{"id":23119,"date":"2024-12-20T17:54:50","date_gmt":"2024-12-20T09:54:50","guid":{"rendered":"http:\/\/139.9.1.231\/?p=23119"},"modified":"2025-07-08T19:54:28","modified_gmt":"2025-07-08T11:54:28","slug":"cosyvoice-2-0","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2024\/12\/20\/cosyvoice-2-0\/","title":{"rendered":"CosyVoice-2.0 \u8bed\u97f3\u751f\u6210\u5927\u6a21\u578b\u5347\u7ea7"},"content":{"rendered":"\n<p class=\"has-text-align-center has-light-gray-background-color has-background\"><strong>CosyVoice 2: Scalable Streaming Speech Synthesis with Large Language Models<\/strong><\/p>\n\n\n\n<p class=\"has-text-align-center has-light-gray-background-color has-background\"><strong>[<a href=\"https:\/\/funaudiollm.github.io\/pdf\/CosyVoice_2.pdf\">Paper<\/a>] [<a href=\"https:\/\/github.com\/FunAudioLLM\/CosyVoice\">Code<\/a>] [<a href=\"https:\/\/www.modelscope.cn\/studios\/iic\/CosyVoice2-0.5B\">Studio<\/a>] [<a href=\"https:\/\/funaudiollm.github.io\/cosyvoice2\/\">HuggingFace<\/a>] [<a href=\"https:\/\/www.modelscope.cn\/models\/iic\/CosyVoice2-0.5B\">ModelScope<\/a>]<\/strong><\/p>\n\n\n\n\n\n<p><strong>CosyVoice<\/strong>\u662f\u963f\u91cc\u5df4\u5df4\u901a\u4e49\u5b9e\u9a8c\u5ba4\u8bed\u97f3\u56e2\u961f\u4e8e\u4eca\u5e747\u6708\u4efd\u5f00\u6e90\u7684\u8bed\u97f3\u751f\u6210\u5927\u6a21\u578b\uff0c\u4f9d\u6258\u5927\u6a21\u578b\u6280\u672f\uff0c\u5b9e\u73b0\u81ea\u7136\u6d41\u7545\u7684\u8bed\u97f3\u751f\u6210\u4f53\u9a8c\u3002\u4e0e\u4f20\u7edf\u8bed\u97f3\u751f\u6210\u6280\u672f\u76f8\u6bd4\uff0c<strong>CosyVoice<\/strong>\u5177\u6709\u97f5\u5f8b\u81ea\u7136\u3001\u97f3\u8272\u903c\u771f\u7b49\u7279\u70b9\u3002\u81ea\u5f00\u6e90\u4ee5\u6765\uff0c<strong>CosyVoice<\/strong>\u51ed\u501f\u9ad8\u54c1\u8d28\u7684\u591a\u8bed\u8a00\u8bed\u97f3\u751f\u6210\u3001\u96f6\u6837\u672c\u8bed\u97f3\u751f\u6210\u3001\u8de8\u8bed\u8a00\u8bed\u97f3\u751f\u6210\u3001\u5bcc\u6587\u672c\u548c\u81ea\u7136\u8bed\u8a00\u7684\u7ec6\u7c92\u5ea6\u63a7\u5236\u80fd\u529b\u83b7\u5f97\u4e86\u5e7f\u5927\u793e\u533a\u5f00\u53d1\u8005\u4eec\u7684\u559c\u7231\u548c\u652f\u6301\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"379\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-56-1024x379.png\" alt=\"\" class=\"wp-image-23176\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-56-1024x379.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-56-300x111.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-56-768x285.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-56.png 1298w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"457\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-58-1024x457.png\" alt=\"\" class=\"wp-image-23178\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-58-1024x457.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-58-300x134.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-58-768x342.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-58.png 1301w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"514\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-57-1024x514.png\" alt=\"\" class=\"wp-image-23177\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-57-1024x514.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-57-300x150.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-57-768x385.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-57.png 1290w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p><strong>CosyVoice\u8fce\u6765\u5168\u9762\u5347\u7ea7<\/strong>2.0\u7248\u672c\uff0c\u63d0\u4f9b\u66f4\u51c6\u3001\u66f4\u7a33\u3001\u66f4\u5feb\u3001 \u66f4\u597d\u7684\u8bed\u97f3\u751f\u6210\u80fd\u529b\u3002<\/p>\n\n\n\n<p><strong>\u8d85\u4f4e\u5ef6\u8fdf\uff1a<\/strong>CosyVoice 2.0<strong>\u63d0\u51fa\u4e86\u79bb\u7ebf\u548c\u6d41\u5f0f\u4e00\u4f53\u5316\u5efa\u6a21\u7684\u8bed\u97f3\u751f\u6210\u5927\u6a21\u578b\u6280\u672f<\/strong>\uff0c<strong>\u652f\u6301\u53cc\u5411\u6d41\u5f0f\u8bed\u97f3\u5408\u6210<\/strong>\uff0c\u5728\u57fa\u672c\u4e0d\u635f\u5931\u6548\u679c\u7684\u60c5\u51b5\u4e0b\u9996\u5305\u5408\u6210<strong>\u5ef6\u8fdf\u53ef\u4ee5\u8fbe\u5230150ms<\/strong>\u3002<\/p>\n\n\n\n<p><strong>\u9ad8\u51c6\u786e\u5ea6\uff1a<\/strong>CosyVoice 2.0\u5408\u6210\u97f3\u9891\u7684\u53d1\u97f3\u9519\u8bef\u76f8\u6bd4\u4e8eCosyVoice 1.0\u76f8\u5bf9\u4e0b\u964d30%\uff5e50%\uff0c\u5728Seed-TTS\u6d4b\u8bd5\u96c6\u7684hard\u6d4b\u8bd5\u96c6\u4e0a\u53d6\u5f97<strong>\u5f53\u524d\u6700\u4f4e\u7684\u5b57\u9519\u8bef\u7387<\/strong>\u3002<strong>\u5408\u6210\u7ed5\u53e3\u4ee4\u3001\u591a\u97f3\u5b57\u3001\u751f\u50fb\u5b57\u4e0a\u5177\u6709\u660e\u663e\u7684\u63d0\u5347\u3002<\/strong><\/p>\n\n\n\n<p><strong>\u5f3a\u7a33\u5b9a\u6027\uff1a<\/strong>CosyVoice 2.0\u5728\u96f6\u6837\u672c\u8bed\u97f3\u751f\u6210\u548c\u8de8\u8bed\u8a00\u8bed\u97f3\u5408\u6210\u4e0a\u80fd\u591f\u51fa\u8272\u5730<strong>\u4fdd\u8bc1\u97f3\u8272\u4e00\u81f4\u6027<\/strong>\uff0c\u7279\u522b\u662f<strong>\u8de8\u8bed\u8a00\u8bed\u97f3\u5408\u6210\u76f8\u6bd4\u4e8e1.0\u7248\u672c\u5177\u6709\u660e\u663e\u63d0\u5347<\/strong>\u3002<\/p>\n\n\n\n<p><strong>\u81ea\u7136\u4f53\u9a8c\uff1a<\/strong>CosyVoice 2.0\u5408\u6210\u97f3\u9891\u7684<strong>\u97f5\u5f8b\u3001\u97f3\u8d28\u3001\u60c5\u611f\u5339\u914d<\/strong>\u76f8\u6bd4\u4e8e1.0\u5177\u6709\u660e\u663e\u63d0\u5347\u3002MOS\u8bc4\u6d4b\u5206\u4ece5.4\u63d0\u5347\u52305.53(<strong>\u76f8\u540c\u8bc4\u6d4b\u67d0\u5546\u4e1a\u5316\u8bed\u97f3\u5408\u6210\u5927\u6a21\u578b\u4e3a5.52<\/strong>)\u3002\u540c\u65f6, CosyVoice 2.0\u5bf9\u4e8e\u6307\u4ee4\u53ef\u63a7\u7684\u97f3\u9891\u751f\u6210\u4e5f\u8fdb\u884c\u4e86\u5347\u7ea7\uff0c\u652f\u6301<strong>\u66f4\u591a\u7ec6\u7c92\u5ea6\u7684\u60c5\u611f\u63a7\u5236<\/strong>\uff0c\u4ee5\u53ca<strong>\u65b9\u8a00\u53e3\u97f3\u63a7\u5236<\/strong>\u3002<\/p>\n\n\n\n<h2><strong>\u258e\u6838\u5fc3\u6a21\u578b\u4e0e\u7b97\u6cd5\u4eae\u70b9<\/strong><\/h2>\n\n\n\n<p>CosyVoice 2.0\u91c7\u7528\u548cCosyVoice 1\u4e00\u81f4\u7684LLM+FM\u7684\u5efa\u6a21\u6846\u67b6\uff0c\u4f46\u662f\u5728\u5177\u4f53\u5b9e\u73b0\u4e0a\u8fdb\u884c\u4e86\u5982\u4e0b\u51e0\u4e2a\u8981\u70b9\u7684\u7b97\u6cd5\u4f18\u5316\uff1a<\/p>\n\n\n\n<p><strong>1\uff09LLM backbone<\/strong>\uff1aCosyVoice 2.0\u91c7\u7528\u9884\u8bad\u7ec3\u597d\u7684\u6587\u672c\u57fa\u5ea7\u5927\u6a21\u578b\uff08Qwen2.5-0.5B\uff09\u66ff\u6362\u4e86\u539f\u6765\u7684Text Encoder + random Transformer\u7684\u7ed3\u6784\u3002\u91c7\u7528LLM\u8fdb\u884c\u521d\u59cb\u5316\u80fd\u591f\u66f4\u597d\u7684\u8fdb\u884c\u6587\u672c\u7684\u8bed\u4e49\u5efa\u6a21\uff0c\u4f7f\u5f97\u5728\u53ef\u63a7\u751f\u6210\uff0c\u97f3\u9891\u548c\u6587\u672c\u7684\u60c5\u611f\u5339\u914d\uff0c\u591a\u97f3\u5b57\u53d1\u97f3\u4e0a\u4f1a\u6709\u660e\u663e\u7684\u6536\u76ca\u3002<\/p>\n\n\n\n<p><strong>2\uff09FSQ Speech Tokenizer<\/strong>\uff1aCosyVoice 1.0\u91c7\u7528VQ\u6765\u63d0\u53d6Supervised semantic codec\uff0c\u7801\u672c\u5927\u5c0f\u4e3a4096\uff0c\u4f46\u662f\u6709\u6548\u7801\u672c\u53ea\u6709963\u3002CosyVoice 2.0\u91c7\u7528\u4e86FSQ\u66ff\u6362VQ\uff0c\u8bad\u7ec3\u4e866561\u7684\u7801\u672c\uff0c\u5e76\u4e14\u7801\u672c100%\u6fc0\u6d3b\u3002FSQ-Speech Tokenizer\u7684\u4f7f\u7528\u4f7f\u5f97CosyVoice 2.0\u5728\u53d1\u97f3\u51c6\u786e\u6027\u4e0a\u6709\u660e\u663e\u63d0\u5347\u3002<\/p>\n\n\n\n<p><strong>3\uff09\u79bb\u7ebf\u548c\u6d41\u5f0f\u4e00\u4f53\u5316\u5efa\u6a21\u65b9\u6848<\/strong>\uff1a\u76ee\u524d\u4e3b\u6d41\u7684\u8bed\u97f3\u751f\u6210\u5927\u6a21\u578b(CosyVoice\uff0c F5-TTS\uff0cMaskGCT\uff0cGPT-SoViTs\u7b49)\u5747\u4e0d\u652f\u6301\u6d41\u5f0f\u8bed\u97f3\u751f\u6210\u3002CosyVoice 2.0\u63d0\u51fa\u4e86\u5982\u56fe2\u6240\u793a\u7684\u79bb\u7ebf\u548c\u6d41\u5f0f\u4e00\u4f53\u5316\u5efa\u6a21\u65b9\u6848\uff0c\u4f7f\u5f97LLM\u548cFM\u5747\u652f\u6301\u6d41\u5f0f\u63a8\u7406\uff0c\u63a5\u65365\u4e2a\u6587\u5b57\u5c31\u53ef\u4ee5\u5408\u6210\u9996\u5305\u97f3\u9891\uff0c\u5ef6\u8fdf\u5927\u81f4\u5728150ms\u3002\u540c\u65f6\u5408\u6210\u97f3\u8d28\u76f8\u6bd4\u4e8e\u79bb\u7ebf\u5408\u6210\u57fa\u672c\u65e0\u635f\u3002<\/p>\n\n\n\n<p><strong>4\uff09\u6307\u4ee4\u53ef\u63a7\u7684\u97f3\u9891\u751f\u6210\u80fd\u529b\u5347\u7ea7<\/strong>\uff1a\u4f18\u5316\u540e\u7684 CosyVoice 2.0 \u5728\u57fa\u6a21\u578b\u548c\u6307\u4ee4\u6a21\u578b\u7684\u6574\u5408\u4e0a\u53d6\u5f97\u4e86\u91cd\u8981\u8fdb\u5c55\uff0c\u4e0d\u4ec5\u5ef6\u7eed\u4e86\u5bf9\u60c5\u611f\u3001\u8bf4\u8bdd\u98ce\u683c\u548c\u7ec6\u7c92\u5ea6\u63a7\u5236\u6307\u4ee4\u7684\u652f\u6301\uff0c\u8fd8\u65b0\u589e\u4e86\u4e2d\u6587\u6307\u4ee4\u7684\u5904\u7406\u80fd\u529b\u3002\u5176\u6307\u4ee4\u63a7\u5236\u529f\u80fd\u7684\u6269\u5c55\u5c24\u4e3a\u663e\u8457\uff0c\u73b0\u5df2\u652f\u6301\u591a\u79cd\u4e3b\u8981\u65b9\u8a00<strong>\uff0c\u5305\u62ec\u7ca4\u8bed\u3001\u56db\u5ddd\u8bdd\u3001\u90d1\u5dde\u8bdd\u3001\u5929\u6d25\u8bdd\u548c\u957f\u6c99\u8bdd\u7b49<\/strong>\uff0c\u4e3a\u7528\u6237\u63d0\u4f9b\u4e86\u66f4\u4e30\u5bcc\u7684\u8bed\u8a00\u9009\u62e9\u3002\u6b64\u5916\uff0cCosyVoice 2.0 \u4e5f<strong>\u5f15\u5165\u4e86\u89d2\u8272\u626e\u6f14\u7684\u529f\u80fd<\/strong>\uff0c\u5982\u80fd\u591f\u6a21\u4eff\u673a\u5668\u4eba\u3001\u5c0f\u732a\u4f69\u5947\u7684\u98ce\u683c\u8bb2\u8bdd\u7b49\u3002\u8fd9\u4e9b\u529f\u80fd\u7684\u63d0\u5347\u8fd8\u4f34\u968f\u7740\u53d1\u97f3\u51c6\u786e\u6027\u548c\u97f3\u8272\u4e00\u81f4\u6027\u7684\u663e\u8457\u6539\u5584\uff0c\u4e3a\u7528\u6237\u5e26\u6765\u4e86\u66f4\u81ea\u7136\u548c\u751f\u52a8\u7684\u8bed\u97f3\u4f53\u9a8c\u3002<\/p>\n\n\n\n<h2><strong>\u258e\u652f\u6301\u7684\u529f\u80fd\uff1a<\/strong><\/h2>\n\n\n\n<p>\ud83c\udfa7&nbsp;<strong>\u97f3\u8272\u590d\u523b<\/strong>\uff1a\u8bed\u97f3\u514b\u9686\/Zero-shot