2c\u4e00\u4e2a\u95ee\u9898\uff0c\u4ee5VALL-E[16]\u4e3a\u4ee3\u8868\u7684\u8bf8\u591azero-shot TTS\u6a21\u578b\u7ed9\u51fa\u4e86\u4e0d\u540c\u7684\u89e3\u51b3\u65b9\u6848\uff0c\u8fd9\u4e9b\u65b9\u6848\u867d\u6709\u4e0d\u540c\uff0c\u4f46\u4e5f\u6709\u4e0d\u53ef\u5ffd\u89c6\u7684\u5171\u540c\u70b9\uff1b\u5bf9\u4e8e\u7b2c\u4e8c\u4e2a\u95ee\u9898\uff0c\u4ee5VoiceLDM[18]\u548cParlerTTS[19]\u4e3a\u4ee3\u8868\u7684text\/prompt-guided zero-shot TTS\u5de5\u4f5c\u7ed9\u51fa\u4e86\u80af\u5b9a\u7684\u7b54\u6848\u3002\u7b80\u5355\u89e3\u91ca\u4e00\u4e0btext\/prompt-guided zero-shot TTS\u662f\u600e\u4e48\u56de\u4e8b\uff0c\u901a\u5e38\u7684\u8bed\u97f3\u5408\u6210\u5c31\u662f\u5c06\u6587\u672c(transcription)\u8f6c\u6362\u6210\u58f0\u97f3\uff0c\u8be5\u4efb\u52a1\u5728transcription\u4e4b\u5916\uff0c\u53c8\u589e\u52a0\u4e86description\u7684\u8f93\u5165\uff0c\u6765\u63cf\u8ff0\u5408\u6210\u8bed\u97f3\u7684\u60c5\u611f\u60c5\u7eea\u3001\u53e3\u97f3\u3001\u8bed\u6c14\u3001\u8bed\u901f\u3001\u97f3\u9ad8\u3001\u8bf4\u8bdd\u73af\u5883\u3001\u6c1b\u56f4\u7b49\u7b49\u4fe1\u606f\u3002\u6211\u4eec\u9010\u4e2a\u6765\u770b\u8fd9\u4e9b\u5de5\u4f5c\u3002<\/p>\n\n\n\n

        Zero-shot TTS<\/strong><\/p>\n\n\n\n

        2023\u5e74\u4ee5\u6765\uff0c\u5b66\u672f\u754c\u548c\u5de5\u4e1a\u754c\u51fa\u4e86\u4e0d\u5c11\u5177\u5907in-context learning(zero-shot\/few-shot)\u80fd\u529b\u7684TTS\u6a21\u578b\u3002\u8fd9\u4e9bTTS\u6a21\u578b\u901a\u5e38\u4f1a\u5c06\u4f4e\u4fe1\u606f\u5bc6\u5ea6\u3001\u957f\u5e8f\u5217\u7684\u8fde\u7eed\u8bed\u97f3\u6570\u636e\u538b\u7f29\u4e3a\u9ad8\u4fe1\u606f\u5bc6\u5ea6\u7684tokens\u6216\u8005latents(\u5176\u5b9e\u5c31\u662f\u7801\u672c\u4e2d\u5177\u4f53\u7684token embedding)\u3002\u8fd9\u4e9b\u6a21\u578b\u672c\u8d28\u4e0a\u505a\u7684\u4e8b\u60c5\u5c31\u662f\uff1a\u5982\u4f55\u9ad8\u6548\u5b9e\u73b0\u8bed\u97f3tokens\/latents\u5230\u97f3\u9891\u6ce2\u5f62\u7684\u6620\u5c04\u3002<\/p>\n\n\n\n

        \u8fd9\u4e9b\u6a21\u578b\u7ed9\u51fa\u7684\u89e3\u51b3\u65b9\u6848\u57fa\u672c\u4e0a\u90fd\u9075\u5faa\u4e00\u4e2a\u51c6\u5219\uff1a\u8bed\u4e49token\u548c\u58f0\u5b66token\u5c42\u6b21\u5316\u89e3\u7801\uff0c\u5148\u8bed\u4e49\u540e\u58f0\u5b66\uff0c\u6216\u8005\u5148\u89e3\u7801\u6210MEL\u518d\u540e\u63a5\u58f0\u7801\u5668\uff0c\u5e76\u4e14\u975e\u5fc5\u8981\u4e0d\u505a\u81ea\u56de\u5f52(\u6bd5\u7adf\u81ea\u56de\u5f52\u4e0a\u7ebf\u867d\u9ad8\uff0c\u4f46\u592a\u5403\u6570\u636e\u4e86)\uff01\u6211\u4eec\u4e00\u4e2a\u4e2a\u6765\u770b\u3002<\/p>\n\n\n\n

        \u57fa\u4e8e\u58f0\u5b66token\u6216\u8bed\u4e49token\u7684\u5de5\u4f5c<\/strong><\/p>\n\n\n\n

        \u5148\u662f\u5fae\u8f6f\u7684VALL-E[16]\u3002\u8fd9\u662fzero-shot TTS\u7684\u5f00\u5c71\u4e4b\u4f5c\uff0c\u9996\u6b21\u5728TTS\u4efb\u52a1\u4e0a\u91c7\u7528\u4e86\u4e0a\u4e07\u5c0f\u65f6\u7684\u6570\u636e\u3002\u5b83\u91c7\u7528Encodec\u5c06\u8bed\u97f3\u8f6c\u6362\u4e3a\u79bb\u6563\u7684token\uff0c\u7136\u540e\u7528GPT\u5728token\u4e0a\u505a\u8bed\u8a00\u6a21\u578b\u7684\u4efb\u52a1\u3002\u4f46\u662f\uff0c\u8bed\u97f3\u6bd5\u7adf\u4e0d\u662f\u6587\u672c\uff0c\u5982\u679c\u76f4\u63a5\u5728\u8bed\u97f3\u7684\u6240\u6709\u7279\u5f81\u4e0a\u90fd\u505a\u81ea\u56de\u5f52\u7684\u8bdd\uff0c\u90a3\u8bad\u7ec3\u7684\u6210\u672c\u4f1a\u76f8\u5f53\u9ad8\u3002\u8003\u8651\u5230Encodec RVQ\u7279\u5f81\u7684\u5c42\u6b21\u6027\uff0c\u4f4e\u5c42\u7279\u5f81\u8868\u793a\u8bed\u4e49\u5185\u5bb9\u8fd9\u6837\u7684\u91cd\u8981\u7279\u5f81\uff0c\u9ad8\u5c42\u7279\u5f81\u5219\u8868\u5f81\u58f0\u5b66\u7ec6\u8282\u3002\u524d\u8005\u5177\u6709\u6bd4\u8f83\u5f3a\u7684\u4e0a\u4e0b\u6587\u4f9d\u8d56\u5173\u7cfb\uff0c\u9002\u5408\u7528\u81ea\u56de\u5f52\u6765\u5efa\u6a21\uff0c\u540e\u8005\u8bf8\u5982\u97f3\u8272\u8fd9\u6837\u7684\u7279\u5f81\uff0c\u5177\u6709\u5168\u5c40\u6027\uff0c\u7528\u975e\u81ea\u56de\u5f52\u7279\u5f81\u4e5f\u53ef\u4ee5\u641e\u5b9a\uff0c\u6240\u4ee5\u5c31\u6709\u4e86VALLE\u4e2d\u81ea\u56de\u5f52+\u975e\u81ea\u56de\u5f52\u7684\u5c42\u6b21\u5efa\u6a21\u65b9\u5f0f\u3002<\/p>\n\n\n\n

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

        \u56fe13\uff1aVALL-E\u7684\u6a21\u578b\u67b6\u6784<\/strong><\/p>\n\n\n\n

        \u5c3d\u7ba1VALL-E[16]\u5728\u7528GPT\u5efa\u6a21token\u7684\u4e0a\u4e0b\u6587\u5173\u7cfb\u7684\u65f6\u5019\uff0c\u57fa\u4e8etoken\u7684\u5c42\u6b21\u5316\u7279\u6027\u505a\u4e86\u5206\u6cbb\u5904\u7406\uff0c\u53ef\u80fd\u662f\u9650\u4e8e\u5f53\u524d\u8bed\u97f3\u6570\u636e\u96c6\u7684\u89c4\u6a21(\u51e0\u4e07\u5c0f\u65f6\u53ef\u80fd\u4e0d\u591f)\uff0c\u8fd9\u79cdGPT\u81ea\u56de\u5f52\u7684\u96be\u5ea6\u8fd8\u662f\u76f8\u5f53\u5927\u7684\uff0c\u89e3\u7801\u8fc7\u7a0b\u5b58\u5728\u5e38\u89c1\u7684\u9519\u8bef\u4f20\u64ad\u73b0\u8c61\uff0c\u9c81\u68d2\u6027\u975e\u5e38\u5dee\uff0c\u6781\u5176\u4e0d\u7a33\u5b9a\u3002\u6839\u636eIlya Sutskever\u6b64\u524d\u5bf9\u4e8e\u81ea\u56de\u5f52\u7684\u8bba\u8ff0\uff0cGPT\u81ea\u56de\u5f52\u76f8\u6bd4\u4e8eBERT\u8fd9\u79cd\u53cc\u5411\u7ed3\u6784\u662f\u975e\u5e38data-hungry\u7684\uff0c\u4e07\u5c0f\u65f6\u7684\u6570\u636e\u53ef\u80fd\u4e0d\u591f\u3002\u6839\u636e\u672c\u4eba\u4ee5\u53ca\u4e00\u4e9b\u540c\u884c\u7684\u7ecf\u9a8c\uff0cVALL-E\u6a21\u578b\u8fd9\u4e00\u7c7b\u7684\u81ea\u56de\u5f52\u6a21\u578b\uff0c\u4e5f\u5305\u62ectortoise-tts[20]\u548cxtts v2\uff0c\u8981\u60f3\u663e\u51fa\u5a01\u529b\uff0c\u81f3\u5c11\u8981\u6709\u5341\u51e0\u4e07\u5c0f\u65f6\u7684\u6570\u636e\u624d\u884c\u3002<\/p>\n\n\n\n

        \u65e2\u7136GPT\u81ea\u56de\u5f52\u7684\u96be\u5ea6\u8fd9\u4e48\u5927\uff0c\u5c31\u6709\u4e0d\u5c11\u4eba\u60f3\u65b9\u8bbe\u6cd5\u5730\u6765\u964d\u4f4eGPT\u5b66\u4e60\u7684\u96be\u5ea6\u4e86\u3002\u4ed6\u4eec\u7684\u89e3\u51b3\u65b9\u6848\u4e5f\u975e\u5e38\u7c7b\u4f3c\uff1a\u7ed9GPT\u63d0\u4f9b\u989d\u5916\u7684\u6761\u4ef6\u4fe1\u606f\u4e0d\u5c31\u884c\u4e86\u3002\u6bd4\u8f83\u5178\u578b\u7684\u5de5\u4f5c\u5c31\u662f\u5fae\u8f6f\u7684RALL-E[21]\u548c\u5409\u5229\u7684HAM-TTS[22]\u3002RALL-E\u5148\u751f\u6210\u4e86\u65f6\u957f\u4fe1\u606f\u548c\u97f3\u9ad8\u4fe1\u606f\uff0c\u4f5c\u4e3aGPT\u81ea\u56de\u5f52\u7684\u5148\u9a8c\uff0c\u4e4b\u6240\u4ee5\u4f1a\u8865\u5145\u65f6\u957f\u548c\u97f3\u9ad8\uff0c\u8fd9\u5927\u6982\u662f\u53d7\u5230FastSpeech2[23]\u8fd9\u6837\u7684\u975e\u81ea\u56de\u5f52\u6a21\u578b\u7684\u542f\u53d1\uff0c\u8fd9\u4e24\u4e2a\u6307\u6807\u7684\u5f15\u5165\uff0c\u6709\u52a9\u4e8e\u63d0\u5347\u5408\u6210\u7684\u9c81\u68d2\u6027\uff1bHAM-TTS\u5219\u662f\u8865\u5145\u4e86\u57fa\u4e8eHuBERT\u7684\u8bed\u4e49\u4fe1\u606f\u3002\u503c\u5f97\u6ce8\u610f\u5730\u662f\uff0cHAM-TTS\u5c06\u6a21\u578b\u7684\u8bad\u7ec3\u6570\u636e\u6269\u5145\u5230\u4e8665\u4e07\u5c0f\u65f6\uff0c\u5176\u4e2d\u670950\u4e07\u5c0f\u65f6\u7684\u6570\u636e\u662f\u5408\u6210\u6570\u636e\u3002\u5408\u6210\u6570\u636e\u4e5f\u80fd\u5927\u5e45\u5ea6\u63d0\u5347\u5408\u6210\u8bed\u97f3\u7684\u97f3\u8d28\u3002<\/strong><\/p>\n\n\n\n

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

        \u56fe14\uff1aRALL-E\u7684\u6a21\u578b\u67b6\u6784\uff0c\u6846\u51fa\u6765\u7684\u5c31\u662f\u8f85\u52a9\u4fe1\u606f<\/strong><\/p>\n\n\n\n

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

        \u56fe15\uff1aHAM-TTS\u7684\u6a21\u578b\u67b6\u6784<\/strong><\/p>\n\n\n\n

        \u8bf4\u5230VALL-E\u7684\u540e\u7eed\u6539\u8fdb\uff0cVoiceCraft\u4e0d\u5f97\u4e0d\u63d0\u3002\u6211\u613f\u610f\u79f0\u4e4b\u4e3a\u201c\u4f18\u96c5\u7684VALL-E\u201d\u3002\u5b83\u7684\u4f18\u96c5\u4e3b\u8981\u4f53\u73b0\u5728\u4e24\u4e2a\u65b9\u9762\uff1acasual masking\u548cdelayed stacking\u3002\u6240\u8c13\u7684causal masking\uff0c\u662f\u4e3a\u4e86\u7528\u81ea\u56de\u5f52GPT\u67b6\u6784\u6765\u505a\u8bed\u97f3\u7f16\u8f91\u4efb\u52a1\uff0c\u5c31\u662f\u628a\u88abmask\u7684\u90e8\u5206\u79fb\u52a8\u5230\u5e8f\u5217\u672b\u5c3e\u53bb\u9884\u6d4b\uff0c\u4e00\u5957\u67b6\u6784\u540c\u65f6\u505a\u5408\u6210\u548c\u7f16\u8f91\u4efb\u52a1\uff1b\u6240\u8c13\u7684delay stacking\uff0c\u662f\u4e3a\u4e86\u9002\u914d\u81ea\u56de\u5f52\u548cRVQ\uff0c\u901a\u8fc7delay\u9519\u4f4d\u8ba9\u5f53\u524d\u7801\u672c\u7684token\u9884\u6d4b\u6b63\u597d\u53ef\u4ee5\u5229\u7528\u524d\u9762\u90a3\u4e9btoken\u7684\u9884\u6d4b\u7ed3\u679c\uff0c\u6bd4\u8d77VALL-E\u90a3\u6837\u81ea\u56de\u5f52\u548c\u975e\u81ea\u56de\u5f52\u7f1d\u5408\u5728\u4e00\u8d77\u7684\u7ed3\u6784\u8981\u4f18\u96c5\u4e0d\u5c11\u3002<\/p>\n\n\n\n

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

        \u56fe16\uff1aVoiceCraft\u7684\u5efa\u6a21\u6d41\u7a0b<\/strong><\/p>\n\n\n\n

        \u57fa\u4e8e\u58f0\u5b66\/\u8bed\u4e49latents\u7684\u5de5\u4f5c<\/strong><\/p>\n\n\n\n

        \u6211\u4eec\u901a\u5e38\u6240\u8bf4\u7684\u8bed\u97f3token\u662f\u79bb\u6563\u7684\u3002\u5982\u679c\u4f7f\u7528\u5bf9\u5e94\u7801\u672c\u4e2d\u7684embedding\u6765\u8868\u793a\u8bed\u97f3\u7684\u8bdd\uff0c\u5b83\u4e5f\u53ef\u4ee5\u662f\u8fde\u7eed\u7684\u4f4e\u7ef4\u5ea6\u7684latent\u53d8\u91cf\u3002\u65e2\u7136\u662f\u4f4e\u7ef4\u5ea6\u7684\u8fde\u7eedlatent\u53d8\u91cf\uff0c\u90a3\u56fe\u50cf\u5408\u6210\u9886\u57df\u4e2d\u5927\u706b\u7684LDM(latent diffusion model\uff0c\u5176\u5b9e\u5c31\u662fstable diffsion 1&2\u91c7\u7528\u7684\u6a21\u578b)\u6a21\u578b[]\u81ea\u7136\u4e5f\u53ef\u4ee5\u7528\u5230\u8bed\u97f3\u7684\u5408\u6210\u4e0a\u3002\u8fd9\u65b9\u9762\u7684\u7ecf\u5178\u5de5\u4f5c\u6709\u5f88\u591a\uff0c\u6bd4\u5982\u8bf4\uff1aNaturalSpeech 2&3[25, 26]\u3001AudioLDM 2[27]\u3001VoiceLDM[18]\u3002\u4f46\u8fd9\u91cc\u9762\u53ea\u6709NaturalSpeech2\u7528\u5230\u4e86\u8bed\u97f3\u79bb\u6563\u5316\u90e8\u5206\u63d0\u53ca\u7684\u58f0\u5b66\/\u8bed\u4e49token\uff0cNaturalSpeech3\u7684\u5c5e\u6027\u5206\u89e3\u5f62\u5f0f\u7684VQ\u66f4\u50cf\u662f\u53e6\u4e00\u79cd\u5f62\u5f0f\u7684RVQ\u3002\u6211\u4eec\u5148\u6765\u770bNaturalSpeech 2&3\uff0c\u5176\u4ed6\u7684\u5de5\u4f5c\u540e\u9762\u518d\u6765\u770b\u3002<\/p>\n\n\n\n