In-context Generation\uff08\u8bed\u97f3\u7eed\u5199\uff09<\/p>\n\n\n\n<p><strong>\ud83c\udfa7&nbsp;<strong>\u591a\u8bed\u8a00\u5408\u6210<\/strong><\/strong>\uff1a<strong>\u4e00\u4e2a\u6587\u672c\uff0c\u5206\u522b\u7528<strong>\u4e2d\u6587\uff0c\u82f1\u6587\uff0c\u65e5\u8bed\uff0c\u97e9\u8bed<\/strong>\u591a\u4e2a\u8bed\u8a00\u6765\u8bf4<\/strong><\/p>\n\n\n\n<p><strong> <strong>\ud83c\udfa7<\/strong> \u6df7\u5408\u8bed\u79cd\u5408\u6210<\/strong>\uff1a\u652f\u6301 \u6587\u672c\u4e2d\u540c\u65f6\u51fa\u73b0 \u4e2d\u6587\/\u82f1\u6587\/\u65e5\u8bed\/\u97e9\u8bed\u7b49\u3002<\/p>\n\n\n\n<p>\ud83c\udfa7&nbsp;<strong>\u591a\u60c5\u611f\u5408\u6210<\/strong>\uff1a#\u538c\u6076# \u4eca\u5929\u53c8\u662f\u6253\u5de5\u4eba\u7684\u4e00\u5929\uff1b#\u6050\u60e7#\u554a\u5df2\u7ecf9\u70b9\u4e86\uff0c\u600e\u4e48\u529e\uff0c\u6211\u8981\u8fdf\u5230\u4e86\uff01#\u6124\u6012#\u90fd\u602a\u6628\u665a\u4ed6\u975e\u8981\u62c9\u6211\u770b\u7535\u5f71\uff0c\u5bb3\u6211\u7761\u665a\u4e86\uff01#\u5e73\u9759#\u4eca\u5e74\u7684\u5e74\u5047\u90fd\u7528\u5149\u4e86\uff0c#\u5f00\u5fc3#\u4e0d\u8fc7\u6ca1\u5173\u7cfb\uff0c\u9a6c\u4e0a\u8981\u653e\u5047\u5566\uff01<\/p>\n\n\n\n<p>\ud83c\udfa7&nbsp;<strong>\u4e0d\u540c\u6307\u4ee4\u5408\u6210<\/strong>\uff1a<\/p>\n\n\n\n<div class=\"wp-block-columns\">\n<div class=\"wp-block-column\" style=\"flex-basis:16.67%\">\n<p>\uff08\u795e\u79d8&lt;|endofprompt|&gt;\u53e4\u8001\u57ce\u5821\u7b3c\u7f69\u5728\u795e\u79d8\u7684\u96fe\u6c14\u4e2d\uff0c\u5438\u5f15\u7740\u65e0\u6570\u5192\u9669\u8005\u524d\u53bb\u63a2\u7d22\u5965\u79d8\u3002<\/p>\n<\/div>\n\n\n\n<div class=\"wp-block-column\" style=\"flex-basis:16.67%\">\n<p>\u5c0f\u732a\u4f69\u5947&lt;|endofprompt|&gt;\u5728\u5fd9\u788c\u4e4b\u4f59\uff0c\u6211\u548c\u670b\u53cb\u50cf\u5c0f\u732a\u4f69\u5947\u4e00\u6837\uff0c\u5e38\u53bb\u516c\u56ed\u4eab\u53d7\u7b80\u5355\u7684\u5feb\u4e50\u3002<\/p>\n<\/div>\n\n\n\n<div class=\"wp-block-column\" style=\"flex-basis:16.67%\">\n<p>\u56db\u5ddd\u8bdd&lt;|endofprompt|&gt;\u800c\u8fd9\u4e9b\u5e7d\u9ed8\u7684\u77ac\u95f4\u4eff\u4f5b\u8ba9\u6211\u7f6e\u8eab\u4e8e\u56db\u5ddd\u7684\u5bbd\u7a84\u5df7\u5b50\uff0c\u4eab\u53d7\u90a3\u4efd\u5b89\u9038\u3002<\/p>\n<\/div>\n\n\n\n<div class=\"wp-block-column\" style=\"flex-basis:16.67%\">\n<p>\u7528\u4f24\u5fc3\u7684\u8bed\u6c14\u8bf4&lt;|endofprompt|&gt;\u6536\u5230\u62d2\u4fe1\u7684\u90a3\u4e00\u523b\uff0c\u6211\u611f\u5230\u65e0\u6bd4\u4f24\u5fc3\u3002\u867d\u7136\u77e5\u9053\u5931\u8d25\u662f\u6210\u957f\u7684\u4e00\u90e8\u5206\uff0c\u4f46\u4ecd\u7136\u96be\u4ee5\u63a9\u9970\u5fc3\u4e2d\u7684\u5931\u843d\u3002<\/p>\n<\/div>\n\n\n\n<div class=\"wp-block-column\" style=\"flex-basis:16.67%\">\n<p>\u6162\u901f&lt;|endofprompt|&gt;\u542c\u7740\u8f7b\u67d4\u7684\u97f3\u4e50\uff0c\u6211\u5728\u753b\u5e03\u4e0a\u6162\u6162\u5730\u6d82\u62b9\u8272\u5f69\uff0c\u8ba9\u6bcf\u4e00\u7b14\u90fd\u5145\u6ee1\u7075\u611f\u548c\u601d\u8003\u3002<\/p>\n<\/div>\n\n\n\n<div class=\"wp-block-column\" style=\"flex-basis:16.67%\">\n<p>\u8ffd\u6c42\u5353\u8d8a\u4e0d\u662f\u7ec8\u70b9\uff0c\u5b83\u9700\u8981\u4f60\u6bcf\u5929\u90fd&lt;strong&gt;\u4ed8\u51fa&lt;\/strong&gt;\u548c&lt;strong&gt;\u7cbe\u8fdb&lt;\/strong&gt;\uff0c\u6700\u7ec8\u624d\u80fd\u8fbe\u5230\u5dc5\u5cf0\u3002\uff09<\/p>\n<\/div>\n<\/div>\n\n\n\n<p>\ud83c\udfa7&nbsp;<strong>\u7ed5\u53e3\u4ee4<\/strong>\uff1a\u9ed1\u5316\u80a5\u53d1\u7070\uff0c\u7070\u5316\u80a5\u53d1\u9ed1\uff0c\u9ed1\u5316\u80a5\u6325\u53d1\u4f1a\u53d1\u9ed1\uff0c\u7070\u5316\u80a5\u6325\u53d1\u4f1a\u53d1\u7070\u3002<\/p>\n\n\n\n<p>\ud83c\udfa7&nbsp;<strong>\u751f\u50fb\u5b57<\/strong>\uff1a\u7162\u7162\u5b51\u7acb \u6c86\u7023\u4e00\u6c23 \u8e3d\u8e3d\u7368\u884c \u918d\u9190\u704c\u9802 \u7dbf\u7dbf\u74dc\u74de \u5949\u70ba\u572d\u81ec \u9f8d\u884c\u9f98\u9f98 <\/p>\n\n\n\n<p>\ud83c\udfa7&nbsp;<strong>\u591a\u97f3\u5b57<\/strong>\uff1a\u5929\u6c14\u6696<strong>\u548c<\/strong>\uff0c\u5c0f\u738b\u5728\u5bb6<strong>\u548c<\/strong>\u6ce5\u62b9\u5899\uff1b<\/p>\n\n\n\n<h2><strong>Demo\u4f53\u9a8c\uff1a<\/strong><\/h2>\n\n\n\n<p><strong>&gt;&gt;&gt;\u521b\u7a7a\u95f4\u5730\u5740<\/strong>\uff1ahttps:\/\/www.modelscope.cn\/studios\/iic\/CosyVoice2-0.5B<\/p>\n\n\n\n<p>\u53ef\u4ee5\u652f\u6301\u7528\u6237\u4e0a\u4f20\u97f3\u9891\u6587\u4ef6\u6216\u5f55\u97f3\u65b9\u5f0f\u8fdb\u884c\u8bed\u97f3\u590d\u523b\u3002\u540c\u65f6\u652f\u6301\u6d41\u5f0f\u63a8\u7406\uff0c\u7528\u6237\u65e0\u9700\u7b49\u5f85\u5168\u90e8\u97f3\u9891\u5408\u6210\u5b8c\u6bd5\u5373\u53ef\u4f53\u9a8c\u6548\u679c\u3002<\/p>\n\n\n\n<p>CosyVoice 2\u652f\u6301\u97f3\u8272\u514b\u9686\u4ee5\u53ca\u81ea\u7136\u8bed\u8a00\u63a7\u5236\u7684\u97f3\u9891\u751f\u6210\uff0c\u53ef\u4ee5\u9009\u62e9\u76f8\u5e94\u7684\u63a8\u7406\u6a21\u5f0f\u3002<\/p>\n\n\n\n<p>1\uff09<strong>3s\u6781\u901f\u590d\u523b<\/strong><\/p>\n\n\n\n<ul><li>\u8f93\u5165\u5f85\u5408\u6210\u6587\u6848<\/li><li>\u9009\u62e9\u662f\u5426\u6d41\u5f0f\u63a8\u7406\uff0c\u6d41\u5f0f\u63a8\u7406\u5177\u6709\u66f4\u4f4e\u7684\u5ef6\u8fdf\uff0c\u79bb\u7ebf\u63a8\u7406\u5177\u6709\u66f4\u597d\u7684\u4e0a\u9650\u6548\u679c<\/li><li>\u4e0a\u4f20prompt\u97f3\u9891\uff0c\u6216\u8005\u5f55\u5236prompt\u97f3\u9891<\/li><li>\u70b9\u51fb\u751f\u6210\u97f3\u9891\uff0c\u7b49\u5f85\u4e00\u4f1a\u513f\u5c31\u4f1a\u542c\u5230\u5408\u6210\u7684\u97f3\u9891\u3002<\/li><\/ul>\n\n\n\n<p>2\uff09<strong>\u81ea\u7136\u8bed\u8a00\u63a7\u5236<\/strong><\/p>\n\n\n\n<ul><li>\u8f93\u5165\u5f85\u5408\u6210\u6587\u6848<\/li><li>\u4e0a\u4f20prompt\u97f3\u9891\uff0c\u6216\u8005\u5f55\u5236prompt\u97f3\u9891<\/li><li>\u8f93\u5165instruct\u6587\u672c\uff1a\u4f8b\u5982\u201c\u7528\u7ca4\u8bed\u8bf4\u8fd9\u53e5\u8bdd\u201d\uff0c\u201c\u7528\u5f00\u5fc3\u7684\u8bed\u6c14\u8bf4\u201d\uff0c\u201c\u6a21\u4eff\u673a\u5668\u4eba\u7684\u58f0\u97f3\u201d\u7b49<\/li><li>\u70b9\u51fb\u751f\u6210\u97f3\u9891\uff0c\u7b49\u5f85\u4e00\u4f1a\u513f\u5c31\u4f1a\u542c\u5230\u5408\u6210\u7684\u97f3\u9891\u3002<\/li><\/ul>\n\n\n\n<h2><strong>\u8865\u5145<\/strong>\uff1a<strong>v1\u7248\u672c<\/strong><\/h2>\n\n\n\n<p class=\"has-text-align-center\"><strong> CosyVoice: A Scalable Multilingual Zero-shot Text-to-speech Synthesizer based on Supervised Semantic Tokens<\/strong><\/p>\n\n\n\n<p class=\"has-text-align-center\"><strong>CosyVoice 1.0<\/strong>:&nbsp;<a href=\"https:\/\/fun-audio-llm.github.io\/\">Demos<\/a>;&nbsp;<a href=\"https:\/\/funaudiollm.github.io\/pdf\/CosyVoice_v1.pdf\">Paper<\/a>;&nbsp;<a href=\"https:\/\/www.modelscope.cn\/studios\/iic\/CosyVoice-300M\">Modelscope<\/a><\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"550\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-49-1024x550.png\" alt=\"\" class=\"wp-image-23123\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-49-1024x550.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-49-300x161.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-49-768x412.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-49.png 1354w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><figcaption>V1\u7248\u672ccosyvoice<\/figcaption><\/figure>\n\n\n\n<h2>\u8bba\u6587\u89e3\u8bfb\uff1a<\/h2>\n\n\n\n<h3>\u6458\u8981<\/h3>\n\n\n\n<p class=\"has-text-align-left\">&nbsp;<strong>CosyVoice<\/strong>\uff0c\u8fd9\u662f\u4e00\u79cd\u57fa\u4e8e\u76d1\u7763\u79bb\u6563\u8bed\u97f3\u6807\u8bb0\u7684\u591a\u8bed\u8a00\u8bed\u97f3\u5408\u6210\u6a21\u578b\u3002\u901a\u8fc7\u4f7f\u7528\u4e24\u79cd\u6d41\u884c\u7684\u751f\u6210\u6a21\u578b\uff1a\u8bed\u8a00\u6a21\u578b \uff08LM\uff09 \u548c\u6d41\u5339\u914d\u8fdb\u884c\u6e10\u8fdb\u5f0f\u8bed\u4e49\u89e3\u7801\uff0cCosyVoice \u5728\u8bed\u97f3\u4e0a\u4e0b\u6587\u5b66\u4e60\u4e2d\u5b9e\u73b0\u4e86\u9ad8\u97f5\u5f8b\u81ea\u7136\u5ea6\u3001\u5185\u5bb9\u4e00\u81f4\u6027\u548c\u8bf4\u8bdd\u4eba\u76f8\u4f3c\u6027\u3002\u6700\u8fd1\uff0c\u591a\u6a21\u6001\u5927\u578b\u8bed\u8a00\u6a21\u578b \uff08LLMs\uff0c\u5176\u4e2d\u54cd\u5e94\u5ef6\u8fdf\u548c\u8bed\u97f3\u5408\u6210\u7684\u5b9e\u65f6\u56e0\u5b50\u5728\u4ea4\u4e92\u4f53\u9a8c\u4e2d\u8d77\u7740\u81f3\u5173\u91cd\u8981\u7684\u4f5c\u7528\u3002\u56e0\u6b64\uff0c\u5728\u8fd9\u9879\u5de5\u4f5c\u4e2d\uff0c\u6211\u4eec\u5f15\u5165\u4e86\u4e00\u79cd\u6539\u8fdb\u7684\u6d41\u5f0f\u8bed\u97f3\u5408\u6210\u6a21\u578b&nbsp;<strong>CosyVoice 2<\/strong>\uff0c\u5e76\u8fdb\u884c\u4e86\u5168\u9762\u548c\u7cfb\u7edf\u7684\u4f18\u5316\u3002\u9996\u5148\uff0c\u6211\u4eec\u5f15\u5165<strong>\u6709\u9650\u6807\u91cf\u91cf\u5316\u6765\u63d0\u9ad8\u8bed\u97f3\u6807\u8bb0\u7684\u7801\u672c\u5229\u7528\u7387<\/strong>\u3002\u5176\u6b21\uff0c\u6211\u4eec<strong>\u7b80\u5316\u4e86\u6587\u672c-\u8bed\u97f3 LM \u7684\u6a21\u578b\u67b6\u6784<\/strong>\uff0c\u4ee5\u4fbf\u53ef\u4ee5\u76f4\u63a5\u4f7f\u7528\u9884\u8bad\u7ec3\u7684 LLMs \u4f5c\u4e3a\u4e3b\u5e72\u3002\u6b64\u5916\uff0c\u6211\u4eec\u8fd8\u8bbe\u8ba1\u4e86\u4e00\u4e2a<strong> chunk-aware \u3010\u6570\u636e\u5757\u611f\u77e5\u3011\u56e0\u679c\u6d41\u5339\u914d\u6a21\u578b<\/strong>\uff0c\u4ee5\u9002\u5e94\u4e0d\u540c\u7684 synthesis \u573a\u666f\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u53ef\u4ee5\u5728<strong>\u5355\u4e2a\u6a21\u578b\u4e2d\u6267\u884c streaming \u548c\u975e streaming <\/strong>\u5408\u6210\u3002\u901a\u8fc7\u5728\u5927\u89c4\u6a21\u591a\u8bed\u8a00\u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u8bad\u7ec3\uff0cCosyVoice 2 \u5b9e\u73b0\u4e86\u4e0e\u4eba\u7c7b\u76f8\u5f53\u7684\u5408\u6210\u8d28\u91cf\uff0c\u5177\u6709\u975e\u5e38\u4f4e\u7684\u54cd\u5e94\u5ef6\u8fdf\u548c\u5b9e\u65f6\u7cfb\u6570\u3002<\/p>\n\n\n\n<p>CosyVoice 2\uff0c\u8fd9\u662f\u4e00\u79cd\u6d41\u5f0f\u96f6\u6837\u672c TTS \u6a21\u578b\uff0c\u5177\u6709\u6539\u8fdb\u7684\u97f5\u5f8b\u81ea\u7136\u5ea6\u3001\u5185\u5bb9\u4e00\u81f4\u6027\u548c\u8bf4\u8bdd\u4eba\u76f8\u4f3c\u6027\u3002\u6211\u4eec\u7684\u8d21\u732e\u5305\u62ec\uff1a<\/p>\n\n\n\n<ul class=\"has-light-gray-background-color has-background\"><li>\u5c06\u6d41\u5f0f\u548c\u975e\u6d41\u5f0f\u5408\u6210\u7edf\u4e00\u5728\u4e00\u4e2a\u6846\u67b6\u4e2d\uff0c\u5e76\u63d0\u51fa\u7edf\u4e00\u7684\u6587\u672c\u8bed\u97f3\u8bed\u8a00\u6a21\u578b\u548c\u5757\u611f\u77e5\u56e0\u679c\u6d41\u5339\u914d\u6a21\u578b\uff0c\u4e0e\u79bb\u7ebf\u6a21\u5f0f\u76f8\u6bd4\uff0c\u5b9e\u73b0\u4e86\u65e0\u635f\u6d41\u5f0f\u5408\u6210<\/li><li>\u901a\u8fc7\u79fb\u9664\u6587\u672c\u7f16\u7801\u5668\u548c\u8bf4\u8bdd\u4eba\u5d4c\u5165\u6765\u7b80\u5316 LM \u67b6\u6784\uff0c\u5141\u8bb8\u9884\u5148\u8bad\u7ec3\u7684\u6587\u672c\u5927\u578b\u8bed\u8a00\u6a21\u578b \uff08LLMs\uff09 \u4f5c\u4e3a\u4e3b\u5e72\uff0c\u589e\u5f3a\u4e0a\u4e0b\u6587\u7406\u89e3\u3002<\/li><li>\u7528\u6709\u9650\u6807\u91cf\u91cf\u5316 \uff08FSQ\uff09 \u66ff\u6362\u8bed\u97f3\u5206\u8bcd\u5668\u4e2d\u7684\u77e2\u91cf\u91cf\u5316 \uff08VQ\uff09\uff0c\u63d0\u9ad8\u7801\u672c\u5229\u7528\u7387\u5e76\u6355\u83b7\u66f4\u591a\u8bed\u97f3\u4fe1\u606f\u3002<\/li><li>\u5347\u7ea7\u6307\u793a\u5f0f TTS \u5bb9\u91cf\u4ee5\u652f\u6301\u66f4\u591a\u6307\u4ee4\uff0c\u5305\u62ec\u60c5\u611f\u3001\u53e3\u97f3\u3001\u89d2\u8272\u98ce\u683c\u548c\u7cbe\u7ec6\u63a7\u5236\u3002\u5728 CosyVoice 2 \u4e2d\uff0c\u6307\u4ee4\u548c\u96f6\u955c\u5934\u5bb9\u91cf\u96c6\u6210\u5230\u4e00\u4e2a\u6a21\u578b\u4e2d\uff0c\u5b9e\u73b0\u66f4\u901a\u7528\u3001\u66f4\u751f\u52a8\u7684\u5408\u6210\u3002<\/li><\/ul>\n\n\n\n<p>\u901a\u8fc7\u4ee5\u4e0a\u7cfb\u7edf\u6027\u7684\u4fee\u6539\u548c\u4f18\u5316\uff0cCosyVoice 2 \u5b9e\u73b0\u4e86\u4eba\u5076\u6821\u9a8c\u7684\u5408\u6210\u8d28\u91cf\uff0c\u5e76\u4e14\u5728 streaming \u6a21\u5f0f\u4e0b\u51e0\u4e4e\u65e0\u635f\u3002\u7edf\u4e00\u6846\u67b6\u653e\u5bbd\u4e86\u90e8\u7f72\u8981\u6c42\uff0c\u4f7f\u5355\u4e2a\u6a21\u578b\u80fd\u591f\u540c\u65f6\u652f\u6301\u6d41\u5f0f\u548c\u975e\u6d41\u5f0f\u5408\u6210\u3002\u5347\u7ea7\u540e\u7684 instructed TTS \u5bb9\u91cf\u4e3a\u7528\u6237\u751f\u6210\u5404\u79cd\u8bed\u97f3\u63d0\u4f9b\u4e86\u66f4\u5f3a\u5927\u3001\u66f4\u8f7b\u677e\u7684\u65b9\u6cd5\u3002\u6b64\u5916\uff0c<strong>\u5757\u611f\u77e5\u6d41\u5339\u914d\u8bbe\u8ba1\u4e5f\u53ef\u4ee5\u5e94\u7528\u4e8e NAR TTS \u6a21\u578b\uff0c\u8fd9\u8868\u660e\u4e86\u6d41\u5f0f NAR \u6a21\u578b\u7684\u6f5c\u529b\u3002<\/strong><\/p>\n\n\n\n<h3>CosyVoice 2<\/h3>\n\n\n\n<figure class=\"wp-block-image size-full is-resized\"><img loading=\"lazy\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-77.png\" alt=\"\" class=\"wp-image-23277\" width=\"687\" height=\"263\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-77.png 940w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-77-300x115.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-77-768x295.png 768w\" sizes=\"(max-width: 687px) 100vw, 687px\" \/><figcaption>\u56fe 1\uff1aCosyVoice 2 \u6982\u8ff0\u3002\uff08a\uff09 \u6f14\u793a\u4e86\u76d1\u7763\u5f0f\u8bed\u97f3\u5206\u8bcd\u5668\uff0c\u5176\u4e2d\u865a\u7ebf\u6a21\u5757\u4ec5\u5728\u8bad\u7ec3\u9636\u6bb5\u4f7f\u7528\u3002\uff08b\uff09 \u662f\u7528\u4e8e\u6d41\u5f0f\u5904\u7406\u548c\u975e\u6d41\u5f0f\u5904\u7406\u5408\u6210\u7684\u7edf\u4e00\u6587\u672c-\u8bed\u97f3\u8bed\u8a00\u6a21\u578b\u3002\u865a\u7ebf\u8868\u793a\u63a8\u7406\u9636\u6bb5\u7684\u81ea\u56de\u5f52\u89e3\u7801\u3002\uff08c\uff09 \u8bf4\u660e\u4e86\u5728\u6982\u7387\u5bc6\u5ea6\u8def\u5f84\u4e0a\u65f6\u95f4\u6b65t\u5bf9\u8bf4\u8bdd\u4eba\u5d4c\u5165&nbsp;\ud835\udc2f\u3001\u8bed\u4e49\u6807\u8bb0&nbsp;\u03bc\u3001\u63a9\u853d\u8bed\u97f3\u7279\u5f81&nbsp;X~\u548c\u4e2d\u95f4\u72b6\u6001&nbsp;Xt\u7684\u56e0\u679c\u6d41\u5339\u914d\u6a21\u578b\u6761\u4ef6\u3002<\/figcaption><\/figure>\n\n\n\n<p>CosyVoice 2 \u5efa\u7acb\u5728\u4e0e\u524d\u4ee3&nbsp;\u7c7b\u4f3c\u7684\u8bbe\u8ba1\u7406\u5ff5\u4e4b\u4e0a\uff0c\u5c06\u8bed\u97f3\u4fe1\u53f7\u7684\u8bed\u4e49\u548c\u58f0\u5b66\u4fe1\u606f\u5206\u79bb\u51fa\u6765\uff0c\u5e76\u72ec\u7acb\u5efa\u6a21\u3002\u8bed\u97f3\u751f\u6210\u8fc7\u7a0b\u88ab\u91cd\u65b0\u5b9a\u4e49\u4e3a\u4e00\u4e2a\u6e10\u8fdb\u7684\u8bed\u4e49\u89e3\u7801\u8fc7\u7a0b\uff0c\u5176\u4e2d\u6761\u4ef6\u4fe1\u606f\u9010\u6e10\u88ab\u7eb3\u5165\u3002\u5177\u4f53\u6765\u8bf4\uff0c<strong>\u6587\u672c-\u8bed\u97f3\u8bed\u8a00\u6a21\u578b \uff08LM\uff09 \u53ea\u5173\u6ce8\u8bed\u4e49\u4fe1\u606f\uff0c\u5c06\u9ad8\u7ea7\u6587\u672c\u6807\u8bb0\u89e3\u7801\u4e3a\u76d1\u7763\u5f0f\u8bed\u4e49\u8bed\u97f3\u6807\u8bb0\u3002\u5728 Flow Matching \u6a21\u578b\u4e2d\uff0c\u901a\u8fc7\u8bf4\u8bdd\u4eba\u5d4c\u5165\u548c\u53c2\u8003\u8bed\u97f3\u5f15\u5165\u58f0\u5b66\u7ec6\u8282\uff08\u4f8b\u5982\u97f3\u8272\uff09\uff0c\u5c06\u8bed\u97f3\u4ee4\u724c\u8f6c\u6362\u4e3a\u7ed9\u5b9a\u8bf4\u8bdd\u4eba\u7684 Mel \u9891\u8c31\u3002\u6700\u540e\uff0c\u9884\u5148\u8bad\u7ec3\u7684\u58f0\u7801\u5668\u6a21\u578b\u6062\u590d\u4e86\u76f8\u4f4d\uff0c\u5c06 Mel \u9891\u8c31\u8f6c\u6362\u56de\u539f\u59cb\u97f3\u9891\u4fe1\u53f7\u3002<\/strong>\u4e0b\u9762\u5c06\u4ece\u6587\u672c\u5206\u8bcd\u5668\u3001\u76d1\u7763\u8bed\u4e49\u8bed\u97f3\u5206\u8bcd\u5668\u3001\u6d41\u5f0f\/\u975e\u6d41\u5f0f\u5408\u6210\u7684\u7edf\u4e00\u6587\u672c\u8bed\u97f3 LM \u548c\u5206\u5757\u611f\u77e5\u6d41\u5f0f\u5339\u914d\u6a21\u578b\u4e94\u4e2a\u65b9\u9762\u4ecb\u7ecd CosyVoice 2 \u7684\u7ec6\u8282\u548c\u6d41\u5f0f\u5408\u6210\u7684\u4fee\u6539\u3002<\/p>\n\n\n\n<h4>Text Tokenizer<\/h4>\n\n\n\n<p>CosyVoice 2 \u76f4\u63a5\u4f7f\u7528\u539f\u59cb\u6587\u672c\u4f5c\u4e3a\u8f93\u5165\uff0c\u5e76\u4f7f\u7528\u57fa\u4e8e BPE \u7684\u6587\u672c\u5206\u8bcd\u5668\u8fdb\u884c\u5206\u8bcd\u3002\u8fd9<strong>\u6d88\u9664\u4e86\u5bf9\u901a\u8fc7\u5b57\u7d20\u5230\u97f3\u7d20 \uff08g2p\uff09 \u8f6c\u6362\u83b7\u53d6\u97f3\u7d20\u7684\u524d\u7aef\u6a21\u578b\u7684\u9700\u6c42<\/strong>\u3002\u8fd9\u79cd\u65b9\u6cd5\u4e0d\u4ec5\u7b80\u5316\u4e86\u6570\u636e\u9884\u5904\u7406\u5de5\u4f5c\u6d41\u7a0b\uff0c\u8fd8\u4f7f\u6a21\u578b\u80fd\u591f\u4ee5\u7aef\u5230\u7aef\u7684\u65b9\u5f0f\u5b66\u4e60\u5404\u79cd\u4e0a\u4e0b\u6587\u4e2d\u5355\u8bcd\u7684\u53d1\u97f3\u3002\u4e0e\u6587\u672c LLMs\uff0cCosyVoice 2 \u5c4f\u853d\u4e86\u4e00\u5bf9\u591a\u7684\u5206\u8bcd\u3002\u8fd9\u53ef\u4ee5\u9632\u6b62\u4ee4\u724c\u7684\u53d1\u97f3\u53d8\u5f97\u8fc7\u957f\uff0c\u5e76\u51cf\u5c11\u7531\u6570\u636e\u7a00\u758f\u5f15\u8d77\u7684\u6781\u7aef\u60c5\u51b5\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u5982\u679c BPE \u4ee4\u724c\u7f16\u7801\u591a\u4e2a\u4e2d\u6587\u5b57\u7b26\uff0c\u5b83\u5c06\u88ab\u5c4f\u853d\u6389\uff0c\u5e76\u4e14\u6bcf\u4e2a\u5b57\u7b26\u5c06\u5728\u5206\u8bcd\u5316\u8fc7\u7a0b\u4e2d\u5355\u72ec\u7f16\u7801\u3002\u5176\u4ed6\u8bed\u8a00\uff08\u5982\u82f1\u8bed\u3001\u65e5\u8bed\u548c\u97e9\u8bed\uff09\u4e0d\u53d7\u7279\u6b8a\u5904\u7406\u3002<\/p>\n\n\n\n<h4>Supervised Semantic Speech Tokenizer<\/h4>\n\n\n\n<p>\u5c06<strong>\u6709\u9650\u6807\u91cf\u91cf\u5316<\/strong> \uff08<strong>FSQ<\/strong>\uff09 \u6a21\u5757\u63d2\u5165<strong> SenseVoice-Large ASR<\/strong> \u6a21\u578b\u7684\u7f16\u7801\u5668\u4e2d\u3002\u5728\u8bad\u7ec3\u9636\u6bb5\uff0c\u8f93\u5165\u8bed\u97f3&nbsp;X&nbsp;\u901a\u8fc7 \u83b7\u5f97&nbsp;Encoder1&nbsp;\u4e2d\u95f4\u8868\u793a\uff0c\u5176\u4e2d&nbsp;Encoder1&nbsp;\u7531\u516d\u4e2a\u5e26\u6709\u65cb\u8f6c\u4f4d\u7f6e\u5d4c\u5165&nbsp;\u7684 Transformer \u5757\u7ec4\u6210\u3002\u7136\u540e\uff0c\u5c06\u4e2d\u95f4\u8868\u793a\u9988\u9001\u5230 FSQ \u6a21\u5757\u8fdb\u884c\u91cf\u5316\uff0c\u5e76\u5c06\u91cf\u5316\u8868\u793a\u4f20\u9012\u7ed9 SenseVoice-Large \u6a21\u5757\u7684\u5176\u4f59\u90e8\u5206\uff0c\u5305\u62ec&nbsp;Encoder2&nbsp;\u548c&nbsp;ASR\u2062Decoder&nbsp;\uff0c\u4ee5\u9884\u6d4b\u76f8\u5e94\u6587\u672c\u6807\u8bb0\u7684\u540e\u9a8c\u6982\u7387\u3002<\/p>\n\n\n\n<h5><em>\u8865\u5145FSQ\uff1a<\/em>FINITE SCALAR QUANTIZATION: VQ-VAE MADE SIMPLE<\/h5>\n\n\n\n<div class=\"wp-block-columns\">\n<div class=\"wp-block-column\" style=\"flex-basis:100%\">\n<p class=\"has-light-gray-background-color has-background\"><em>2023 \u5e74 google \u53d1\u8868\u7684\u6587\u7ae0\uff0c\u53ef\u4ee5\u7528\u4e8e\u6587\u672c\u3001\u89c6\u9891\u751f\u6210\u9886\u57df\u4e2d\u3002\u63d0\u51fa\u4e00\u79cd\u79f0\u4e3a\u6709\u9650\u6807\u91cf\u91cf\u5316\uff08FSQ\uff09\u7684\u7b80\u5355\u65b9\u6848\u6765\u66ff\u6362 VQ-VAEs \u4e2d\u7684\u5411\u91cf\u91cf\u5316\uff08VQ\uff09\u3002\u89e3\u51b3\u4f20\u7edf VQ \u4e2d\u7684\u4e24\u4e2a\u4e3b\u8981\u95ee\u9898\uff1a<strong>\u9700\u8981\u907f\u514d codebook collapse \uff08\u7801\u7801\u672c\u574d\u584c\uff1a\u5c31\u662f\u5927\u90e8\u5206codebook elements\u90fd\u6ca1\u6709\u88ab\u7528\u5230<\/strong>\uff0c\u591a\u4e2a\u7801\u5b57\u53ef\u80fd\u4f1a\u53d8\u5f97\u975e\u5e38\u76f8\u4f3c\uff0c\u751a\u81f3\u5b8c\u5168\u76f8\u540c\u3002<strong>\uff09\u7684\u8f85\u52a9\u635f\u5931\u3010\u5728\u635f\u5931\u51fd\u6570\u4e2d\u6dfb\u52a0\u4e00\u4e2a\u8f85\u52a9\u635f\u5931\u9879\uff0c\u8feb\u4f7f\u7801\u5b57\u4e4b\u95f4\u4fdd\u6301\u4e00\u5b9a\u7684\u8ddd\u79bb\uff0c\u6216\u8005\u4fdd\u6301\u5176\u591a\u6837\u6027\u3011\u3001\u5927 codebook size \u60c5\u51b5\u4e0b\u7801\u672c\u5229\u7528\u7387\u4f4e\u3002<\/strong>FSQ \u4f5c\u7528\uff1a<strong>\u6d88\u9664\u8f85\u52a9\u635f\u5931\u3001\u63d0\u9ad8\u7801\u672c\u5229\u7528\u7387\u3001\u4f5c\u4e3a VQ \u7684\u53ef\u66ff\u6362\u7ec4\u4ef6<\/strong><\/em>\u3002<\/p>\n\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter size-large is-resized\"><img loading=\"lazy\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-78-1024x341.png\" alt=\"\" class=\"wp-image-23297\" width=\"554\" height=\"184\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-78-1024x341.