        \u9996\u5148\u662fNaturalSpeech 2[26]\uff0c\u5b83\u57fa\u672c\u4e0a\u5c31\u662fVALL-E\u7684\u8fde\u7eed\u7248\u672c\u3002\u5b83\u7528\u7684latent\u4e5f\u662f\u6765\u81eaEncodec\uff0c\u5bf9\u5176\u4e2d\u4e0d\u540c\u5c42\u6b21\u7684latent\u505a\u4e86\u6c42\u548c\uff0c\u7136\u540e\u5c06\u5176\u4f5c\u4e3a\u6269\u6563\u6a21\u578b\u7684\u8bad\u7ec3\u76ee\u6807\u3002\u503c\u5f97\u6ce8\u610f\u5730\u662f\uff0c\u6269\u6563\u6a21\u578b\u548cFastSpeech2\u4e00\u6837\u4e5f\u7528\u4e86\u65f6\u957f\u548c\u97f3\u9ad8\u4f5c\u4e3a\u5408\u6210\u7684\u5148\u9a8c\u6761\u4ef6\u3002\u8fd9\u4e00\u70b9\u4e5f\u88ab\u540e\u6765\u7684RALL-E\u91c7\u7528\u3002\u8be5\u5de5\u4f5c\u4e2d\u7684\u6269\u6563\u6a21\u578b\u91c7\u7528WaveNet\u5b9e\u73b0\uff0c\u540c\u65f6\u9884\u6d4b\u4e0d\u52a0\u566a\u7684latent\u548c\u540e\u9a8c\u5747\u503c\uff0c\u548c\u56fe\u50cf\u5408\u6210\u9886\u57df\u7684\u6269\u6563\u6a21\u578b\u5728\u5b9e\u73b0\u65b9\u5f0f\u4e0a\u8fd8\u662f\u6709\u6240\u4e0d\u540c\u7684\u3002<\/p>\n\n\n\n

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

        \u56fe17\uff1aNaturalSpeech2\u7684\u6a21\u578b\u67b6\u6784<\/strong><\/p>\n\n\n\n

        \u7136\u540e\u662fNaturalSpeech 3[26]\uff0c\u8fd8\u662f\u975e\u81ea\u56de\u5f52\u7684\uff0c\u800c\u4e14\u975e\u81ea\u56de\u5f52\u7684\u6b63\u7edf\u6027\u5473\u9053\u66f4\u52a0\u6d53\u539a\uff0c\u501f\u7528\u4e86\u4e0d\u5c11FastSpeech2\u548cmegatts1&2(\u540e\u9762\u4f1a\u8bb2)[27, 28]\u7684\u8bbe\u8ba1\u601d\u60f3\u3002\u50cfmegatts 1&2\u4e00\u6837\uff0c\u540c\u6837\u91c7\u7528(\u81ea)\u76d1\u7763\u4fe1\u53f7\u5bf9\u8bed\u97f3token\u7f16\u7801\u7684\u5185\u5bb9\u505a\u4e86\u9650\u5236\uff0c\u800c\u4e0d\u518d\u50cf\u662fVALL-E\/NaturalSpeech2\u90a3\u6837\u4e00\u628a\u6293\u3002\u76f8\u5e94\u5730\uff0c\u8bed\u97f3token\u5316\u7684\u65b9\u6cd5\u4e5f\u7528VQ\u5c31\u884c\u3002\u5177\u4f53\u800c\u8a00\uff0c\u6587\u7ae0\u5c06\u8bed\u97f3\u4fe1\u53f7\u5206\u89e3\u4e3a\u65f6\u957f\u3001\u5185\u5bb9\u3001\u97f5\u5f8b\u548c\u7ec6\u8282\u56db\u4e2a\u90e8\u5206\uff0c\u7136\u540e\u6bcf\u4e2a\u90e8\u5206\u7528\u79bb\u6563\u5316\u7684\u6269\u6563\u6a21\u578b\u6765\u5efa\u6a21\u3002\u4e0d\u8fc7\uff0c\u539f\u6587\u4f7f\u7528GRL\u6765\u4fc3\u8fdb\u8bed\u97f3\u5c5e\u6027\u7684\u5206\u89e3\uff0c\u8fd9\u4e00\u70b9\u7684\u9760\u8c31\u7a0b\u5ea6\u5b58\u7591\u3002\u6211\u4e5f\u5c1d\u8bd5\u8fc7\u6587\u7ae0\u7684FACodec\uff0c\u4f46\u6548\u679c\u5f88\u5dee\u3002\u4e09\u7ea7\u6269\u6563\u6a21\u578b\u7ea7\u8054\u7684\u7ed3\u6784\uff0c\u9884\u6d4b\u8d77\u6765\u4f3c\u4e4e\u4e5f\u975e\u5e38\u9ebb\u70e6\u3002<\/p>\n\n\n\n

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

        \u56fe18\uff1aNaturalSpeech3\u7684\u6a21\u578b\u67b6\u6784<\/strong><\/p>\n\n\n\n

        \u57fa\u4e8eMEL\u8c31+VQ\u7684TOKEN\u7684\u5de5\u4f5c<\/strong><\/p>\n\n\n\n

        \u5f53\u7136\uff0c\u4e5f\u6709\u4e0d\u5c11\u5de5\u4f5c\u7528\u4e86MEL\u8c31\u4f5c\u4e3a\u4e2d\u95f4\u7279\u5f81\uff0c\u7136\u540e\u5728\u6885\u5c14\u8c31\u7684\u57fa\u7840\u4e0a\uff0c\u6216\u662f\u7528VQ\u63d0\u4f9b\u79bb\u6563token\uff0c\u6216\u662f\u7528CNN\u6765\u63d0\u53d6\u8fde\u7eedlatent\u3002\u5bf9\u4e8eMEL+VQ\u7684\u5de5\u4f5c\uff0c\u6709tortoise-tts[20]\u3001xtts 1&2\u3001megatts1&2[28, 29]\u3001base TTS[30]\u3002\u5bf9\u4e8eMEL+latents\u7684\u5de5\u4f5c\uff0c\u6709\uff1aAudioLDM 1&2[27]\u3001StyleTTS 1&2[31, 32]\u3002\u6211\u4eec\u6765\u7b80\u5355\u770b\u770b\u662f\u5b83\u4eec\u662f\u600e\u4e48\u505a\u7684\u3002<\/p>\n\n\n\n

        Tortoise-tts<\/strong>[20]\u3002\u8be5\u5de5\u4f5c\u662f\u8457\u540d\u7684\u5f00\u6e90\u82f1\u6587TTS\u6a21\u578b\u3002\u5176\u4f5c\u8005\u76ee\u524d\u5728OpenAI\u5c31\u804c\uff0c\u540c\u65f6\u4e5f\u662fGPT-4o\u7684\u91cd\u8981Contributor(\u4ed6\u81ea\u4e2a\u513f\u5728\u535a\u5ba2\u4e2d\u8bf4\u7684)\u3002<\/strong>Tortoise-tts\u4f7f\u7528MEL+VQVAE\u7684\u65b9\u6cd5\u5f97\u5230\u8bed\u97f3\u7684MEL token\uff0c\u7136\u540e\u5bf9MEL token\u4ee5\u53catext token\u505aGPT\u81ea\u56de\u5f52\u5efa\u6a21\u3002\u5bf9\u4e8e\u8bed\u97f3\u7684\u89e3\u7801\uff0c\u81ea\u7136\u4e5f\u662f\u5206\u4e3a\u4e24\u6b65\uff1a\u5148\u662f\u7528\u6269\u6563\u6a21\u578b\u5c06MEL token\u8f6c\u6362\u4e3aMEL\u8c31\uff0c\u8fd9\u4e00\u6b65\u548c\u6587\u751f\u56fe\u5f88\u50cf\uff0c\u7528\u6269\u6563\u6a21\u578b\u662f\u5f88\u81ea\u7136\u7684\u9009\u62e9\uff1b\u7136\u540e\u7528\u58f0\u7801\u5668\u5c06MEL\u8c31\u8f6c\u6362\u4e3a\u97f3\u9891\u6ce2\u5f62\u3002tortoise-tts\u548cVALL-E\u7684\u4e3b\u4f53\u90fd\u662f\u81ea\u56de\u5f52\u5efa\u6a21\uff0c\u4e8c\u8005\u7684\u4e0d\u540c\u4e3b\u8981\u5728\u4e8etoken\u7684\u4e0d\u540c\u3002<\/p>\n\n\n\n

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

        \u56fe19\uff1atortoise-tts\u7684\u6a21\u578b\u67b6\u6784<\/strong><\/p>\n\n\n\n

        MegaTTS 1&2<\/strong>[28, 29]\u3002\u5b57\u8282\u8df3\u52a8\u7684MegaTTS\u7cfb\u5217\u5bf9\u8bed\u97f3token\u7f16\u7801\u4fe1\u606f\u505a\u4e86\u663e\u5f0f\u7684\u4fe1\u606f\u538b\u7f29\u5904\u7406\uff0c\u8ba9\u8bed\u97f3token\u4ec5\u7f16\u7801\u4e0a\u4e0b\u6587\u4f9d\u8d56\u5f3a\u7684\u97f5\u5f8b\u4fe1\u606f\uff0c\u7136\u540e\u7528GPT\u81ea\u56de\u5f52\u6765\u5efa\u6a21\u8bed\u97f3\u7684\u97f5\u5f8b\u3002\u5bf9\u4e8e\u5176\u4ed6\u65b9\u9762\u7684\u4fe1\u606f\uff0c\u6a21\u578b\u7684\u5904\u7406\u663e\u5f97\u8f83\u4e3a\u5e38\u89c4\uff1a\u97f3\u8272\u4e00\u822c\u5177\u6709\u5168\u5c40\u6027\uff0c\u4f7f\u7528\u5355\u4e00\u7684\u97f3\u8272\u7f16\u7801\u5668\u4ece\u53c2\u8003\u97f3\u9891\u4e2d\u63d0\u53d6\u5c31\u6027\uff1b\u5bf9\u4e8e\u6587\u672c\u8bed\u4e49\u5185\u5bb9\u7684\u5904\u7406\uff0c\u6a21\u578b\u5728\u5f88\u5927\u7a0b\u5ea6\u4e0a\u53c2\u8003\u4e86\u975e\u81ea\u56de\u5f52\u7684FastSpeech 2\u3002<\/p>\n\n\n\n

        \u5bf9\u4e8e\u8bed\u97f3\u7684\u89e3\u7801\uff0c\u4e5f\u662f\u5206\u4e3a\u4e24\u6b65\uff1a\u5148\u901a\u8fc7MEL decoder\u8fd8\u539f\u4e3aMEL\u8c31\uff0c\u7136\u540e\u901a\u8fc7\u58f0\u7801\u5668\u89e3\u7801\u4e3a\u97f3\u9891\u6ce2\u5f62\u3002MegaTTS 2\u548c1\u603b\u4f53\u4e0a\u7c7b\u4f3c\uff0c\u5728\u97f3\u8272\u7f16\u7801(\u97f3\u7d20\u7ea7\u7f16\u7801\u3001\u591a\u6761\u53c2\u8003\u97f3\u9891)\u3001\u8bed\u97f3\u63d0\u793a\u957f\u5ea6(\u6269\u5c55\u540cspeaker\u8bed\u97f3\u4e0a\u4e0b\u6587\u957f\u5ea6\u786ctrain\uff0c\u97f3\u9891prompt\u957f\u5ea6\u66f4\u957f)\u548c\u65f6\u957f\u5efa\u6a21(\u4e5f\u7528GPT\u81ea\u56de\u5f52)\u4e0a\u505a\u4e86\u6539\u8fdb\uff0c\u540c\u65f6\u5806\u4e86\u66f4\u5927\u89c4\u6a21\u7684\u6570\u636e\u3002\u526a\u6620\u7684\u540e\u7aefTTS\u6a21\u578b\u7528\u7684\u5c31\u662fmegatts2\u3002\u8be5\u5de5\u4f5c\u5728\u5404\u8bba\u6587\u7684\u8bc4\u6d4b\u4e2d\u8868\u73b0\u4e5f\u90fd\u4e0d\u9519\u3002<\/p>\n\n\n\n

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

        \u56fe20\uff1amegatts1\u7684\u6a21\u578b\u67b6\u6784<\/strong><\/p>\n\n\n\n

        \u57fa\u4e8eMEL\u8c31+VAE\u7684latents\u7684\u5de5\u4f5c<\/strong><\/p>\n\n\n\n

        AudioLDM 1&2<\/strong>[27]\u3002AudioLDM 1&2\u4f7f\u7528\u7684\u8bed\u97f3latents\u662f\u4e00\u81f4\u7684\uff0c\u5747\u901a\u8fc7MEL+VAE\u83b7\u5f97\u3002\u65e2\u7136\u662f\u8fde\u7eed\u7684latents\uff0c\u4f7f\u7528\u6269\u6563\u6a21\u578b\u6765\u5efa\u6a21\u4e5f\u5408\u60c5\u5408\u7406\u3002\u89e3\u7801\u8fc7\u7a0b\u4e5f\u76f8\u5f53\u7b80\u5355\uff1aVAE decoder\u83b7\u5f97\u6885\u5c14\u8c31\uff0c\u7136\u540e\u7528\u58f0\u7801\u5668\u8f6c\u6362\u4e3a\u97f3\u9891\u6ce2\u5f62\u3002\u8be5\u7cfb\u5217\u5de5\u4f5c\u7684\u6838\u5fc3\u521b\u65b0\u70b9\u662f\u5229\u7528\u591a\u6a21\u6001\u6a21\u578b\u7edf\u4e00\u4e86\u6269\u6563\u6a21\u578b\u6761\u4ef6\u8f93\u5165\u4fa7\u7684\u4fe1\u606f\uff1aAudioLDM 1\u7528CLAP\u7edf\u4e00\u4e86\u6587\u672c\u6a21\u6001\u548c\u97f3\u9891\u6a21\u6001\uff0c\u7528\u5355\u6a21\u6001\u7684\u97f3\u9891\u6570\u636e\u5c31\u80fd\u5b8c\u6210\u6a21\u578b\u7684\u8bad\u7ec3\uff1bAudioLDM 2\u5219\u5305\u542b\u4e86\u56fe\u50cf\u3001\u6587\u672c\u3001\u8f6c\u5f55\u6587\u672c\u7b49\u66f4\u591a\u6a21\u6001\uff0c\u6a21\u578b\u6cdb\u7528\u6027\u4e5f\u66f4\u5f3a\uff0c\u65e2\u80fd\u505a\u8bed\u97f3\u5408\u6210\uff0c\u4e5f\u80fd\u505a\u97f3\u4e50\u751f\u6210\u3001\u97f3\u9891\u4e8b\u4ef6\u751f\u6210\u3002<\/p>\n\n\n\n

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

        \u56fe21\uff1aAudioLDM 1\u7684\u6a21\u578b\u67b6\u6784<\/strong><\/p>\n\n\n\n

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

        \u56fe22\uff1aAudioLDM2\u7684\u6a21\u578b\u67b6\u6784<\/strong><\/p>\n\n\n\n

        StyleTTS 1&2<\/strong>[31, 32]\u3002StyleTTS\u7cfb\u5217\u7684\u6a21\u578b\u4e00\u4f17zero-shot TTS\u6a21\u578b\u663e\u5f97\u6bd4\u8f83\u8001\u6d3e\uff0c\u6574\u4f53\u7ed3\u6784\u57fa\u672c\u4e0a\u6cbf\u88ad\u4e86\u975e\u81ea\u56de\u5f52\u7684FastSpeech 2\uff0c\u4e0d\u540c\u4e4b\u5904\u5728\u4e8e\u589e\u52a0\u4e86\u57fa\u4e8e\u53c2\u8003\u97f3\u9891\u62bd\u53d6\u7684\u98ce\u683c\u4fe1\u606f\u3002\u8bf4\u662f\u98ce\u683c\uff0c\u5176\u5b9e\u8ddfmegatts\u7684\u97f3\u8272\u5f88\u50cf\u3002StyleTTS 2\u7684\u5de5\u4f5c\u5219\u5c06\u98ce\u683c\u8fdb\u4e00\u6b65\u62c6\u5206\u6210\u58f0\u5b66\u98ce\u683c\u548c\u97f5\u5f8b\u98ce\u683c\u3002\u8bad\u7ec3\u65f6\u7684\u98ce\u683c\u4fe1\u606f\u7531\u97f3\u9891\u63d0\u4f9b\uff0c\u63a8\u65ad\u65f6\u7684\u98ce\u683c\u4fe1\u606f\u5219\u7531\u6269\u6563\u6a21\u578b\u63d0\u4f9b\u3002StyleTTS 2\u901a\u8fc7\u4e00\u4e2a\u6269\u6563\u6a21\u578b\u6865\u63a5\u4e86\u6587\u672c\u97f5\u5f8b\u548c\u8bed\u97f3\u98ce\u683c\u4e4b\u95f4\u7684\u8054\u7cfb\uff0c\u6446\u8131\u63a8\u65ad\u65f6\u5bf9\u53c2\u8003\u97f3\u9891\u7684\u4f9d\u8d56\u3002\u4e0d\u7528\u53c2\u8003\u97f3\u9891\u5176\u5b9e\u5bf9\u4ea7\u54c1\u7684\u610f\u4e49\u8fd8\u633a\u5927\u7684\uff0c\u8981\u90fd\u7528\u73b0\u5b9e\u4e16\u754c\u4e2d\u771f\u4eba\u5c24\u5176\u662f\u540d\u4eba\u7684\u58f0\u97f3\u4f5c\u4e3a\u53c2\u8003\u97f3\u9891\uff0c\u90a3\u8fd9\u52bf\u5fc5\u4f1a\u5f15\u8d77\u7248\u6743\u7ea0\u7eb7\u3002\u8fd9\u79cd\u7ea0\u7eb7\u5728\u56fd\u5185\u56fd\u5916\u90fd\u6709\u76f8\u5173\u7684\u4e8b\u4ef6\u3002\u6700\u8fd1\u5be1\u59d0\u6295\u8bc9OpenAI\u7684\u4e8b\u4ef6\u5c31\u662f\u4e00\u4f8b\u3002<\/p>\n\n\n\n

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

        \u56fe23\uff1aStyleTTS 1\u7684\u6a21\u578b\u67b6\u6784<\/strong><\/p>\n\n\n\n

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

        \u56fe24\uff1aStyleTTS 2\u7684\u6a21\u578b\u67b6\u6784<\/strong><\/p>\n\n\n\n

        TTS\u5bf9\u6307\u4ee4\u7684\u9075\u5faa<\/strong><\/p>\n\n\n\n

        SLM\u4e0d\u4ec5\u8981\u5408\u6210\u5408\u4e4e\u4e0a\u4e0b\u6587\u8bed\u4e49\u7684\u9ad8\u8868\u73b0\u529b\u8bed\u97f3\uff0c\u5408\u6210\u7684\u8bed\u97f3\u8fd8\u8981\u7b26\u5408\u7528\u6237\u7684\u5373\u65f6\u8981\u6c42\u3002\u4e00\u4e9btext-guided zero-shot TTS\u7684\u5de5\u4f5c\u503c\u5f97\u53c2\u8003\u3002\u8fd9\u4e9b\u5de5\u4f5c\u4e00\u822c\u90fd\u662f\u5728\u5df2\u6709\u7684zero-shot TTS\u6a21\u578b\u6216\u8005text-to-audio\u6a21\u578b\u4e0a\u6539\u9020\u800c\u6765\uff0c\u540c\u65f6\u5438\u6536transcription\u548cdescription\u4e24\u8def\u6761\u4ef6\u3002\u5176\u4e2d\u7684\u91cd\u70b9\u8fd8\u662f\u5728\u4e8e\u6570\u636e\u96c6\u7684\u6784\u5efa\u3002\u8fd9\u65b9\u9762\u7684\u5de5\u4f5c\u6709\uff1aPromptTTS[33]\u3001InstructTTS[34]\u3001ParlerTTS[19]\u3001VoiceLDM[18]\u548cAudiobox[35]\u3002\u6211\u4eec\u4e3b\u8981\u8c08\u8c08ParlerTTS\u548cVoiceLDM\u3002<\/p>\n\n\n\n