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-78-300x100.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-78-768x256.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-78.png 1117w\" sizes=\"(max-width: 554px) 100vw, 554px\" \/><\/figure><\/div>\n\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter size-full is-resized\"><img loading=\"lazy\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-79.png\" alt=\"\" class=\"wp-image-23299\" width=\"496\" height=\"230\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-79.png 945w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-79-300x139.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-79-768x357.png 768w\" sizes=\"(max-width: 496px) 100vw, 496px\" \/><\/figure><\/div>\n\n\n\n<p class=\"has-light-gray-background-color has-background\"><em>\u4f20\u7edf\u7684\u7f16\u7801\u5668\u6240\u5f97\u5230\u7684\u8868\u5f81\u5411\u91cfz\u4e2d\u7684\u6bcf\u4e00\u4e2a\u5143\u7d20(\u6807\u91cf)\u7684\u503c\u5e76\u6ca1\u6709\u4e00\u4e2a\u660e\u786e\u7684\u8fb9\u754c\uff0c\u4e5f\u5c31\u662f\u8bf4z\u5728\u7279\u5f81\u7a7a\u95f4\u4e2d\u4e0d\u53d7\u4efb\u4f55\u7ea6\u675f\u3002\u90a3\u4e48\uff0c\u4f5c\u8005\u5c31\u60f3\u5230\u4e86\u4e3az\u4e2d\u7684\u6bcf\u4e2a\u6807\u91cf\u90fd\u8bbe\u5b9a\u597d\u53d6\u503c\u7684\u8303\u56f4\u548c\u80fd\u591f\u53d6\u503c\u7684\u4e2a\u6570\u3002\u5047\u8bbe\u6709\u4e00\u4e2ad\u7ef4\u7279\u5f81\u5411\u91cfz\uff0c\u5c06\u6bcf\u4e2a\u6807\u91cfzi\u200b\u90fd\u9650\u5236\u53ea\u80fd\u53d6L\u4e2a\u503c\uff0c\u5c06zi\u200b\u2192\u230aL\/2\u230btanh(zi\u200b)\u7136\u540e\u56db\u820d\u4e94\u5165\u4e3a\u4e00\u4e2a\u6574\u6570\u503c\u3002\u4f8b\u5982\u56fe\u4e2d\u6240\u793a\uff0c\u53d6d=3\uff0cL=3\uff0c\u4ee3\u8868C={(\u22121,\u22121,\u22121),(\u22121,\u22121,0),&#8230;,(1,1,1)}\uff0c\u4e00\u5171\u670927\u79cd\u7ec4\u5408\uff0c\u5373\u4e00\u4e2a3\u7ef4\u5411\u91cf\u7684\u6bcf\u4e2a\u6807\u91cf\u90fd\u6709\u4e09\u79cd\u503c\u7684\u53d6\u6cd5\u3002\u503c\u5f97\u4e00\u63d0\u7684\u662f\uff0cFSQ\u4e2d\u7684codebook\u4e0d\u50cfVQ-VAE\u90a3\u6837\u662f\u663e\u5f0f\u5b58\u5728\u7684\uff0c\u800c\u662f\u9690\u5f0f\u7684\uff0c\u7f16\u7801\u5668\u76f4\u63a5\u8f93\u51fa\u91cf\u5316\u540e\u7684\u7279\u5f81\u5411\u91cfz^\u3002\u56e0\u6b64\uff0cFSQ\u4e5f\u5c31\u6ca1\u6709\u4e86VQ-VAE\u635f\u5931\u7684\u540e\u4e24\u9879\u4e86\u3002<\/em><\/p>\n\n\n\n<p class=\"has-light-gray-background-color has-background\"><em>FSQ \u7684\u4f18\u70b9\u662f\u4e0d\u4f1a\u906d\u53d7\u7801\u672c\u574d\u584c\uff08codebook collapse\uff09\uff0c\u5e76\u4e14\u4e0d\u9700\u8981 VQ \u4e2d\u4e3a\u4e86\u907f\u514d\u7801\u672c\u574d\u584c\u800c\u4f7f\u7528\u7684\u590d\u6742\u673a\u5236\uff08\u627f\u8bfa\u635f\u5931\u3001\u7801\u672c\u91cd\u65b0\u64ad\u79cd\u3001\u7801\u5206\u5272\u3001\u71b5\u60e9\u7f5a\u7b49\uff09<\/em><\/p>\n<\/div>\n<\/div>\n\n\n\n<p>\u5728 FSQ \u6a21\u5757\u4e2d\uff0c\u9996\u5148&nbsp;H&nbsp;\u5c06\u4e2d\u95f4\u8868\u793a\u6295\u5f71\u5230&nbsp;D&nbsp;\u7ef4\u4f4e\u79e9\u7a7a\u95f4\u4e2d\uff0c\u5e76\u4e14\u6bcf\u4e2a\u7ef4\u5ea6\u7684\u503c\u901a\u8fc7\u6709\u754c\u56db\u820d\u4e94\u5165\u64cd\u4f5cROUND \u91cf\u5316\u5230\u533a\u95f4[\u2212K,K]\u3002\u7136\u540e\uff0c\u91cf\u5316\u540e\u7684\u4f4e\u79e9\u8868\u793a H\u02c9 \u88ab\u6295\u5f71\u56de\u539f\u59cb\u7ef4\u5ea6 H~\uff0c\u4ee5\u4f9b\u540e\u7eed\u6a21\u5757\u4f7f\u7528\u3002<\/p>\n\n\n\n<p>\u5728\u8bad\u7ec3\u9636\u6bb5\uff0c&nbsp;straight-through estimation\u7528\u4e8e\u8fd1\u4f3c FSQ \u6a21\u5757 \u548c&nbsp;Encoder1&nbsp;\u7684\u68af\u5ea6\u3002 \u8bed\u97f3\u6807\u8bb0&nbsp;\u03bci&nbsp;\u53ef\u4ee5\u901a\u8fc7\u8ba1\u7b97&nbsp;(2\u2062K+1)&nbsp;-ary \u7cfb\u7edf\u4e2d\u91cf\u5316\u7684\u4f4e\u79e9\u8868\u793a&nbsp;h\u00afi&nbsp;\u7684\u7d22\u5f15\u6765\u83b7\u5f97\uff1a<\/p>\n\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter size-full is-resized\"><img loading=\"lazy\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-80.png\" alt=\"\" class=\"wp-image-23311\" width=\"200\" height=\"57\"\/><\/figure><\/div>\n\n\n\n<p><strong>\u8bed\u97f3\u5206\u8bcd\u5668\u4ee5 25 Hz \u7684\u4ee4\u724c\u901f\u7387\u5de5\u4f5c\uff0c\u5373\u6bcf\u79d2 25 \u4e2a\u8bed\u97f3\u4ee4\u724c\u3002<\/strong><\/p>\n\n\n\n<h4>Unified Text-Speech Language Model<\/h4>\n\n\n\n<p>CosyVoice 2 \u4e2d\uff0c\u4f7f\u7528\u9884\u8bad\u7ec3\u7684\u6587\u672c<strong>Qwen2.5-0.5B<\/strong>&nbsp;\u4f5c\u4e3a\u6587\u672c-\u8bed\u97f3\u8bed\u8a00\u6a21\u578b\uff0c\u4ee5\u8f93\u5165\u6587\u672c\u4f5c\u4e3a\u63d0\u793a\u81ea\u52a8\u56de\u5f52\u751f\u6210\u8bed\u97f3\u6807\u8bb0\u3002\u4e0e\u5176\u4ed6 LM \u7c7b\u4f3c\uff0c\u6587\u672c\u8bed\u97f3 LM \u4e5f\u91c7\u7528\u4e0b\u4e00\u4e2a\u6807\u8bb0\u9884\u6d4b\u65b9\u6848\u8fdb\u884c\u8bad\u7ec3\uff0c\u5982\u56fe&nbsp;1\uff08b\uff09 \u6240\u793a\u3002\u4e0e\u4e4b\u524d\u7684 CosyVoice \u4e0d\u540c\uff0c\u6211\u4eec<strong>\u53bb\u9664\u4e86\u8bf4\u8bdd\u4eba\u5d4c\u5165<\/strong>\uff0c\u4ee5\u907f\u514d\u4fe1\u606f\u6cc4\u6f0f\u3002\u66f4\u91cd\u8981\u7684\u662f\uff0c\u6211\u4eec\u53d1\u73b0\u8fd9\u79cd\u8bdd\u8bed\u7ea7\u522b\u7684\u5411\u91cf\u4e0d\u4ec5\u5305\u542b\u8bf4\u8bdd\u4eba\u8eab\u4efd\u4fe1\u606f\uff0c\u8fd8\u5305\u542b\u8bed\u8a00\u548c\u526f\u8bed\u8a00\u4fe1\u606f\uff0c\u8fd9\u4f1a\u5f71\u54cd\u6587\u672c-\u8bed\u97f3\u8bed\u8a00\u6a21\u578b\u7684\u97f5\u5f8b\u81ea\u7136\u6027\u548c\u8de8\u8bed\u8a00\u80fd\u529b\u3002\u6b64\u5916\uff0c\u6211\u4eec\u8fd8\u653e\u5f03\u4e86\u4e4b\u524d CosyVoice \u7684\u6587\u672c\u7f16\u7801\u5668\uff0c\u56e0\u4e3a\u6211\u4eec\u53d1\u73b0 Qwen2.5-0.5B \u6a21\u578b\u5df2\u7ecf\u8db3\u591f\u5f3a\u5927\uff0c\u53ef\u4ee5\u5bf9\u9f50\u6587\u672c\u548c\u8bed\u97f3\u6807\u8bb0\uff0c\u56e0\u6b64\u4e0d\u518d\u9700\u8981\u6587\u672c\u7f16\u7801\u5668\u3002<\/p>\n\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter size-full is-resized\"><img loading=\"lazy\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-81.png\" alt=\"\" class=\"wp-image-23325\" width=\"607\" height=\"351\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-81.png 742w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-81-300x174.png 300w\" sizes=\"(max-width: 607px) 100vw, 607px\" \/><\/figure><\/div>\n\n\n\n<p>\u53d7\u76ca\u4e8e\u6587\u672c\u8bed\u97f3 LM \u7684\u7b80\u5355\u6027\uff0c\u6211\u4eec\u53ef\u4ee5\u4e3a<strong>\u6d41\u5f0f\u548c\u975e\u6d41\u5f0f\u5408\u6210\u6784\u5efa\u4e00\u4e2a\u7edf\u4e00\u7684\u6a21\u578b<\/strong>\u3002\u5728\u8fd9\u91cc\uff0c<strong>\u201c\u6d41\u6a21\u5f0f\u201d\u610f\u5473\u7740\u8f93\u5165\u6587\u672c\u4ee5\u8fde\u7eed\u6d41\u7684\u5f62\u5f0f\u63a5\u6536\uff0c\u800c\u4e0d\u662f\u63d0\u524d\u88ab\u79f0\u4e3a\u5b8c\u6574\u7684\u53e5\u5b50<\/strong>\u3002\u5728 CosyVoice 2 \u4e2d\uff0c<strong>\u63a8\u6d41\u6a21\u5f0f\u548c\u975e\u63a8\u6d41\u6a21\u5f0f\u7684\u533a\u522b\u53ea\u662f LM \u7684\u5e8f\u5217\u6784\u5efa\u65b9\u5f0f<\/strong>\uff1a<\/p>\n\n\n\n<ul><li>\u5bf9\u4e8e\u975e\u6d41\u6a21\u5f0f\uff0c\u201c\u5e8f\u5217\u5f00\u59cb\u201dS\u3001\u6240\u6709\u6587\u672c\u6807\u8bb0\u3001\u201c\u8bed\u97f3\u8f6c\u6362\u201dT\u6807\u8bb0\u3001\u6240\u6709\u8bed\u97f3\u6807\u8bb0\u548c\u201c\u5e8f\u5217\u7ed3\u675f\u201dE\u6309\u987a\u5e8f\u8fde\u63a5\uff0c\u5982\u56fe 2 \u5e95\u90e8\u6240\u793a\u3002Ignore \u6807\u8bb0\u8868\u793a\u5ffd\u7565\u5b83\u4eec\u7684\u635f\u5931\uff0c\u540c\u65f6\u6700\u5c0f\u5316\u4ea4\u53c9\u71b5\u76ee\u6807\u51fd\u6570\u3002<\/li><li>\u5728<strong>\u6d41\u5f0f\u6a21\u5f0f\u4e0b\uff0c<\/strong>\u6211\u4eec\u5c06\u6587\u672c\u548c\u8bed\u97f3\u6807\u8bb0\u6309\u9884\u5b9a\u4e49\u7684\u6bd4\u4f8b N:M \u6df7\u5408\uff0c\u5373\u6bcf N \u4e2a\u6587\u672c\u6807\u8bb0\u540e\u8ddf\u7740 M \u4e2a\u8bed\u97f3\u6807\u8bb0\uff0c\u5982\u56fe 2 \u7684\u9876\u90e8\u6240\u793a\u3002\u5982\u679c\u4e0b\u4e00\u4e2a\u6807\u8bb0\u662f\u6587\u672c\u6807\u8bb0\uff0c\u6a21\u578b\u4f1a\u9884\u6d4b\u4e00\u4e2a\u586b\u5145\u6807\u8bb0\uff08\u800c\u4e0d\u662f\u6587\u672c\u6807\u8bb0\uff09\uff0c\u8be5<strong>\u586b\u5145\u6807\u8bb0\u8868\u793a\u63a5\u4e0b\u6765\u7684 N\u4e2a\u6587\u672c\u6807\u8bb0\u5e94\u8be5\u5728\u63a8\u7406\u9636\u6bb5\u8fdb\u884c\u8fde\u63a5<\/strong>\u3002\u3010\u65b9\u4fbf\u63a8\u7406\u65f6\u5019\u83b7\u53d6\u8f93\u51fa\u7684\u8bed\u4e49token\u3011\uff0c\u5f53\u6587\u672c\u6807\u8bb0\u7528\u5c3d\u65f6\uff0c\u201c\u8bed\u97f3\u8f6e\u6b21\u201d\u6807\u8bb0 T\u548c\u5269\u4f59\u7684\u8bed\u97f3\u6807\u8bb0\u4f1a\u88ab\u987a\u5e8f\u8fde\u63a5\uff0c\u5f62\u6210\u6d41\u5f0f\u6a21\u5f0f\u4e0b\u7684\u6df7\u5408\u6587\u672c-\u8bed\u97f3\u6807\u8bb0\u5e8f\u5217\u3002<\/li><\/ul>\n\n\n\n<p>\u901a\u8fc7<strong>\u540c\u65f6\u5728\u4e0a\u8ff0\u4e24\u4e2a\u5e8f\u5217\u4e0a\u8bad\u7ec3\u6587\u672c-\u8bed\u97f3 LM<\/strong>\uff0c\u6211\u4eec\u53ef\u4ee5\u5728\u5355\u4e2a\u7edf\u4e00\u6a21\u578b\u4e2d\u6267\u884c\u6d41\u5f0f\u548c\u975e\u6d41\u5f0f\u8bed\u97f3\u751f\u6210\u3002\u5728\u5b9e\u9645\u573a\u666f\u4e2d\uff0c\u4f8b\u5982<strong>\u8bf4\u8bdd\u4eba\u5fae\u8c03 \uff08SFT\uff09 \u548c\u4e0a\u4e0b\u6587\u5b66\u4e60 \uff08ICL\uff09<\/strong>\uff0c\u63a8\u7406\u5e8f\u5217\u6709\u6240\u4e0d\u540c\uff0c\u5982\u4e0b\u6240\u793a\uff1a<\/p>\n\n\n\n<p>&nbsp;<strong>ICL\uff0c\u975e\u6d41\u5f0f<\/strong>\uff1a\u5728 ICL \u4e2d\uff0c<strong>LM \u9700\u8981\u6765\u81ea\u53c2\u8003\u97f3\u9891\u7684\u63d0\u793a\u6587\u672c\u548c\u8bed\u97f3\u6807\u8bb0<\/strong>\uff0c<strong>\u4ee5\u6a21\u4eff\u91cd\u97f3\u3001\u97f5\u5f8b\u3001\u60c5\u611f\u548c\u98ce\u683c\u3002<\/strong>\u5728\u975e\u6d41\u5f0f\u5904\u7406\u6a21\u5f0f\u4e0b\uff0c\u63d0\u793a\u548c\u8981\u5408\u6210\u7684\u6587\u672c\u6807\u8bb0\u8fde\u63a5\u4e3a\u6574\u4e2a\u5b9e\u4f53\uff0c\u63d0\u793a\u8bed\u97f3\u6807\u8bb0\u88ab\u89c6\u4e3a\u9884\u5148\u751f\u6210\u7684\u7ed3\u679c\u5e76\u56fa\u5b9a\uff1a\u201cS \u3001 prompt_text\u3001 text \u3001T\u3001 prompt_speech\u201d\u3002LM \u7684\u81ea\u56de\u5f52\u751f\u6210\u4ece\u6b64\u7c7b\u5e8f\u5217\u5f00\u59cb\uff0c\u76f4\u5230\u68c0\u6d4b\u5230 \u201cEnd of sequence\u201d \u6807\u8bb0\u3002<\/p>\n\n\n\n<p>&nbsp;<strong>ICL\uff0c\u6d41\u5f0f\u5904\u7406<\/strong>\uff1a\u5728\u6b64\u65b9\u6848\u4e2d\uff0c\u6211\u4eec<strong>\u5047\u8bbe\u8981\u751f\u6210\u7684\u6587\u672c\u662f\u5df2\u77e5\u7684<\/strong>\uff0c\u5e76\u4e14<strong>\u8bed\u97f3\u4ee4\u724c\u5e94\u4ee5\u6d41\u5f0f\u5904\u7406\u65b9\u5f0f\u751f\u6210<\/strong>\u3002\u540c\u6837\uff0c\u6211\u4eec\u5c06 prompt \u548c to-generate \u6587\u672c\u89c6\u4e3a\u4e00\u4e2a\u6574\u4f53\u3002\u7136\u540e\uff0c\u6211\u4eec\u5c06\u5176\u4e0e\u63d0\u793a\u8bed\u97f3\u6807\u8bb0\u6df7\u5408\uff0c\u6bd4\u4f8b\u4e3a&nbsp;N&nbsp;\uff1a&nbsp;M&nbsp;\uff1a \u201cS\uff0c mixed_text_speech\uff0cT\uff0cremaining_speech\u201d\u3002\u5982\u679c\u6587\u672c\u957f\u5ea6\u5927\u4e8e\u63d0\u793a\u8bed\u97f3 Token \u7684\u957f\u5ea6\uff0cLM \u5c06\u751f\u6210 \u201cfilling token\u201d\u3002\u5728\u8fd9\u79cd\u60c5\u51b5\u4e0b\uff0c\u6211\u4eec<strong>\u624b\u52a8\u586b\u5145&nbsp;N\u4e2a\u6587\u672c\u6807\u8bb0<\/strong>\u3002\u5982\u679c\u6587\u672c\u4ee4\u724c\u7528\u5b8c\uff0c\u5c06\u6dfb\u52a0\u201cTurn of speech\u201d T \u4ee4\u724c\u3002\u5728\u6d41\u5f0f\u5904\u7406\u6a21\u5f0f\u4e0b\uff0c\u6211\u4eec\u4e3a\u6bcf\u4e2a&nbsp;M&nbsp;\u4ee4\u724c\u8fd4\u56de\u751f\u6210\u7ed3\u679c\uff0c\u76f4\u5230\u68c0\u6d4b\u5230 E\u4e3a\u6b62\u3002<\/p>\n\n\n\n<p><strong>SFT\uff0c\u975e\u6d41\u5f0f<\/strong>\uff1a\u5728 SFT \u573a\u666f\u4e2d\uff0c<strong>LM \u9488\u5bf9\u7279\u5b9a\u8bf4\u8bdd\u4eba\u8fdb\u884c\u5fae\u8c03\uff0c\u4e0d\u518d\u9700\u8981\u63d0\u793a\u6587\u672c\u548c\u8bed\u97f3\u3002<\/strong>\u56e0\u6b64\uff0c\u521d\u59cb\u5e8f\u5217\u975e\u5e38\u7b80\u5355\uff1a\u201c S\uff0c \u6587\u672c \uff0cT \u201d\u3002\u4ece\u6b64\u5f00\u59cb\uff0c\u6587\u672c-\u8bed\u97f3 LM \u53ef\u4ee5\u81ea\u52a8\u56de\u5f52\u751f\u6210\u8bed\u97f3\u6807\u8bb0\uff0c\u76f4\u5230 E\u7ed3\u675f\u3002<\/p>\n\n\n\n<p><strong>SFT, Streaming:<\/strong>&nbsp;\u5728 SFT \u7684\u6d41\u6a21\u5f0f\u4e0b\uff0c\u6211\u4eec\u4ece\u4ee5\u4e0b\u5e8f\u5217\u5f00\u59cb\u8bed\u97f3\u751f\u6210\uff1a\u201c S,first_N_text\u201d\u3002\u7136\u540e\uff0cLM \u5c06\u751f\u6210&nbsp;M&nbsp;\u8bed\u97f3\u4ee4\u724c\uff0c\u6211\u4eec\u624b\u52a8\u586b\u5145\u4e0b\u4e00N\u4e2a&nbsp;\u6587\u672c\u4ee4\u724c\u3002\u6211\u4eec\u91cd\u590d\u4e0a\u8ff0\u8fc7\u7a0b\uff0c\u76f4\u5230\u6240\u6709\u6587\u672c\u6807\u8bb0\u90fd\u7528\u5b8c\uff0c\u7136\u540e\u6dfb\u52a0 T \u3002\u6ce8\u610f\uff0cspeech-to-speech\u591a\u6a21\u6001\u5927\u8bed\u8a00\u6a21\u578b\u4e5f\u53ef\u4ee5\u91c7\u7528\u8fd9\u79cd\u6a21\u5f0f\uff0c\u4ee5\u83b7\u5f97\u6781\u4f4e\u7684\u5ef6\u8fdf\u3002<\/p>\n\n\n\n<h4>Chunk-aware Flow Matching<\/h4>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"907\" height=\"523\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-82.png\" alt=\"\" class=\"wp-image-23360\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-82.png 907w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-82-300x173.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-82-768x443.png 768w\" sizes=\"(max-width: 907px) 100vw, 907px\" \/><\/figure>\n\n\n\n<p><strong>\u91c7\u7528\u6885\u5c14\u9891\u8c31\u56fe\u4f5c\u4e3a\u58f0\u5b66\u7279\u5f81\uff0c\u5e27\u7387\u4e3a 50 Hz<\/strong>\uff0c<strong>\u91c7\u6837\u7387\u4e3a 24000<\/strong>\u3002\u7531\u4e8e\u8bed\u97f3\u4ee4\u724c\u548c Mel \u7279\u5f81\u4e4b\u95f4\u7684\u5e27\u901f\u7387\u4e0d\u5339\u914d\uff0c\u6211\u4eec\u4ee5 <strong>2 \u7684\u6bd4\u7387\u5bf9\u8bed\u97f3\u4ee4\u724c\u8fdb\u884c\u4e0a\u91c7\u6837<\/strong>\uff0c<strong>\u4ee5\u5339\u914d Mel \u9891\u8c31\u56fe\u7684\u5e27\u901f\u7387\u3002<\/strong>\u5728\u4e0a\u91c7\u6837\u64cd\u4f5c\u4e4b\u524d\uff0c\u6211\u4eec<strong>\u6dfb\u52a0\u4e86\u4e00\u4e2a\u989d\u5916\u7684\u524d\u77bb\u5377\u79ef\u5c42\uff0c\u4ee5\u4fbf\u4e3a\u4ee5\u4e0b\u56e0\u679c\u6a21\u5757\u63d0\u4f9b\u672a\u6765\u4fe1\u606f\u3002\u524d\u77bb\u5c42\u7531\u53f3\u586b\u5145\u7684\u4e00\u7ef4\u5377\u79ef\u5b9e\u73b0\uff0c\u5176\u586b\u5145\u5927\u5c0f\u4e3a &nbsp;P\uff0c&nbsp;\u5377\u79ef\u6838\u5927\u5c0f\u4e3a&nbsp;P+1&nbsp;\u3002\u5728\u6b64\u4e4b\u540e\uff0c\u51e0\u4e2achunk-aware causal Transformer blocks\u6765\u5bf9\u9f50\u8bed\u97f3\u6807\u8bb0\u7684\u8868\u793a\u7a7a\u95f4\u4ee5\u5339\u914d\u58f0\u5b66\u7279\u5f81\u3002<\/strong><\/p>\n\n\n\n<p>\u968f\u540e\uff0c\u6211\u4eec\u7684<strong>\u76ee\u6807\u662f\u5c06\u8bed\u97f3\u6807\u8bb0\u8fdb\u4e00\u6b65\u89e3\u7801\u4e3a\u7531\u8bf4\u8bdd\u4eba\u5d4c\u5165\u548c\u53c2\u8003\u8bed\u97f3\u6307\u5b9a\u7684 Mel \u9891\u8c31\u56fe<\/strong>\u3002 \u4e3a\u4e86\u5b9e\u73b0\u8fd9\u4e00\u76ee\u6807\uff0c\u6211\u4eec\u91c7\u7528\u6761\u4ef6\u6d41\u5339\u914d \uff08CFM\uff09 \u6a21\u578b\u5bf9 Mel \u9891\u8c31\u56fe\u8fdb\u884c\u91c7\u6837\uff0c<strong>\u7ed9\u5b9a\u8bed\u97f3\u6807\u8bb0\u3001\u53c2\u8003\u8bed\u97f3\u548c\u8bf4\u8bdd\u4eba\u5d4c\u5165\u4f5c\u4e3a\u6761\u4ef6<\/strong>\u3002 \u5728 CFM \u6a21\u578b\u4e2d\uff0c\u76ee\u6807 Mel \u4e09\u7ef4\u9891\u8c31\u56fe\u7684\u5206\u5e03\u7531\u6765\u81ea\u5148\u9a8c\u5206\u5e03&nbsp;p0\u2062(X)&nbsp;\u548c\u6570\u636e\u5206\u5e03&nbsp;q\u2062(X)&nbsp;\u7684\u6982\u7387\u5bc6\u5ea6\u8def\u5f84 \u6765\u63cf\u8ff0\u3002\u6982\u7387\u5bc6\u5ea6\u8def\u5f84\u53ef\u4ee5\u7531\u77ac\u6001\u5411\u91cf\u573a\u5b9a\u4e49\u3002 \u4e3a\u4e86\u63d0\u9ad8\u91c7\u6837\u6548\u7387\uff0c\u6211\u4eec\u91c7\u7528\u6700\u4f73\u4f20\u8f93 \uff08OT\uff09 \u6d41\u6765\u5339\u914d\u77e2\u91cf\u573a&nbsp;\u3002<\/p>\n\n\n\n<p>\u5728\u8bad\u7ec3\u9636\u6bb5\uff0c\u63a9\u7801 Mel \u9891\u8c31\u56fe\u662f\u901a\u8fc7\u968f\u673a\u63a9\u7801 70% \u5230 100% \u7684\u6700\u7ec8\u5e27\u6765\u83b7\u5f97\u7684&nbsp;X1&nbsp;\u3002\u81f3\u4e8e\u63a8\u8bba\uff0c\u5b83\u7531\u4ece\u53c2\u8003\u8bed\u97f3\u4e2d\u63d0\u53d6\u7684 Mel \u9891\u8c31\u56fe\u63d0\u4f9b\u3002 \u901a\u8fc7\u6700\u5c0f\u5316\u9884\u6d4b ODE \u548c\u771f\u5b9e ODE \u4e4b\u95f4\u7684 L1 \u635f\u5931\uff0c\u6211\u4eec\u53ef\u4ee5\u6309\u5982\u4e0b\u65b9\u5f0f\u4f18\u5316 UNet \u53c2\u6570&nbsp;\u03b8&nbsp;\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"843\" height=\"58\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-83.png\" alt=\"\" class=\"wp-image-23369\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-83.png 843w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-83-300x21.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-83-768x53.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-83-825x58.png 825w\" sizes=\"(max-width: 843px) 100vw, 843px\" \/><\/figure>\n\n\n\n<p><strong>\u56e0\u679c\u6d41\u5339\u914d\u6a21\u578b\uff1a\u5f53\u524d\u7684\u6d41\u5339\u914d\u6a21\u578b\u603b\u662f\u4ee5\u79bb\u7ebf\u6a21\u5f0f\u5de5\u4f5c\uff0c\u5373\u53ea\u6709\u5728\u751f\u6210\u6240\u6709\u8bed\u97f3\u6807\u8bb0\u540e\uff0cMel\u8c31\u56fe\u624d\u80fd\u88ab\u91c7\u6837\uff0c\u8fd9\u5bf9\u4e8e\u6d41\u5f0f\u5408\u6210\u5e76\u4e0d\u53cb\u597d\u3002<\/strong>\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\uff0c\u6211\u4eec\u5c06<strong>\u591a\u6b65\u6d41\u4f30\u8ba1\u89c6\u4e3a\u4e00\u4e2a\u5806\u53e0\u7684\u6df1\u5ea6\u795e\u7ecf\u7f51\u7edc\uff0c\u8be5\u7f51\u7edc\u91cd\u590d\u4f7f\u7528 UNet \u5341\u6b21\u3002<\/strong>\u56e0\u6b64\uff0c<strong>\u901a\u8fc7\u4f7f\u5c55\u5f00\u7684\u795e\u7ecf\u7f51\u7edc\u5177\u6709\u56e0\u679c\u6027\uff0c\u6211\u4eec\u53ef\u4ee5\u5c06\u5176\u5e94\u7528\u4e8e\u6d41\u5f0f\u5408\u6210\u3002<\/strong>\u6211\u4eec\u6784\u9020\u4e86\u56db\u4e2a\u63a9\u7801\uff0c\u4ee5\u6ee1\u8db3\u4e0d\u540c\u7684\u5e94\u7528\u60c5\u51b5\uff1a<\/p>\n\n\n\n<p><strong>Non-causal Mask<\/strong>&nbsp;<\/p>\n\n\n\n<p>\u975e\u56e0\u679c\u63a9\u7801 \uff08Non-causal Mask\uff09 \u7528\u4e8e\u79bb\u7ebf\u6a21\u5f0f\uff0c\u901a\u8fc7\u6ee1\u8db3\u6240\u6709\u6761\u4ef6\u5e27\u6765\u5b9e\u73b0\u6700\u4f73\u6027\u80fd\u3002\u975e\u56e0\u679c\u63a9\u7801\u9002\u7528\u4e8e\u5bf9\u5ef6\u8fdf\u4e0d\u654f\u611f\u7684\u60c5\u51b5\u3002<\/p>\n\n\n\n<p><strong> Full-causal Mask<\/strong> <\/p>\n\n\n\n<p>&nbsp;\u5168\u56e0\u679c\u63a9\u7801 \uff08Full-causal Mask\uff09 \u9002\u7528\u4e8e\u9700\u8981\u6781\u4f4e\u5ef6\u8fdf\u7684\u573a\u666f\uff0c\u5176\u4e2d\u53ea\u80fd\u89c2\u770b\u8fc7\u53bb\u7684\u5e27\u3002<\/p>\n\n\n\n<p><strong>Chunk-M&nbsp;Mask<\/strong><\/p>\n\n\n\n<p>\u53ef\u4ee5\u901a\u8fc7\u727a\u7272\u66f4\u591a\u5ef6\u8fdf\u6765\u5b9e\u73b0\u79bb\u7ebf\u6a21\u5f0f\u7684\u8fd1\u4f3c\u6027\u80fd\uff0c\u8fd9\u53ef\u7528\u4e8e\u7ea7\u8054\u751f\u6210 Chunk \u4ee5\u83b7\u5f97\u66f4\u597d\u7684\u6027\u80fd\u3002<\/p>\n\n\n\n<p><strong>Chunk-<\/strong>2\u2062M<strong>&nbsp;Mask<\/strong><\/p>\n\n\n\n<p>\u53ef\u4ee5\u901a\u8fc7\u727a\u7272\u66f4\u591a\u5ef6\u8fdf\u6765\u5b9e\u73b0\u79bb\u7ebf\u6a21\u5f0f\u7684\u8fd1\u4f3c\u6027\u80fd\uff0c\u8fd9\u53ef\u7528\u4e8e\u7ea7\u8054\u751f\u6210 Chunk \u4ee5\u83b7\u5f97\u66f4\u597d\u7684\u6027\u80fd\u3002<\/p>\n\n\n\n<p>\u5bf9\u4e8e\u5c0f\u6279\u91cf\u4e2d\u7684\u6bcf\u4e2a\u8bad\u7ec3\u6848\u4f8b\uff0c\u6211\u4eec\u4ece\u5747\u5300\u5206\u5e03\u4e0b\u7684\u4e0a\u8ff0\u56db\u4e2a\u63a9\u7801\u4e2d\u968f\u673a\u91c7\u6837\u4e00\u4e2a\u63a9\u7801\u3002\u8fd9\u6837\uff0c\u4e00\u4e2a\u6d41\u5339\u914d\u6a21\u578b\u53ef\u4ee5\u517c\u5bb9\u4e0d\u540c\u7684\u573a\u666f\uff0c\u964d\u4f4e\u90e8\u7f72\u590d\u6742\u5ea6\u3002\u8fd9\u79cd\u5757\u611f\u77e5\u8bad\u7ec3\u7684\u53e6\u4e00\u4e2a\u4f18\u70b9\u662f\uff0c\u5177\u6709\u66f4\u591a\u4e0a\u4e0b\u6587\u7684\u63a9\u7801\u53ef\u4ee5\u4f5c\u4e3a\u5177\u6709\u8f83\u5c11\u4e0a\u4e0b\u6587\u7684\u63a9\u7801\u7684\u6559\u5e08\uff0c\u53d7\u76ca\u4e8e\u9690\u542b\u7684\u81ea\u6211\u84b8\u998f\u65b9\u6848\u3002<\/p>\n\n\n\n<h4>\u6d41\u5f0f\u5904\u7406\u6a21\u5f0f\u7684\u5ef6\u8fdf\u5206\u6790<\/h4>\n\n\n\n<p>\u9996\u5305\u5ef6\u8fdf\u662f\u6d41\u5f0f\u5408\u6210\u6a21\u578b\u7684\u4e00\u4e2a\u91cd\u8981\u6307\u6807\uff0c\u5b83\u663e\u7740\u5f71\u54cd\u7528\u6237\u4f53\u9a8c\uff0c\u5c24\u5176\u662f\u5728\u57fa\u4e8e LLM \u7684\u8bed\u97f3\u804a\u5929\u5e94\u7528\u7a0b\u5e8f\u4e2d\uff0c\u4f8b\u5982 GPT-4o&nbsp;\u3002&nbsp;\u5728 TTS \u7684\u4e0a\u4e0b\u6587\u4e2d\uff0c\u8981\u5408\u6210\u7684\u6587\u672c\u662f\u4e8b\u5148\u5df2\u77e5\u7684\uff0c\u5ef6\u8fdf\u6765\u81ea\u8bed\u97f3\u4ee4\u724c\u751f\u6210\u3001\u6885\u5c14\u9891\u8c31\u56fe\u91cd\u5efa\u548c\u6ce2\u5f62\u5408\u6210\u7b49\u65b9\u9762\u3002 \u56e0\u6b64\uff0cCosyVoice 2 \u7684\u9996\u5305\u5ef6\u8fdf&nbsp;LT\u2062T\u2062S&nbsp;\u53ef\u4ee5\u5f97\u5230\u5982\u4e0b\uff1a<\/p>\n\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter size-full is-resized\"><img loading=\"lazy\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-84.png\" alt=\"\" class=\"wp-image-23383\" width=\"319\" height=\"26\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-84.png 442w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-84-300x24.png 300w\" sizes=\"(max-width: 319px) 100vw, 319px\" \/><\/figure><\/div>\n\n\n\n<p>\u5176\u4e2d&nbsp;dl\u2062m&nbsp;\uff0c\u8868\u793a LM \u751f\u6210\u4e00\u4e2a\u8bed\u97f3\u8bcd\u5143\u7684\u8ba1\u7b97\u65f6\u95f4\uff0c&nbsp;df\u2062m&nbsp;\u8868\u793a Flow Matching \u6a21\u578b\u751f\u6210\u4e00\u4e2a\u8bed\u97f3\u8bcd\u5143\u7684\u6885\u5c14\u9891\u8c31\u56fe\u5e27\u7684\u8ba1\u7b97\u65f6\u95f4\uff0c&nbsp;dv\u2062o\u2062c&nbsp;\u8868\u793a\u58f0\u7801\u5668\u5408\u6210\u4e00\u4e2a\u8bed\u97f3\u8bcd\u5143\u5bf9\u5e94\u7684\u6ce2\u5f62\u7684\u8ba1\u7b97\u65f6\u95f4\u3002 <strong>\u5728\u57fa\u4e8e LLM \u7684\u8bed\u97f3\u804a\u5929\u4e0a\u4e0b\u6587\u4e2d\uff0c\u8fd8\u5e94\u8003\u8651 first-package-required \u6587\u672c\u7684\u957f\u5ea6\uff0cfirst-package \u5ef6\u8fdf&nbsp;LC\u2062h\u2062a\u2062t&nbsp;\u5982\u4e0b<\/strong>\uff1a<\/p>\n\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter size-full is-resized\"><img loading=\"lazy\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-85.png\" alt=\"\" class=\"wp-image-23386\" width=\"219\" height=\"31\"\/><\/figure><\/div>\n\n\n\n<p>\u5176\u4e2d&nbsp;dl\u2062l\u2062m&nbsp;\u8868\u793a LLM \u751f\u6210\u4e00\u4e2a\u6587\u672c Token \u7684\u8ba1\u7b97\u65f6\u95f4\u3002\u8bf7\u6ce8\u610f\uff0c\u7531\u4e8e\u591a\u5b57\u7b26\u6807\u8bb0\u5728 CosyVoice 2 \u7684\u6587\u672c\u5206\u8bcd\u5668\u4e2d\u88ab\u5c4f\u853d\uff0c\u56e0\u6b64\u6587\u672c LLMs\u603b\u662f\u6bd4 CosyVoice 2 \u4f7f\u7528\u7684\u6587\u672c\u6807\u8bb0\u7f16\u7801\u66f4\u957f\u7684\u539f\u59cb\u6587\u672c\u3002\u56e0\u6b64\uff0c\u7b2c\u4e00\u4e2a\u5305\u7684\u5ef6\u8fdf&nbsp;LC\u2062h\u2062a\u2062t&nbsp;\u5fc5\u987b\u4f4e\u4e8e&nbsp;N\u22c5dl\u2062l\u2062m&nbsp;\u548c &nbsp;LT\u2062T\u2062S\u7684\u548c\u3002<\/p>\n\n\n\n<h4>Instructed Generation<\/h4>\n\n\n\n<p>\u4e3a\u4e86\u589e\u5f3a CosyVoice 2 \u7684\u53ef\u63a7\u6027\uff0c\u6211\u4eec\u5c06 indirected \u6570\u636e\u96c6\u96c6\u6210\u5230\u57fa\u7840\u8bad\u7ec3\u96c6\u4e2d\u3002\u6211\u4eec\u6536\u96c6\u4e86 1500 \u5c0f\u65f6\u7684\u5b9a\u5411\u8bad\u7ec3\u6570\u636e\uff0c\u5176\u4e2d\u5305\u62ec\u81ea\u7136\u8bed\u8a00\u6307\u4ee4\u548c\u7ec6\u7c92\u5ea6\u6307\u4ee4\uff0c\u5982\u8868\u6240\u793a\u3002\u5bf9\u4e8e\u81ea\u7136\u8bed\u8a00\u6307\u4ee4\uff0c\u6211\u4eec\u5728\u8981\u5408\u6210\u7684\u8f93\u5165\u6587\u672c\u4e4b\u524d\u9884\u7f6e\u81ea\u7136\u8bed\u8a00\u63cf\u8ff0\u548c\u7279\u6b8a\u7ed3\u675f\u6807\u8bb0\u201c&lt;|endofprompt|&gt;\u201d\u3002\u8fd9\u4e9b\u63cf\u8ff0\u6db5\u76d6\u60c5\u611f\u3001\u8bed\u901f\u3001\u89d2\u8272\u626e\u6f14\u548c\u65b9\u8a00\u7b49\u65b9\u9762\u3002\u5bf9\u4e8e\u7ec6\u7c92\u5ea6\u7684\u6307\u4ee4\uff0c\u6211\u4eec\u5728\u6587\u672c\u6807\u8bb0\u4e4b\u95f4\u63d2\u5165\u4eba\u58f0\u7206\u53d1\uff0c\u4f7f\u7528\u201c[laughter]\u201d\u548c\u201c[breath]\u201d\u7b49\u6807\u8bb0\u3002\u6b64\u5916\uff0c\u6211\u4eec\u5c06 vocal feature \u6807\u7b7e\u5e94\u7528\u4e8e\u77ed\u8bed;\u4f8b\u5982\uff0c\u201c&lt;strong&gt;XXX&lt;\/strong&gt;\u201d\u8868\u793a\u5f3a\u8c03\u67d0\u4e9b\u8bcd\uff0c\u800c\u201c&lt;laughter&gt;XXX&lt;\/laughter&gt;\u201d\u8868\u793a\u7b11\u58f0\u8bf4\u8bdd\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"1017\" height=\"601\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-86.png\" alt=\"\" class=\"wp-image-23395\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-86.png 1017w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-86-300x177.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-86-768x454.png 768w\" sizes=\"(max-width: 1017px) 100vw, 1017px\" \/><\/figure>\n\n\n\n<h4>Multi-Speaker Fine-tuning<\/h4>\n\n\n\n<p>\u5728\u7279\u5b9a\u8bf4\u8bdd\u4eba \uff08SFT\uff09 \u4e0a\u5fae\u8c03\u9884\u8bad\u7ec3\u6a21\u578b\u53ef\u4ee5\u8fdb\u4e00\u6b65\u63d0\u9ad8\u751f\u6210\u8d28\u91cf\u548c\u8bf4\u8bdd\u4eba\u76f8\u4f3c\u5ea6\u3002\u5728\u672c\u62a5\u544a\u4e2d\uff0c\u6211\u4eec\u4ecb\u7ecd\u4e86\u591a\u626c\u58f0\u5668\u5fae\u8c03 \uff08mSFT\uff09\uff0c<strong>\u5176\u4e2d\u9884\u8bad\u7ec3\u6a21\u578b\u540c\u65f6\u5728\u591a\u626c\u58f0\u5668\u4e0a\u8fdb\u884c\u5fae\u8c03\uff0c\u800c\u4e0d\u662f\u5728\u5355\u4e2a\u626c\u58f0\u5668\u4e0a\u8fdb\u884c\u5fae\u8c03\u3002\u8fd9\u79cd\u65b9\u6cd5\u53ef\u786e\u4fdd\u8de8\u591a\u4e2a\u8bf4\u8bdd\u4eba\u7684\u5168\u9762\u97f5\u5f8b\u548c\u53d1\u97f3\u8986\u76d6\uff0c\u5e76\u51cf\u5c11\u9884\u8bad\u7ec3\u6a21\u578b\u53ef\u80fd\u51fa\u73b0\u7684\u707e\u96be\u6027\u9057\u5fd8\u3002<\/strong>\u4e3a\u907f\u514d\u4e0d\u540c\u8bf4\u8bdd\u4eba\u4e4b\u95f4\u7684\u97f3\u8272\u6df7\u6dc6\uff0c\u6211\u4eec<strong>\u5728\u7279\u5b9a\u8bf4\u8bdd\u4eba\u7684\u8f93\u5165\u6587\u672c\u524d\u52a0\u4e0a\u8bf4\u8bdd\u4eba\u63d0\u793a\u6807\u8bb0\u201cSpeaker A&lt;|endofprompt|&gt;\u201d\u3002\u5982\u679c\u8bad\u7ec3\u6837\u672c\u672a\u6807\u8bb0\u4e3a\u8bf4\u8bdd\u4eba\uff0c\u5219\u4f7f\u7528\u7279\u6b8a\u6807\u7b7e\u201cunknown&lt;|endofprompt|&gt;\u201d\u3002\u5728\u6574\u4e2a\u591a\u8bf4\u8bdd\u4eba\u5fae\u8c03\u8fc7\u7a0b\u4e2d\uff0c\u5b66\u4e60\u7387\u8bbe\u7f6e\u4e3a 1e-5\u3002<\/strong><\/p>\n\n\n\n<h4>Reinforcement Learning for SFT<\/h4>\n\n\n\n<p>\u5f3a\u5316\u5b66\u4e60\u662f\u5927\u578b\u8bed\u8a00\u6a21\u578b\u8bad\u7ec3\u4e2d\u5e38\u7528\u7684\u65b9\u6cd5\uff0c\u5b83\u53ef\u4ee5\u4f7f LM \u8f93\u51fa\u4e0e\u4eba\u7c7b\u504f\u597d\u4fdd\u6301\u4e00\u81f4\u3002 \u5728 CosyVoice 2 \u4e2d\uff0c\u6211\u4eec\u91c7\u7528 ASR \u7cfb\u7edf\u7684\u8bf4\u8bdd\u4eba\u76f8\u4f3c\u5ea6 \uff08SS\uff09 \u548c\u8bc6\u522b\u8bcd\u9519\u8bef\u7387 \uff08WER\uff09 \u4f5c\u4e3a\u5956\u52b1\u51fd\u6570\uff0c\u4ee5\u63d0\u9ad8\u5fae\u8c03\u9636\u6bb5\u7684\u8bf4\u8bdd\u4eba\u76f8\u4f3c\u5ea6\u548c\u53d1\u97f3\u51c6\u786e\u6027\u3002\u6211\u4eec\u4f7f\u7528 WER \u548c SS \u6765\u533a\u5206\u9996\u9009\u6837\u54c1&nbsp;xw&nbsp;\u548c\u4e0d\u5408\u683c\u6837\u54c1&nbsp;xl&nbsp;\uff0c\u5e76\u901a\u8fc7\u76f4\u63a5\u504f\u597d\u4f18\u5316 \uff08DPO\uff09 &nbsp;\u4f18\u5316 TTS \u7cfb\u7edf\uff0c\u5982\u4e0b\u6240\u793a\uff1a<\/p>\n\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter size-full is-resized\"><img loading=\"lazy\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-87.png\" alt=\"\" class=\"wp-image-23402\" width=\"477\" height=\"49\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-87.png 723w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-87-300x32.png 300w\" sizes=\"(max-width: 477px) 100vw, 477px\" \/><\/figure><\/div>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"958\" height=\"294\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-88.png\" alt=\"\" class=\"wp-image-23405\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-88.png 958w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-88-300x92.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-88-768x236.png 768w\" sizes=\"(max-width: 958px) 100vw, 958px\" \/><\/figure>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"960\" height=\"64\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-89.png\" alt=\"\" class=\"wp-image-23407\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-89.png 960w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-89-300x20.