        ParlerTTS[19]\u3002VALL-E\/VoiceCraft\u7684\u589e\u5f3a\u7248\uff0c\u901a\u8fc7T5\u7f16\u7801\u5668\u548ccross-attention\u65c1\u8def\u5f15\u5165\u4e86\u63cf\u8ff0\u6027\u6587\u672c\u7684\u4fe1\u606f\u3002\u8be5\u5de5\u4f5c\u7684\u76ee\u7684\u662f\u60f3\u4f7f\u7528\u81ea\u7136\u8bed\u8a00prompt\u6765\u6307\u5b9a\u8bf4\u8bdd\u98ce\u683c\u548c\u73af\u5883\u4fe1\u606f\uff0c\u6446\u8131\u5bf9\u53c2\u8003\u97f3\u9891\u7684\u4f9d\u8d56\u3002\u63cf\u8ff0\u6027\u6807\u7b7e\u6587\u672c\u7684\u6536\u96c6\u8fc7\u7a0b\u4e5f\u663e\u5f97\u76f8\u5f53\u6734\u7d20\uff1a\u901a\u8fc7\u5b9a\u5236\u5316\u7684\u76d1\u7763\u5f0f\u6a21\u578b\u83b7\u53d6\u8bed\u97f3\u6570\u636e\u7684\u53e3\u97f3\u7279\u5f81\u3001\u5f55\u97f3\u8d28\u91cf\u7279\u5f81\u3001\u97f3\u9ad8\u8bed\u901f\u7279\u5f81\u3002\u7136\u540e\u7528LLM\u5c06\u8fd9\u4e9b\u7279\u5f81\u8f6c\u6362\u4e3a\u81ea\u7136\u8bed\u8a00\u7684\u63cf\u8ff0\u3002\u5728\u6211\u770b\u6765\uff0c\u8fd9\u4e2a\u5de5\u4f5c\u6709\u8fd9\u4e48\u51e0\u70b9\u5c40\u9650\u6027\u5427\uff1a\u5176\u4e00\uff0c\u7f3a\u4e4f\u60c5\u7eea\u6807\u7b7e\uff1b\u5176\u4e8c\uff0c\u8bed\u97f3\u63cf\u8ff0\u6027\u6807\u7b7e\u7684\u6536\u96c6\u5e76\u4e0d\u5177\u5907\u901a\u7528\u6027\uff0c\u8f83\u4e3a\u7e41\u7410\uff0c\u8fdc\u4e0d\u5982\u4e00\u4e2a\u5f3a\u5927\u7684\u591a\u6a21\u6001\u8bed\u97f3\u7406\u89e3\u6a21\u578b\u6765\u5f97\u5b9e\u5728\u3002\u6587\u7ae0demo\u867d\u7136\u8fbe\u5230\u4e86\u9884\u671f\u7684\u6548\u679c\uff0c\u4f46\u573a\u666f\u4f3c\u4e4e\u5c40\u9650\u5728\u6717\u8bfb\u7684\u60c5\u666f\u4e2d\u3002<\/p>\n\n\n\n

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

        \u56fe25\uff1aParlerTTS\u7684\u6a21\u578b\u67b6\u6784<\/strong><\/p>\n\n\n\n

        VoiceLDM[18]\u3002\u5728VoiceLDM1\u7684\u57fa\u7840\u4e0a\u589e\u52a0\u4e86\u8f6c\u5f55\u6587\u672c\u7684\u8f93\u5165\u3002\u8fd9\u4e2a\u5de5\u4f5c\u548cAudioLDM 1\u5f88\u50cf\uff0c\u540c\u6837\u4f7f\u7528CLAP\u6ce8\u5165\u8bed\u97f3\u7684\u63cf\u8ff0\u6027\u4fe1\u606f\u3002\u4e0d\u540c\u5730\u662f\uff0c\u4e3a\u4e86\u505aTTS\u4efb\u52a1\uff0c\u8be5\u5de5\u4f5c\u901a\u8fc7cross-attention\u65c1\u8def\u589e\u52a0\u4e86transcription\u7684\u4fe1\u606f\u3002<\/p>\n\n\n\n

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

        \u56fe26\uff1aVoiceLDM\u7684\u6a21\u578b\u67b6\u6784<\/strong><\/p>\n\n\n\n

        TTS\u603b\u7ed3<\/strong><\/p>\n\n\n\n

        \u6797\u6797\u603b\u603b\u8bf4\u4e86\u8fd9\u4e48\u591azero-shot\u7684TTS\u65b9\u6cd5\uff0c\u6211\u60f3\u8bf4\u660e\u7684\u7ed3\u8bba\u6709\u8fd9\u4e48\u51e0\u70b9\uff1a<\/p>\n\n\n\n

        1. \u5728LLM\u5927\u884c\u5176\u9053\u3001scaling law\u5927\u663e\u795e\u5a01\u7684\u65f6\u4ee3\uff0cTTS\u6a21\u578b\u7684\u8bad\u7ec3\u6570\u636e\u89c4\u6a21\u5df2\u7ecf\u7a81\u7834\u4e86\u4e07\u5c0f\u65f6\uff0c\u751a\u81f3\u8fbe\u5230\u4e86\u6570\u5341\u4e07\u5c0f\u65f6\u7684\u7ea7\u522b\u3002\u5728\u5927\u6570\u636e\u7684\u52a0\u6301\u4e0b\uff0cTTS\u4efb\u52a1\u4e0a\u4e5f\u6d8c\u73b0\u51fa\u4e86in-context learning\u80fd\u529b\u3002<\/li>
        2. \u8bed\u97f3\u4fe1\u606f\u7684\u89e3\u7801\u901a\u5e38\u90fd\u8981\u5c42\u6b21\u5316\u6216\u8005\u591a\u6b65\u8fdb\u884c\uff0c\u4e0d\u80fd\u4e00\u6b65\u5230\u4f4d\u3002\u81ea\u56de\u5f52\u3001\u6269\u6563\u6a21\u578b\u548c\u6d41\u5339\u914d\u90fd\u80fd\u5728TTS\u4e2d\u53d1\u6325\u4f5c\u7528\uff1b<\/li>
        3. \u501f\u9274NLP instruction fine-tuning\u548c\u6587\u751f\u56fe\u7684\u7ecf\u9a8c\uff0cTTS\u6a21\u578b\u540c\u6837\u53ef\u4ee5\u9075\u5faa\u6587\u672c\u6307\u4ee4\u6216\u8005\u8bed\u97f3\u6307\u4ee4\uff0c\u5408\u6210\u7b26\u5408\u7528\u6237\u5373\u65f6\u8981\u6c42\u7684\u8bed\u97f3\uff0c\u6446\u8131\u5bf9\u53c2\u8003\u97f3\u9891\u7684\u4f9d\u8d56\uff0c\u8fd9\u6216\u8bb8\u4e5f\u80fd\u89c4\u907f\u4e00\u4e9b\u77e5\u8bc6\u4ea7\u6743\u7684\u56f0\u6270(\u6bd4\u5982\u6700\u8fd1\u6709\u540d\u7684\u5be1\u59d0\u6295\u8bc9OpenAI\u4e8b\u4ef6)\u3002\u540c\u65f6\uff0c\u7528\u6237\u4e5f\u80fd\u5728\u5bf9\u8bdd\u8fc7\u7a0b\u4e2d\u968f\u65f6\u5207\u6362\u8bed\u97f3\u56de\u590d\u7684\u98ce\u683c\uff0c\u8fd9\u4e00\u70b9\u5728OpenAI\u7684demo\u4e2d\u6709\u5f88\u660e\u786e\u7684\u4f53\u73b0\u3002\u53e6\u5916\uff0c\u4e0d\u77e5\u9053\u5927\u5bb6\u6709\u6ca1\u6709\u6ce8\u610f\uff0cGPT-4o\u5408\u6210\u7684\u8bed\u97f3\u662f\u53ef\u4ee5\u662f\u653e\u6620\u6240\u5904\u7684\u58f0\u5b66\u73af\u5883\u7684\uff1a\u6709\u4e00\u6bb5\u8bed\u97f3\u80cc\u540e\u4f3c\u4e4e\u662f\u6709\u94a2\u7434\u58f0\u7684\u3002<\/li>
        4. text-guided zero-shot TTS\u5728\u6a21\u578b\u67b6\u6784\u4e0a\u548czero-shot TTS\u6709\u975e\u5e38\u5927\u7684\u76f8\u4f3c\u6027\u3002\u4f46\u8bad\u7ec3\u6570\u636e\u53ef\u80fd\u8f83\u4e3a\u7f3a\u4e4f\u3002\u5148\u5f00\u53d1zero-shot TTS\uff0c\u518d\u7528\u7c7b\u4f3cSALMONN\u90a3\u6837\u7684\u591a\u6a21\u6001\u7406\u89e3\u6a21\u578b\u6765\u6253\u6807\u7b7e(\u7c7b\u4f3cDALLE3\u7684\u505a\u6cd5)\uff0c\u8fd9\u6837\u6570\u636e\u96c6\u6784\u9020\u65b9\u5f0f\uff0c\u53ef\u80fd\u4f1a\u662f\u66f4\u597d\u7684\u9009\u62e9\u3002<\/li><\/ol>\n\n\n\n