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-89-768x51.png 768w\" sizes=\"(max-width: 960px) 100vw, 960px\" \/><\/figure>\n\n\n\n<h3>Experimental<\/h3>\n\n\n\n<h4>Training Data for Speech Tokenizer<\/h4>\n\n\n\n<p>\u4e00\u4e2a 200000 \u5c0f\u65f6\u7684\u6570\u636e\u96c6\u7528\u4e8e\u8bad\u7ec3\u8bed\u97f3\u5206\u8bcd\u5668\uff0c\u5e76\u5c06\u89c4\u8303\u5316\u8f6c\u5f55\u4f5c\u4e3a\u6807\u7b7e\u3002\u8be6\u7ec6\u7684\u6570\u636e\u4fe1\u606f\u5982\u8868\u6240\u793a\u3002\u8bad\u7ec3\u6570\u636e\u6765\u81ea\u4e09\u79cd\u4e0d\u540c\u7684\u8d44\u6e90\uff1a\u5f00\u6e90 ASR \u6570\u636e\u96c6\u3001\u5185\u90e8\u5de5\u4e1a\u6570\u636e\u96c6\u548c TTS \u751f\u6210\u6570\u636e\u96c6\u3002\u867d\u7136\u6211\u4eec\u5728\u8bad\u7ec3\u8bed\u97f3\u5206\u8bcd\u5668\u65f6\u53ea\u4f7f\u7528\u4e86\u4e2d\u82f1\u6587\u6570\u636e\uff0c\u5982\u8868&nbsp;\u6240\u793a\uff0c\u4f46\u968f\u540e\u7684\u5b9e\u9a8c\u8868\u660e\uff0c<strong>\u8bed\u97f3\u5206\u8bcd\u5668\u5bf9\u5176\u4ed6\u8bed\u8a00\u5177\u6709\u96f6\u955c\u5934\u80fd\u529b\u3002\u5b83\u8fd8\u53ef\u7528\u4e8e\u65e5\u8bed\u548c\u97e9\u8bed\u7b49\u8bed\u8a00\u7684\u8bed\u97f3\u5408\u6210<\/strong>\u3002<\/p>\n\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter size-full\"><img loading=\"lazy\" width=\"556\" height=\"147\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-90.png\" alt=\"\" class=\"wp-image-23413\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-90.png 556w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-90-300x79.png 300w\" sizes=\"(max-width: 556px) 100vw, 556px\" \/><\/figure><\/div>\n\n\n\n<h4>Training Data for CosyVoice 2<\/h4>\n\n\n\n<p>CosyVoice 2 \u4e0e\u4e4b\u524d\u7684\u7248\u672c\u5171\u4eab\u76f8\u540c\u7684\u8bad\u7ec3\u6570\u636e\u3002\u6211\u4eec\u9996\u5148\u4f7f\u7528\u5185\u90e8\u8bed\u97f3\u5904\u7406\u5de5\u5177\u6536\u96c6\u7eaf\u8bed\u97f3\u6570\u636e\u3002\u968f\u540e\uff0cParaformer&nbsp;\u548c SenseVoice\u5206\u522b\u7528\u4e8e\u751f\u6210\u4e2d\u6587\u548c\u5176\u4ed6\u8bed\u8a00\u7684\u4f2a\u6587\u672c\u6807\u7b7e\u3002\u6211\u4eec\u8fd8\u91c7\u7528\u5185\u90e8\u529b\u5bf9\u9f50\u6a21\u578b\u6765\u8fc7\u6ee4\u6389\u4f4e\u8d28\u91cf\u7684\u6570\u636e\u5e76\u63d0\u9ad8\u6807\u70b9\u7b26\u53f7\u7684\u51c6\u786e\u6027\u3002\u8868 3 \u63d0\u4f9b\u4e86\u6570\u636e\u8be6\u7ec6\u4fe1\u606f\uff1a<\/p>\n\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter size-full\"><img loading=\"lazy\" width=\"514\" height=\"208\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-91.png\" alt=\"\" class=\"wp-image-23416\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-91.png 514w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-91-300x121.png 300w\" sizes=\"(max-width: 514px) 100vw, 514px\" \/><\/figure><\/div>\n\n\n\n<h4>Experimental Results<\/h4>\n\n\n\n<p><strong>Evaluations on Speech Tokenizer<\/strong><\/p>\n\n\n\n<p>\u7406\u60f3\u7684\u8bed\u97f3\u5206\u8bcd\u5668\u5e94\u8be5\u6709\u6548\u5730\u5229\u7528\u7801\u672c\uff0c\u4ee5\u9ad8\u4fdd\u771f\u5ea6\u4fdd\u7559\u4fe1\u606f\uff0c\u5e76\u5c55\u793a\u8bf4\u8bdd\u4eba\u7684\u72ec\u7acb\u6027\u3002\u5728\u8fd9\u90e8\u5206\uff0c\u6211\u4eec\u4ece\u56db\u4e2a\u65b9\u9762\u8bc4\u4f30\u6211\u4eec\u7684\u76d1\u7763\u8bed\u97f3\u5206\u8bcd\u5668\uff1a1\uff09 \u7801\u672c\u5229\u7528\u7387;2\uff09 \u6574\u4e2a\u7f16\u7801\u5668\u5185\u7684 ASR \u9519\u8bef\u7387;3\uff09 \u4e0d\u540c\u8bf4\u8bdd\u4eba\u7684\u4ee4\u724c\u53ef\u89c6\u5316;4\uff09 \u8bf4\u8bdd\u4eba\u8bc6\u522b\u57f9\u8bad\u3002\u8868&nbsp;<a href=\"https:\/\/arxiv.org\/html\/2412.10117v2#S4.T4\">4<\/a>&nbsp;\u663e\u793a\u4e86\u7801\u7c3f\u5229\u7528\u7387\u548c ASR \u9519\u8bef\u7387\u3002\u4e8b\u5b9e\u8bc1\u660e\uff0c\u57fa\u4e8e FSQ \u7684\u5206\u8bcd\u5668\u5145\u5206\u5229\u7528\u4e86\u7801\u672c\uff0c\u4ece ASR \u65b9\u9762\u7ef4\u62a4\u4e86\u66f4\u6709\u6548\u7684\u4fe1\u606f\uff0c\u8868\u660e FSQ \u7ef4\u62a4\u7684\u8bed\u4e49\u4fe1\u606f\u66f4\u591a\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"1018\" height=\"235\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-92.png\" alt=\"\" class=\"wp-image-23419\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-92.png 1018w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-92-300x69.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-92-768x177.png 768w\" sizes=\"(max-width: 1018px) 100vw, 1018px\" \/><\/figure>\n\n\n\n<p><strong>Comparison Results with Baselines\uff1a<\/strong><\/p>\n\n\n\n<p>\u6211\u4eec\u9996\u5148\u5728\u6709\u9650\u7684\u82f1\u8bed\u6587\u672c\u57df\u4e0a\u8bc4\u4f30\u4e86\u6211\u4eec\u7684 CosyVoice 2 \u6a21\u578b\uff0c\u5e76\u5c06\u5176\u4e0e\u51e0\u4e2a\u5f00\u6e90\u6a21\u578b\u8fdb\u884c\u4e86\u6bd4\u8f83\uff0c\u4f8b\u5982 ChatTTS\u3001GPT-SoVITs&nbsp;\u3001OpenVoice\u3001ParlerTTS\u3001EmotiVoice&nbsp;\u53ca\u5176\u524d\u8eab CosyVoice\u3002\u5ba2\u89c2\u7ed3\u679c\u5982\u8868&nbsp;<a href=\"https:\/\/arxiv.org\/html\/2412.10117v2#S4.T5\">5<\/a>&nbsp;\u6240\u793a\uff0c\u5305\u62ec\u5185\u5bb9\u4e00\u81f4\u6027 \uff08WER\uff09\u3001\u8bed\u97f3\u8d28\u91cf \uff08NMOS\uff09 \u548c\u8bf4\u8bdd\u4eba\u76f8\u4f3c\u5ea6 \uff08SS\uff09\u3002\u4ece\u8868\u4e2d\u6211\u4eec\u53ef\u4ee5\u770b\u5230\uff0cCosyVoice 2 \u5728 Librispeech \u6d4b\u8bd5\u6e05\u7406\u96c6\u4e0a\u5b9e\u73b0\u4e86\u6700\u5148\u8fdb\u7684\u6027\u80fd\uff0c\u8d85\u8d8a\u4e86\u6240\u6709\u57fa\u7ebf\u6a21\u578b\u548c\u6240\u6709\u8bc4\u4f30\u6307\u6807\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0cCosyVoice 2 \u751a\u81f3\u8868\u73b0\u51fa\u6bd4\u4eba\u7c7b\u8bdd\u8bed\u66f4\u9ad8\u7684\u5185\u5bb9\u4e00\u81f4\u6027\u3001\u8bed\u97f3\u8d28\u91cf\u548c\u8bf4\u8bdd\u4eba\u76f8\u4f3c\u5ea6\uff0c\u8fd9\u8868\u660e\u5176\u4eba\u7c7b\u5947\u5076\u6821\u9a8c\u7684\u5408\u6210\u8d28\u91cf\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"1011\" height=\"871\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-93.png\" alt=\"\" class=\"wp-image-23421\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-93.png 1011w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-93-300x258.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-93-768x662.png 768w\" sizes=\"(max-width: 1011px) 100vw, 1011px\" \/><figcaption>\u8868 5\uff1aLibriSpeech \u6d4b\u8bd5\u6e05\u7406\u57fa\u7ebf\u5b50\u96c6\u548c CosyVoice 2 \u4e0a\u7684\u5185\u5bb9\u4e00\u81f4\u6027 \uff08WER\uff09\u3001\u8bf4\u8bdd\u4eba\u76f8\u4f3c\u5ea6 \uff08SS\uff09 \u548c\u8bed\u97f3\u8d28\u91cf \uff08NMOS\uff09 \u7ed3\u679c\u3002Whisper-Large V3 \u7528\u4f5c ASR \u6a21\u578b\uff0c\u5e76\u4e14\u5728 WER \u8ba1\u7b97\u4e4b\u524d\u6392\u9664\u6807\u70b9\u7b26\u53f7\u3002<\/figcaption><\/figure>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"1018\" height=\"442\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-94.png\" alt=\"\" class=\"wp-image-23424\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-94.png 1018w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-94-300x130.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-94-768x333.png 768w\" sizes=\"(max-width: 1018px) 100vw, 1018px\" \/><\/figure>\n\n\n\n<p><strong>Modular Ablation Study<\/strong>\uff1a<\/p>\n\n\n\n<p>\u6211\u4eec\u5bf9\u6587\u672c\u8bed\u97f3\u8bed\u8a00\u6a21\u578b\u8fdb\u884c\u4e86\u6a21\u5757\u5316\u6d88\u878d\u7814\u7a76\uff0c\u4ee5\u8bc4\u4f30\u6211\u4eec\u7684\u4fee\u6539\u7684\u5f71\u54cd\uff0c\u5305\u62ec LLM \u521d\u59cb\u5316\u3001\u5220\u9664\u8bf4\u8bdd\u4eba\u5d4c\u5165\u548c\u5229\u7528 FSQ\u3002\u8868&nbsp;<a href=\"https:\/\/arxiv.org\/html\/2412.10117v2#S4.T7\">7<\/a>&nbsp;\u5c55\u793a\u4e86 CosyVoice 2 \u5728\u524d\u4ee3\u4ea7\u54c1\u7684\u57fa\u7840\u4e0a\u7684\u9010\u6b65\u53d1\u5c55\u3002\u901a\u8fc7\u5c06\u968f\u673a\u521d\u59cb\u5316\u7684\u8bed\u8a00\u6a21\u578b\u66ff\u6362\u4e3a\u9884\u8bad\u7ec3\u7684 LLM\uff09\uff0c\u6211\u4eec\u5728 test-zh \u548c test-hard \u96c6\u4e0a\u7684\u5185\u5bb9\u4e00\u81f4\u6027\u5206\u522b\u5b9e\u73b0\u4e86 18.46% \u548c 15.40% \u7684\u76f8\u5bf9\u6539\u8fdb&nbsp;\u3002\u63a5\u4e0b\u6765\uff0c\u6211\u4eec\u4ece\u6587\u672c\u8f6c\u8bed\u97f3\u8bed\u8a00\u6a21\u578b\u4e2d\u5220\u9664\u4e86\u8bf4\u8bdd\u4eba\u5d4c\u5165\uff0c\u8fd9\u6709\u52a9\u4e8e\u9632\u6b62\u4e0a\u4e0b\u6587\u5b66\u4e60\u4e2d\u7684\u4fe1\u606f\u6cc4\u9732\u548c\u5e72\u6270\u3002\u8fd9\u4e00\u53d8\u5316\u5bfc\u81f4\u5185\u5bb9\u9519\u8bef\u663e\u8457\u51cf\u5c11\uff0c\u540c\u65f6\u4fdd\u6301\u8bf4\u8bdd\u4eba\u76f8\u4f3c\u6027\uff0c\u8868\u660e\u5185\u5bb9\u4fe1\u606f\u4e3b\u8981\u7531 LM \u5efa\u6a21\uff0c\u8bf4\u8bdd\u4eba\u4fe1\u606f\u4e3b\u8981\u7531\u6d41\u5339\u914d\u6a21\u578b\u6062\u590d\u3002\u6700\u540e\uff0c\u901a\u8fc7\u5c06 VQ \u66ff\u6362\u4e3a FSQ\uff0c\u6211\u4eec\u5b9e\u73b0\u4e86 CosyVoice 2 \u6a21\u578b\uff0c\u6ce8\u610f\u5230\u66f4\u9ad8\u7684\u5185\u5bb9\u4e00\u81f4\u6027\u548c\u4e0d\u53d8\u7684\u8bf4\u8bdd\u4eba\u76f8\u4f3c\u5ea6\u3002\u901a\u8fc7\u5145\u5206\u5229\u7528\u7801\u672c\uff0cFSQ \u53ef\u4ee5\u6355\u83b7\u66f4\u591a\u7684\u5185\u5bb9\u4fe1\u606f\u548c\u4e0a\u4e0b\u6587\u53d8\u5316\uff0c\u4ece\u800c\u66f4\u597d\u5730\u534f\u8c03\u6587\u672c\u548c\u8bed\u97f3\u4ee4\u724c\u3002\u6b64\u5916\uff0c\u6211\u4eec\u901a\u8fc7\u5728\u57fa\u4e8e FSQ \u7684\u8bed\u97f3\u6807\u8bb0\u5668\u7684\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u5c06\u97f3\u9ad8\u635f\u5931\u4f5c\u4e3a\u7ea6\u675f\u6761\u4ef6\u8fdb\u884c\u4e86\u6bd4\u8f83\u5b9e\u9a8c\u3002\u6211\u4eec\u53d1\u73b0\u8fd9\u79cd\u65b9\u6cd5\u53ef\u4ee5\u63d0\u9ad8\u4e0b\u6e38 TTS \u4efb\u52a1\u7684\u6027\u80fd\uff0c\u5982\u8868&nbsp;<a href=\"https:\/\/arxiv.org\/html\/2412.10117v2#S4.T7\">7<\/a>&nbsp;\u7684\u6700\u540e\u4e00\u884c\u6240\u793a\u3002\u5728 CosyVoice \u7684\u672a\u6765\u7248\u672c\u4e2d\uff0c\u6211\u4eec\u8ba1\u5212\u8fdb\u884c\u66f4\u8be6\u7ec6\u7684\u5b9e\u9a8c\u548c\u5206\u6790\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"984\" height=\"307\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-95.png\" alt=\"\" class=\"wp-image-23425\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-95.png 984w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-95-300x94.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-95-768x240.png 768w\" sizes=\"(max-width: 984px) 100vw, 984px\" \/><\/figure>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"1018\" height=\"307\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-96.png\" alt=\"\" class=\"wp-image-23427\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-96.png 1018w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-96-300x90.