          \u53e6\u5916\uff0c\u5bf9\u4e8e\u8bed\u97f3\u7684\u89e3\u7801\u65b9\u6848\uff0c\u6211\u503e\u5411\u4e8e\u662f\u8fd9\u6837\u7684\uff1a<\/p>\n\n\n\n

          1. \u5982\u679c\u8981\u505a\u6d41\u5f0f\u63a8\u7406\uff0c\u5916\u63a5\u7c7b\u4f3cHIFIGAN\u8fd9\u6837\u7684\u58f0\u7801\u5668\u7684\u65b9\u5f0f\u53ef\u80fd\u4e0d\u662f\u597d\u7684\u9009\u62e9\u3002HIFIGAN\u5e76\u4e0d\u5929\u7136\u652f\u6301\u6d41\u5f0f\u89e3\u7801\u3002\u76f8\u53cd\u5730\uff0c\u8bf8\u5982SoundStream\u548cEncodec\u8fd9\u6837\u7684\u65b9\u6cd5\uff0c\u540c\u65f6\u6709\u6d41\u5f0f\u53d8\u4f53\u548c\u975e\u6d41\u5f0f\u53d8\u4f53\uff1b<\/li>
          2. \u5148\u505a\u8bed\u4e49token\u7684\u89e3\u7801\uff0c\u8fd9\u4e2a\u89e3\u7801\u5927\u6982\u7387\u662f\u81ea\u56de\u5f52\u89e3\u7801\u3002\u8bed\u4e49token\u6bd5\u7adf\u662f\u5efa\u6a21\u4e0a\u4e0b\u6587\u4f9d\u8d56\u5173\u7cfb\uff0c\u81ea\u56de\u5f52\u65b9\u6cd5\u5df2\u7ecf\u5728NLP\u4e0a\u8bc1\u660e\u4e86\u8fd9\u4e00\u70b9\uff1b<\/li>
          3. \u7136\u540e\u505a\u58f0\u5b66token\u7684\u89e3\u7801\uff0c\u6269\u6563\u6216\u8005flow-matching\u53ef\u80fd\u662f\u66f4\u597d\u7684\u9009\u62e9\u3002\u6269\u6563\u6a21\u578b\u6216\u8005\u6d41\u5339\u914d\u53ef\u4ee5\u5f88\u597d\u5730\u4fee\u8865\u8bed\u97f3\u7684\u7ec6\u8282\uff1b<\/li><\/ol>\n\n\n\n

            \u5f53\u7136\uff0c\u9664\u4e86\u4e0a\u9762\u8bb2\u5230\u7684\uff0czero-shot TTS\u8fd8\u6709\u5f88\u591a\u503c\u5f97\u7814\u7a76\u7684\u65b9\u6cd5\u3002\u9650\u4e8e\u7bc7\u5e45\uff0c\u4ec5\u5217\u4e3e\u4e8e\u6b64\uff0c\u4e0d\u518d\u8be6\u8ff0\uff1aHierSpeech++[36]\u3001base TTS[30]\u3001Voicebox\/Audiobox[35]\u3001UniAudio[37]\u3001Make-a-Voice[38]\u7b49\u7b49\u3002<\/p>\n\n\n\n

            \u5176\u4ed6\u95ee\u9898<\/strong><\/h2>\n\n\n\n

            \u5bf9\u4e8eGPT-4o\u6a21\u578b\uff0c\u5982\u679c\u4ec5\u4ec5\u805a\u7126\u4e8e\u8bed\u97f3\u591a\u6a21\u6001\uff0c\u8fd8\u6709\u4e0b\u9762\u7684\u95ee\u9898\u503c\u5f97\u5173\u6ce8\uff1a<\/p>\n\n\n\n

            1. \u8bed\u97f3\u4ea4\u4e92\u5982\u4f55\u505a\u5230\u4f4e\u5ef6\u8fdf\uff1f\u5927\u6982\u7387\u8981\u6c42\u6d41\u5f0f\u5207\u7247\u5904\u7406\uff0c\u4e3b\u8981\u5de5\u4f5c\u5728\u4e8e\u5de5\u7a0b\u4f18\u5316\uff0c\u7528C++\u91cd\u5199\u7b97\u5b50\u3002\u63a8\u7406\u6846\u67b6\u7684\u8bdd\uff0c\u7528tensorrt\u3001mnn\u8fd9\u4e9b\u90fd\u884c\u3002\u4e0a\u4e0b\u6587\u6240\u8ff0\u7684\u97f3\u9891\u79bb\u6563\u5316\u65b9\u6cd5\uff0c\u8bf8\u5982SoundStream\u548cEncodec\uff0c\u5176\u5b9e\u4e5f\u652f\u6301\u6d41\u5f0f\u5904\u7406\u3002<\/li>
            2. \u8bed\u97f3\u5bf9\u8bdd\u4e2d\u7684\u6253\u65ad\u5982\u4f55\u5b9e\u73b0\uff1f\u4e2a\u4eba\u8ba4\u4e3a\u6709\u4e24\u79cd\u53ef\u80fd\u7684\u65b9\u6848\uff1aturn-based\u548c\u6d41\u5f0f\u5904\u7406\u3002\u6240\u8c13\u7684turn-based\u65b9\u6848\uff0c\u662f\u6bd4\u8f83\u5de5\u7a0b\u5316\u7684\uff0c\u7b80\u7b54\u6982\u62ec\u4e00\u4e0b\u5c31\u662f\uff1a\u68c0\u6d4b\u662f\u5426\u6709\u505c\u987f\uff0c\u5982\u679c\u4e00\u6bb5\u65f6\u95f4\u5185\u6ca1\u6709\u58f0\u97f3\uff0c\u6a21\u578b\u5c31\u5f00\u59cb\u8fd4\u56de\u8bed\u97f3\u56de\u590d\u3002\u53e6\u4e00\u79cd\u6d41\u5f0f\u65b9\u6848\uff0c\u5219\u662f\uff1a\u6a21\u578b\u4e00\u76f4\u5728\u63a5\u53d7\u7528\u6237\u7684\u6d41\u5f0f\u8bed\u97f3\u8f93\u5165\uff0c\u5224\u65ad\u662f\u5426\u5e94\u8be5\u8f93\u51fa\u8bed\u97f3\u56de\u590d\uff0c\u4e00\u4e2a\u5145\u5206\u8bad\u7ec3\u7684\u6a21\u578b\u5e94\u8be5\u662f\u80fd\u591f\u51c6\u786e\u9884\u6d4b\u51fa\u8bed\u97f3\u8bcd\u8868\u4e2d\u7684[START]\u548c[END]\u7684\u3002<\/li><\/ol>\n\n\n\n

              \u5bf9\u6e38\u620f\u914d\u97f3\u4e1a\u52a1\u7684\u601d\u8003<\/strong><\/h2>\n\n\n\n

              text\/prompt-guided zero-shot TTS\u65b9\u6cd5\u5bf9\u6e38\u620f\u7684AI\u914d\u97f3\u610f\u4e49\u91cd\u5927\u3002\u4e3b\u8981\u4f53\u73b0\u5728\uff1a<\/p>\n\n\n\n

              1. \u7528\u81ea\u7136\u8bed\u8a00\u63d0\u793a\u53bb\u5408\u6210\u97f3\u8272\u7a33\u5b9a\u7684\u8bed\u97f3\uff0c\u6446\u8131\u5bf9\u53c2\u8003\u97f3\u9891\u7684\u4f9d\u8d56\uff0c\u5728\u4e1a\u52a1\u4e2d\u80fd\u591f\u66f4\u52a0\u7075\u6d3b\uff0c\u81f3\u5c11\u6bd4\u514b\u9686\u5df2\u6709\u4eba\u7269\/\u89d2\u8272\u7684\u8bed\u97f3\u7684\u65b9\u5f0f\u66f4\u52a0\u65b9\u4fbf\uff0c\u66f4\u4e0d\u5bb9\u6613\u51fa\u620f\u3002\u4e3e\u4e2a\u4f8b\u5b50\uff0c\u5728\u5f00\u653e\u4e16\u754c\u5267\u60c5\u7c7b\u6e38\u620f\u7684\u7814\u53d1\u9636\u6bb5\uff0c\u6211\u4eec\u4f1a\u8bbe\u5b9a\u4e00\u4e9bprofile\u8d4b\u4e88NPC\uff0c\u8ba9\u73a9\u5bb6\u8ddfNPC\u804a\u5929\u3002\u6211\u4eec\u66fe\u7ecf\u7528\u514b\u9686\u300a\u539f\u795e\u300b\u3001\u300a\u5d29\u574f\uff1a\u661f\u7a79\u94c1\u9053\u300b\u5df2\u6709\u89d2\u8272\u7684\u65b9\u5f0f\u8d4b\u4e88\u8fd9\u4e9bNPC\u89d2\u8272\u8bed\u97f3\uff0c\u4f46\u653e\u5728\u90a3\u4e9b\u6b27\u7f8e\u80cc\u666f\u7684NPC\u4e2d\uff0c\u5c31\u662f\u5f88\u6709\u8fdd\u548c\u611f\uff0c\u6ca1\u6709\u73b0\u5b9e\u4e16\u754c\u4e2d\u7684accent\uff0c\u4e0d\u591fdecent\u3002<\/li>
              2. \u5267\u60c5\u4efb\u52a1\u4e2d\u7684\u914d\u97f3\u4f1a\u66f4\u52a0\u771f\u4eba\u5316\u3001\u66f4\u6709\u6c89\u6d78\u611f\u3002\u8fc7\u5e74\u671f\u95f4\u8fc7\u300a\u5d29\u574f\uff1a\u661f\u7a79\u94c1\u9053\u300b\u82b1\u706b\u548c\u9ed1\u5929\u9e45\u7684\u540c\u884c\u4efb\u52a1\u7684\u65f6\u5019\uff0c\u90e8\u5206NPC\u89d2\u8272\u4f1a\u6709\u516d\u516c\u4e3b\u7684\u7ffb\u8bd1\u8154\uff0c\u8fd9\u662f\u82b1\u706b\u884c\u4e8e\u6b22\u6109\u547d\u9014\u7684\u6076\u8da3\u5473\uff0c\u7a7a\u6c14\u4e2d\u987f\u65f6\u5145\u6ee1\u4e86\u5feb\u6d3b\u7684\u5473\u9053\u3002\u5982\u679c\u8d70bv2\u3001gsv\u7684\u8bed\u97f3\u514b\u9686\u65b9\u6848\uff0c\u5e94\u8be5\u662f\u5f88\u96be\u6709\u8fd9\u79cd\u6548\u679c\u7684\u3002\u800c\u4e14\uff0c\u73a9\u5bb6\u5728\u5267\u60c5\u4efb\u52a1\u4e2d\u52bf\u5fc5\u4f1a\u7ecf\u8fc7\u4e0d\u540c\u7684\u5730\u52bf\u5730\u8c8c\uff0c\u81f3\u5c11\u5ba4\u5185\u3001\u5ba4\u5916\u7684\u58f0\u97f3\u542c\u8d77\u6765\u662f\u6709\u4e0d\u540c\u7684\u3002\u5ba4\u5185\u7684\u58f0\u97f3\u81f3\u5c11\u4f1a\u6709\u56de\u54cd\u3001\u6df7\u54cd\u7684\u5427\u3002\u8fd9\u79cd\u611f\u89c9\u8bed\u97f3\u514b\u9686\u65b9\u6848\u4e5f\u662f\u65e0\u6cd5\u505a\u5230\u7684\u3002<\/li><\/ol>\n\n\n\n
                \"\"\/<\/figure>\n\n\n\n

                \u5168\u6587\u603b\u7ed3<\/strong><\/h2>\n\n\n\n

                \u603b\u7ed3\u4e00\u4e0b\u672c\u6587\u8bf4\u8c08\u7684\u5185\u5bb9\uff0c\u6211\u8ba4\u4e3aGPT-4o\u8bed\u97f3\u591a\u6a21\u6001\u7684\u5b9e\u73b0\u53ef\u80fd\u662f\u8d70\u4e86\u4ee5\u4e0b\u7684\u6280\u672f\u8def\u7ebf\uff1a<\/p>\n\n\n\n

                1. audio & text tokenizer\u7684\u5b9e\u73b0\u5e94\u8be5\u662f\u8bed\u97f3\u79bb\u6563\u5316\u90e8\u5206\u6240\u7528\u7684\u6280\u672f\uff0c\u4f8b\u5982SoundStream\u3001Encodec\u3001SpeechTokenizer\uff0c\u6216\u8005\u662fMEL+VQ\u6700\u540e\u914d\u5408\u58f0\u7801\u5668\u6765\u89e3\u7801\uff1b\u53c2\u8003zero-shot TTS\u3001AudioLM\/AudioPaLM\u3001SpeechGPT-Gen\u7b49\u5de5\u4f5c\u7684\u7ed3\u679c\uff0cLLM\u4e2d\u8bed\u97f3token\u7684\u89e3\u7801\u5e94\u8be5\u662f\u8981\u8d70\u5c42\u6b21\u5316\u6216\u8005\u591a\u6b65\u7684\u65b9\u6cd5\uff0c\u5148\u89e3\u7801\u8bed\u4e49\u7279\u5f81\uff0c\u518d\u89e3\u7801\u58f0\u5b66\u7279\u5f81\uff0c\u6216\u8005\u662f\u5148\u89e3\u7801MEL\uff0c\u518d\u52a0\u4e00\u4e2aHIFIGAN\u8fd9\u6837\u7684\u58f0\u7801\u5668\u3002\u53e6\u5916\uff0c\u5982\u679c\u505aaudio\/speech\/music\u8fd9\u6837\u7684\u901a\u7528\u58f0\u5408\u6210\u7684\u8bdd\uff0c\u53ef\u80fd\u4e5f\u80fd\u901a\u8fc7prompt\u6765\u63a7\u5236\u3002AudioLDM2\u867d\u7136\u505a\u4e86\u8fd9\u65b9\u9762\u7684\u5de5\u4f5c\uff0c\u4f46audio\/music\u548cspeech\u7684\u53c2\u6570\u5176\u5b9e\u662f\u4e0d\u4e00\u6837\u7684\uff0c\u8bf4\u5230\u5e95\u8fd8\u4e0d\u662f\u540c\u4e00\u4e2a\u6a21\u578b\u3002<\/li>
                2. \u5bf9\u4e8e\u6307\u4ee4\u5fae\u8c03\uff0c\u6570\u636e\u96c6\u7684\u6784\u9020\u975e\u5e38\u91cd\u8981\uff0c\u5927\u6982\u7387\u8981\u7528\u5230\u5408\u6210\u6570\u636e\u3002\u5176\u4e00\uff0c\u7f51\u7edc\u4e0a\u9ad8\u8d28\u91cf\u8bed\u97f3\u6570\u636e\u7684\u91cf\u7ea7\u8fdc\u8fdc\u4e0d\u53ca\u6587\u672c\uff0c\u76f4\u63a5\u62ffASR\u6570\u636e\u6765\u505a\u80af\u5b9a\u4f1a\u5f71\u54cd\u6a21\u578b\u5408\u6210\u8bed\u97f3\u7684\u97f3\u8d28\uff1b\u5176\u4e8c\uff0c\u5927\u8bed\u8a00\u6a21\u578b\u5408\u6210\u7684instruction\u5f80\u5f80\u89e6\u53ca\u4e0d\u5230\u8bed\u97f3\u7684\u7ec6\u7c92\u5ea6\u7279\u5f81\uff0c\u8fd9\u6837\u7684instruction\u5176\u5b9e\u65e0\u6cd5\u51c6\u786e\u8be6\u5c3d\u5730\u63cf\u8ff0text\u548cspeech\u4e4b\u95f4\u7684\u5173\u7cfb\u3002\u56e0\u800c\uff0c\u9700\u8981\u5f15\u5165\u5f3a\u5927\u7684zero-shot TTS\u6a21\u578b\u5408\u6210\u9ad8\u8d28\u91cf\u8bed\u97f3\uff0c\u7136\u540e\u7528\u591a\u6a21\u6001\u8bed\u97f3\u7406\u89e3\u6a21\u578b\u6765\u4e3a\u5408\u6210\u8bed\u97f3\u6253\u6807\u7b7e\uff0c\u5f53\u7136\u4e5f\u53ef\u4ee5\u8bc4\u5206\u505a\u7b5b\u9009\u4ec0\u4e48\u7684\u3002<\/li>
                3. \u6700\u540e\u662f\u8981\u8ba9\u5927\u6a21\u578b\u7684\u8f93\u51fa\u5bf9\u9f50\u4eba\u7c7b\u7684\u504f\u597d\u3002\u8fd9\u65b9\u9762\u7684\u65b9\u6cd5\u6709\u5f88\u591a\uff0c\u6709DPO\u3001PPO\u4ec0\u4e48\u7684\uff0c\u90fd\u53ef\u4ee5\u7528\u3002<\/li><\/ol>\n\n\n\n
                  \"\"\/<\/figure>\n\n\n\n

                  \u56fe27\uff1a\u5168\u6587\u603b\u7ed3\uff0c\u53ef\u80fd\u7684roadmap<\/strong><\/p>\n\n\n\n

                  \u53c2\u8003\u6587\u732e<\/strong><\/h2>\n\n\n\n

                  [1] Baevski A, Zhou Y, Mohamed A, et al. wav2vec 2.0: A framework for self-supervised learning of speech representations[J]. Advances in neural information processing systems, 2020, 33: 12449-12460.<\/em><\/p>\n\n\n\n

                  [2] Hsu W N, Bolte B, Tsai Y H H, et al. Hubert: Self-supervised speech representation learning by masked prediction of hidden units[J]. IEEE\/ACM Transactions on Audio, Speech, and Language Processing, 2021, 29: 3451-3460.<\/em><\/p>\n\n\n\n

                  [3] Chung Y A, Zhang Y, Han W, et al. W2v-bert: Combining contrastive learning and masked language modeling for self-supervised speech pre-training[C]\/\/2021 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU). IEEE, 2021: 244-250.<\/em><\/p>\n\n\n\n

                  [4] Van Den Oord A, Vinyals O. Neural discrete representation learning[J]. Advances in neural information processing systems, 2017, 30.<\/em><\/p>\n\n\n\n

                  [5] Zeghidour N, Luebs A, Omran A, et al. Soundstream: An end-to-end neural audio codec[J]. IEEE\/ACM Transactions on Audio, Speech, and Language Processing, 2021, 30: 495-507.<\/em><\/p>\n\n\n\n

                  [6] D\u00e9fossez A, Copet J, Synnaeve G, et al. High fidelity neural audio compression[J]. arXiv preprint arXiv:2210.13438, 2022.<\/em><\/p>\n\n\n\n

                  [7] Zhang X, Zhang D, Li S, et al. Speechtokenizer: Unified speech tokenizer for speech large language models[J]. arXiv preprint arXiv:2308.16692, 2023.<\/em><\/p>\n\n\n\n

                  [8] Borsos Z, Marinier R, Vincent D, et al. Audiolm: a language modeling approach to audio generation[J]. IEEE\/ACM Transactions on Audio, Speech, and Language Processing, 2023.<\/em><\/p>\n\n\n\n

                  [9] Rubenstein P K, Asawaroengchai C, Nguyen D D, et al. Audiopalm: A large language model that can speak and listen[J]. arXiv preprint arXiv:2306.12925, 2023.<\/em><\/p>\n\n\n\n

                  [10] Changli Tang, Wenyi Yu, Guangzhi Sun, Xianzhao Chen, Tian Tan, Wei Li, Lu Lu, Zejun Ma, Chao Zhang. SALMONN: Towards Generic Hearing Abilities for Large Language Models<\/em><\/p>\n\n\n\n

                  [11] Zhang D, Li S, Zhang X, et al. Speechgpt: Empowering large language models with intrinsic cross-modal conversational abilities[J]. arXiv preprint arXiv:2305.11000, 2023.<\/em><\/p>\n\n\n\n

                  [12] Zhang D, Zhang X, Zhan J, et al. SpeechGPT-Gen: Scaling Chain-of-Information Speech Generation[J]. arXiv preprint arXiv:2401.13527, 2024.<\/em><\/p>\n\n\n\n

                  [13] Zhang D, Li Z, Li S, et al. SpeechAlign: Aligning Speech Generation to Human Preferences[J]. arXiv preprint arXiv:2404.05600, 2024.<\/em><\/p>\n\n\n\n

                  [14] Chen Q, Chu Y, Gao Z, et al. Lauragpt: Listen, attend, understand, and regenerate audio with gpt[J]. arXiv preprint arXiv:2310.04673, 2023.<\/em><\/p>\n\n\n\n

                  [15] Wu S, Fei H, Qu L, et al. Next-gpt: Any-to-any multimodal llm[J]. arXiv preprint arXiv:2309.05519, 2023.<\/em><\/p>\n\n\n\n

                  [16] Wang C, Chen S, Wu Y, et al. Neural codec language models are zero-shot text to speech synthesizers[J]. arXiv preprint arXiv:2301.02111, 2023.<\/em><\/p>\n\n\n\n

                  [17] Anil R, Dai A M, Firat O, et al. Palm 2 technical report[J]. arXiv preprint arXiv:2305.10403, 2023.<\/em><\/p>\n\n\n\n

                  [18] Lee Y, Yeon I, Nam J, et al. VoiceLDM: Text-to-Speech with Environmental Context[C]\/\/ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024: 12566-12571.<\/em><\/p>\n\n\n\n

                  [19] Lyth D, King S. Natural language guidance of high-fidelity text-to-speech with synthetic annotations[J]. arXiv preprint arXiv:2402.01912, 2024.<\/em><\/p>\n\n\n\n

                  [20] Betker J. Better speech synthesis through scaling[J]. arXiv preprint arXiv:2305.07243, 2023.<\/em><\/p>\n\n\n\n

                  [21] Xin D, Tan X, Shen K, et al. RALL-E: Robust Codec Language Modeling with Chain-of-Thought Prompting for Text-to-Speech Synthesis[J]. arXiv preprint arXiv:2404.03204, 2024.<\/em><\/p>\n\n\n\n

                  [22] Wang C, Zeng C, Zhang B, et al. HAM-TTS: Hierarchical Acoustic Modeling for Token-Based Zero-Shot Text-to-Speech with Model and Data Scaling[J]. arXiv preprint arXiv:2403.05989, 2024.<\/em><\/p>\n\n\n\n

                  [23] Ren Y, Hu C, Tan X, et al. Fastspeech 2: Fast and high-quality end-to-end text to speech[J]. arXiv preprint arXiv:2006.04558, 2020.<\/em><\/p>\n\n\n\n

                  [24] Rombach R, Blattmann A, Lorenz D, et al. High-resolution image synthesis with latent diffusion models[C]\/\/Proceedings of the IEEE\/CVF conference on computer vision and pattern recognition. 2022: 10684-10695.<\/em><\/p>\n\n\n\n

                  [25] Shen K, Ju Z, Tan X, et al. Naturalspeech 2: Latent diffusion models are natural and zero-shot speech and singing synthesizers[J]. arXiv preprint arXiv:2304.09116, 2023.<\/em><\/p>\n\n\n\n

                  [26] Ju Z, Wang Y, Shen K, et al. NaturalSpeech 3: Zero-shot speech synthesis with factorized codec and diffusion models[J]. arXiv preprint arXiv:2403.03100, 2024.<\/em><\/p>\n\n\n\n

                  [27] Liu H, Tian Q, Yuan Y, et al. AudioLDM 2: Learning holistic audio generation with self-supervised pretraining[J]. arXiv preprint arXiv:2308.05734, 2023.<\/em><\/p>\n\n\n\n

                  [28] Jiang Z, Ren Y, Ye Z, et al. Mega-tts: Zero-shot text-to-speech at scale with intrinsic inductive bias[J]. arXiv preprint arXiv:2306.03509, 2023.<\/em><\/p>\n\n\n\n

                  [29] Jiang Z, Liu J, Ren Y, et al. Mega-tts 2: Zero-shot text-to-speech with arbitrary length speech prompts[J]. arXiv preprint arXiv:2307.07218, 2023.<\/em><\/p>\n\n\n\n

                  [30] \u0141ajszczak M, C\u00e1mbara G, Li Y, et al. BASE TTS: Lessons from building a billion-parameter text-to-speech model on 100K hours of data[J]. arXiv preprint arXiv:2402.08093, 2024.<\/em><\/p>\n\n\n\n

                  [31] Li Y A, Han C, Mesgarani N. Styletts: A style-based generative model for natural and diverse text-to-speech synthesis[J]. arXiv preprint arXiv:2205.15439, 2022.<\/em><\/p>\n\n\n\n

                  [32] Li Y A, Han C, Raghavan V, et al. Styletts 2: Towards human-level text-to-speech through style diffusion and adversarial training with large speech language models[J]. Advances in Neural Information Processing Systems, 2024, 36.<\/em><\/p>\n\n\n\n

                  [33] Guo Z, Leng Y, Wu Y, et al. Prompttts: Controllable text-to-speech with text descriptions[C]\/\/ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023: 1-5.<\/em><\/p>\n\n\n\n

                  [34] Yang D, Liu S, Huang R, et al. Instructtts: Modelling expressive TTS in discrete latent space with natural language style prompt[J]. arXiv preprint arXiv:2301.13662, 2023.<\/em><\/p>\n\n\n\n

                  [35] Vyas A, Shi B, Le M, et al. Audiobox: Unified audio generation with natural language prompts[J]. arXiv preprint arXiv:2312.15821, 2023.<\/em><\/p>\n\n\n\n

                  [36] Lee S H, Choi H Y, Kim S B, et al. HierSpeech++: Bridging the Gap between Semantic and Acoustic Representation of Speech by Hierarchical Variational Inference for Zero-shot Speech Synthesis[J]. arXiv preprint arXiv:2311.12454, 2023.<\/em><\/p>\n\n\n\n

                  [37] Yang D, Tian J, Tan X, et al. Uniaudio: An audio foundation model toward universal audio generation[J]. arXiv preprint arXiv:2310.00704, 2023.<\/em><\/p>\n\n\n\n

                  [38] Huang R, Zhang C, Wang Y, et al. Make-a-voice: Unified voice synthesis with discrete representation[J]. arXiv preprint arXiv:2305.19269, 2023.<\/em><\/p>\n","protected":false},"excerpt":{"rendered":"

                  5\u670814\u65e5\u51cc\u6668\uff0cOpenAI\u63a8\u51fa\u4e86\u6700\u65b0\u7684\u751f\u6210\u6a21\u578bGPT-4o\uff0c\u5e26\u6765\u4e86\u4e00\u7cfb\u5217\u9707\u64bc\u7684\u529f\u80fd\uff0c\u7528\u6280\u672f\u5f7b\u5e95\u98a0\u8986\u4e86\u4ea7\u54c1\u5f62\u6001 … \u7ee7\u7eed\u9605\u8bfbGPT-4o\u80cc\u540e\u7684\u8bed\u97f3\u6280\u672f<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[21,40,4,9,38,34],"tags":[],"_links":{"self":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/15743"}],"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=15743"}],"version-history":[{"count":27,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/15743\/revisions"}],"predecessor-version":[{"id":19118,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/15743\/revisions\/19118"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=15743"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=15743"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=15743"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}