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-96-768x232.png 768w\" sizes=\"(max-width: 1018px) 100vw, 1018px\" \/><\/figure>\n\n\n\n<p><strong>Results on Japanese and Korean Benchmarks<\/strong><\/p>\n\n\n\n<p>\u9664\u4e86\u4e2d\u6587\u548c\u82f1\u6587\uff0cCosyVoice 2 \u8fd8\u652f\u6301\u65e5\u8bed\u548c\u97e9\u8bed\u3002\u6211\u4eec\u5728\u6784\u5efa\u7684\u65e5\u8bed\u548c\u97e9\u8bed\u6d4b\u8bd5\u96c6\u4e0a\u8bc4\u4f30\u4e86\u5185\u5bb9\u4e00\u81f4\u6027\u3001\u8bf4\u8bdd\u4eba\u76f8\u4f3c\u5ea6\u548c\u8bed\u97f3\u8d28\u91cf\u3002\u5982\u8868&nbsp;<a href=\"https:\/\/arxiv.org\/html\/2412.10117v2#S4.T9\">9<\/a>&nbsp;\u6240\u793a\uff0c\u5728\u6240\u6709\u8bc4\u4f30\u6307\u6807\u4e2d\uff0cCosyVoice 2 \u5728\u97e9\u8bed\u4e0a\u7684\u8868\u73b0\u660e\u663e\u4f18\u4e8e\u65e5\u8bed\u3002\u8fd9\u79cd\u5dee\u5f02\u4e3b\u8981\u662f\u7531\u4e8e\u65e5\u8bed\u548c\u4e2d\u6587\u4e4b\u95f4\u7684\u5b57\u7b26\u96c6\u91cd\u53e0\uff0c\u8fd9\u5bfc\u81f4\u65e5\u8bed\u4e0a\u4e0b\u6587\u4e2d\u7684\u4e2d\u6587\u53d1\u97f3\u3002\u5728\u672a\u6765\u7684\u5de5\u4f5c\u4e2d\uff0c\u6211\u4eec\u8ba1\u5212\u63a2\u7d22\u589e\u5f3a\u591a\u8bed\u8a00\u5408\u6210\u7684\u8bed\u8a00\u4e0a\u4e0b\u6587\u7684\u65b9\u6cd5\u3002\u7531\u4e8e\u97e9\u8bed\u4e0e\u5176\u4ed6\u8bed\u8a00\u6ca1\u6709\u5b57\u7b26\u91cd\u53e0\uff0c\u56e0\u6b64\u5176\u8bed\u97f3\u5408\u6210\u6027\u80fd\u8981\u597d\u5f97\u591a\u3002\u53e6\u4e00\u4e2a\u95ee\u9898\u662f\u6570\u636e\u4e0d\u5e73\u8861\u3002\u6211\u4eec\u76f8\u4fe1\uff0c\u589e\u52a0\u8bad\u7ec3\u6570\u636e\u91cf\u53ef\u4ee5\u8fdb\u4e00\u6b65\u63d0\u9ad8\u65e5\u8bed\u548c\u97e9\u8bed\u7684\u7efc\u5408\u6027\u80fd\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"1015\" height=\"253\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-97.png\" alt=\"\" class=\"wp-image-23428\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-97.png 1015w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-97-300x75.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-97-768x191.png 768w\" sizes=\"(max-width: 1015px) 100vw, 1015px\" \/><\/figure>\n\n\n\n<p><em><strong>Results on Instructed Generation<\/strong><\/em>\uff1a<\/p>\n\n\n\n<p>\u4e3a\u4e86\u8bc4\u4f30 instructed generation \u7684\u6027\u80fd\uff0c\u6211\u4eec\u521b\u5efa\u4e86\u4e00\u4e2a\u5305\u542b 290 \u4e2a\u6837\u672c\u7684\u4e2d\u6587\u6d4b\u8bd5\u96c6\u3002\u8fd9\u7ec4\u6307\u4ee4\u5305\u62ec 29 \u79cd\u7c7b\u578b\u7684\u6307\u4ee4\uff0c\u5982\u8868&nbsp;<a href=\"https:\/\/arxiv.org\/html\/2412.10117v2#S2.SS6\">2.6<\/a>&nbsp;\u6240\u793a\uff0c\u6bcf\u79cd\u6307\u4ee4\u90fd\u6709 10 \u79cd\u4e0d\u540c\u7684\u8f93\u5165\u6587\u672c\u3002\u6211\u4eec\u4f7f\u7528\u6765\u81ea 5 \u4e2a\u8bf4\u8bdd\u4eba\uff083 \u4e2a\u5973\u6027\u548c 2 \u4e2a\u7537\u6027\uff09\u7684 5 \u4e2a\u97f3\u9891\u63d0\u793a\u548c\u8bf4\u8bdd\u4eba\u5d4c\u5165\u4f5c\u4e3a flow matching \u6a21\u578b\u7684\u6761\u4ef6\u3002\u6211\u4eec\u7684\u6d4b\u8bd5\u4ee5\u79bb\u7ebf\u6a21\u5f0f\u8fdb\u884c\u3002\u6211\u4eec\u5ba2\u89c2\u5730\u8bc4\u4f30\u4e86\u5185\u5bb9\u4e00\u81f4\u6027 \uff08CER\uff09\u3001\u8bf4\u8bdd\u4eba\u76f8\u4f3c\u5ea6 \uff08SS\uff09 \u548c\u8bed\u97f3\u8d28\u91cf \uff08NMOS\uff09\u3002\u4e3b\u89c2\u4e0a\uff0c\u6211\u4eec\u4f7f\u7528\u6559\u5b66\u5e73\u5747\u610f\u89c1\u5206\u6570 \uff08MOS-I\uff09 \u8bc4\u4f30\u6559\u5b66\u7684\u51c6\u786e\u6027\u548c\u81ea\u7136\u6027\uff0c\u8303\u56f4\u4ece 1 \u5230 5\u3002\u6bcf\u4e2a\u6837\u672c\u7531 10 \u540d\u4ee5\u4e2d\u6587\u4e3a\u6bcd\u8bed\u7684\u4eba\u8fdb\u884c\u8bc4\u4f30\uff0c\u5206\u6570\u4ee5 0.5 \u4e3a\u589e\u91cf\u5206\u914d\u3002\u8bc4\u4f30\u6807\u51c6\u4fa7\u91cd\u4e8e\u8bed\u97f3\u662f\u5426\u9075\u5b88\u6240\u6709\u6307\u5b9a\u7684\u6307\u4ee4\uff0c\u4f8b\u5982\u60c5\u611f\u8868\u8fbe\u3001\u8bed\u901f\u8c03\u6574\u3001\u65b9\u8a00\u4f7f\u7528\u548c\u89d2\u8272\u626e\u6f14\u3002\u7cbe\u7ec6\u7684\u63a7\u5236\uff08\u5305\u62ec\u63d2\u5165\u7b11\u58f0\u3001\u7b11\u58f0\u8bf4\u8bdd\u3001\u547c\u5438\u63a7\u5236\u548c\u5f3a\u8c03\uff09\u5c06\u8bc4\u4f30\u5176\u81ea\u7136\u6027\u548c\u51c6\u786e\u6027\u3002\u5982\u8868&nbsp;<a href=\"https:\/\/arxiv.org\/html\/2412.10117v2#S4.T10\">10<\/a>&nbsp;\u6240\u793a\uff0cCosyVoice 2 \u8868\u73b0\u51fa\u5353\u8d8a\u7684\u5185\u5bb9\u4e00\u81f4\u6027 \uff08CER\uff09\u3001\u8bf4\u8bdd\u4eba\u76f8\u4f3c\u6027 \uff08SS\uff09 \u4ee5\u53ca\u6307\u4ee4\u63a7\u5236 \uff08MOS-I\uff09 \u7684\u51c6\u786e\u6027\u548c\u81ea\u7136\u6027\uff0c\u540c\u65f6\u4fdd\u6301\u4e86\u4e0e CosyVoice-Inspire \u76f8\u5f53\u7684\u8bed\u97f3\u8d28\u91cf\u3002\u5f53\u4ece CosyVoice 2 \u4e2d\u5220\u9664\u8f93\u5165\u6307\u4ee4\u65f6\uff0cMOS-I \u660e\u663e\u4e0b\u964d;\u7136\u800c\uff0c\u5728\u5185\u5bb9\u4e00\u81f4\u6027 \uff08CER\uff09 \u3001\u8bf4\u8bdd\u4eba\u76f8\u4f3c\u5ea6 \uff08SS\uff09 \u548c\u8bed\u97f3\u8d28\u91cf \uff08NMOS\uff09 \u65b9\u9762\u89c2\u5bdf\u5230\u6539\u5584\u3002\u8fd9\u8868\u660e\u6307\u4ee4\u53ef\u63a7\u6027\u5f88\u96be\u4ece\u5185\u5bb9\u6587\u672c\u4e2d\u9690\u5f0f\u51fa\u73b0\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"366\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-98-1024x366.png\" alt=\"\" class=\"wp-image-23431\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-98-1024x366.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-98-300x107.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-98-768x274.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-98.png 1045w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p><strong>Results on Speaker Fine-tuned Models<\/strong><\/p>\n\n\n\n<p>\u5728\u5fae\u8c03\u9636\u6bb5\uff0c\u6211\u4eec\u5bf9\u540c\u4e00\u626c\u58f0\u5668\u7684\u626c\u58f0\u5668\u5d4c\u5165\u91c7\u7528\u65e0\u76d1\u7763\u805a\u7c7b\uff0c\u4ee5\u786e\u4fdd\u626c\u58f0\u5668\u97f3\u8272\u7684\u7a33\u5b9a\u6027\u3002\u6211\u4eec\u5df2\u7ecf\u8bc1\u660e\uff0c\u53ea\u6709 400 \u4e2a\u97f3\u9891\u8bb0\u5f55\u7684\u76ee\u6807\u8bf4\u8bdd\u4eba\u53ef\u4ee5\u5b9e\u73b0\u76f8\u5f53\u597d\u7684\u8bed\u97f3\u5408\u6210\u6027\u80fd\uff0c\u5728\u4e0d\u540c\u8bf4\u8bdd\u4eba\u4e4b\u95f4\u89c2\u5bdf\u5230\u7684\u5ba2\u89c2\u6307\u6807\u4ec5\u5b58\u5728\u8f7b\u5fae\u5dee\u5f02\uff0c\u5982\u56fe&nbsp;<a href=\"https:\/\/arxiv.org\/html\/2412.10117v2#S4.F6\">6<\/a>&nbsp;\u6240\u793a\u3002\u6211\u4eec\u7684\u5b9e\u9a8c\u8868\u660e\uff0c\u5927\u591a\u6570\u8bf4\u8bdd\u4eba\u53ef\u4ee5\u7ee7\u627f\u96f6\u955c\u5934 TTS \u6a21\u578b\u7684\u7a33\u5065\u4e0a\u4e0b\u6587\u7406\u89e3\u548c\u611f\u77e5\uff0c\u4ece\u800c\u81ea\u7136\u5730\u8868\u8fbe\u5404\u79cd\u60c5\u7eea\u548c\u60c5\u7eea\u4ee5\u54cd\u5e94\u8f93\u5165\u6587\u672c\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"1012\" height=\"447\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-99.png\" alt=\"\" class=\"wp-image-23433\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-99.png 1012w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-99-300x133.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-99-768x339.png 768w\" sizes=\"(max-width: 1012px) 100vw, 1012px\" \/><\/figure>\n\n\n\n<p><strong>LLM Fine-tuning with Reinforcement Learning<\/strong><\/p>\n\n\n\n<p>\u5c3d\u7ba1 SFT \u53ef\u4ee5\u63d0\u9ad8\u5927\u591a\u6570\u626c\u58f0\u5668\u7684\u6027\u80fd\uff0c\u4f46 Spk E \u7684\u7ed3\u679c\u4ecd\u7136\u6bd4\u57fa\u672c\u6a21\u578b\u5dee\uff0c\u5c24\u5176\u662f\u5728\u82f1\u8bed\u4e0a\u3002\u56e0\u4e3a Spk E \u7684\u58f0\u97f3\u66f4\u590d\u6742\uff0c\u8bf4\u8bdd\u901f\u5ea6\u66f4\u5feb\u3002\u6b64\u5916\uff0c\u53ea\u6709 Chinese \u5f55\u97f3\u53ef\u7528\u4e8e Spk E\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u5728 Spk E \u4e0a\u5e94\u7528\u5f3a\u5316\u5b66\u4e60\u4ee5\u8fdb\u4e00\u6b65\u6539\u8fdb\u3002\u5bf9\u4e8e DPO\uff0c\u6211\u4eec\u901a\u8fc7 SFT \u6a21\u578b\u5408\u6210\u4e86 10,000 \u4e2a\u6837\u672c\u5bf9\uff0c\u4ee5\u6539\u53d8 ASR \u548c SS \u5956\u52b1\u5bf9 LM \u7684\u504f\u597d\u504f\u5dee\u3002\u6211\u4eec\u8fd8\u4f7f\u7528\u53ef\u5fae\u5206\u7684 ASR \u5956\u52b1\u6765\u4f18\u5316 LM \u53c2\u6570\u3002\u5728 RL \u4e4b\u540e\uff0c\u6211\u4eec\u5728 Spk E \u6d4b\u8bd5\u96c6\u4e0a\u7528\u5185\u5bb9\u4e00\u81f4\u6027 \uff08WER\uff09\u3001\u8bf4\u8bdd\u4eba\u76f8\u4f3c\u5ea6 \uff08SS\uff09 \u548c\u8bed\u97f3\u8d28\u91cf \uff08NMOS\uff09 \u8bc4\u4f30\u6a21\u578b\uff0c\u5e76\u8fdb\u4e00\u6b65\u8bc4\u4f30 SeedTTS \u6d4b\u8bd5\u96c6\u4e0a\u7684 WER\uff0c\u4ee5\u63a2\u7d22\u6a21\u578b\u662f\u5426\u53ef\u4ee5\u4fdd\u6301\u5bf9\u57df\u5916\u6216\u8de8\u8bed\u8a00\u8f93\u5165\u6587\u672c\u7684\u9c81\u68d2\u6027\u3002\u7ed3\u679c\u5982\u8868&nbsp;<a href=\"https:\/\/arxiv.org\/html\/2412.10117v2#S4.T11\">11<\/a>&nbsp;\u6240\u793a<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"997\" height=\"346\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-100.png\" alt=\"\" class=\"wp-image-23436\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-100.png 997w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-100-300x104.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/12\/image-100-768x267.png 768w\" sizes=\"(max-width: 997px) 100vw, 997px\" \/><\/figure>\n\n\n\n<p>\u4e0e\u9884\u5148\u8bad\u7ec3\u7684\u57fa\u7840\u6a21\u578b\u76f8\u6bd4\uff0cSFT \u6a21\u578b\u663e\u793a\u51fa\u66f4\u9ad8\u7684\u8bf4\u8bdd\u4eba\u76f8\u4f3c\u5ea6\u548c\u8bed\u97f3\u8d28\u91cf\uff0c\u4f46\u662f\uff0cWER \u53ef\u80fd\u6bd4\u57fa\u7840\u6a21\u578b\u5dee\u3002\u6211\u4eec\u53d1\u73b0\uff0c\u57fa\u672c\u6a21\u578b\u5408\u6210\u7684\u97f3\u9891\u603b\u662f\u6bd4 SFT \u548c\u771f\u5b9e\u503c\u6162\uff0c\u8fd9\u5bf9 ASR \u7cfb\u7edf\u66f4\u53cb\u597d\u3002\u5bf9\u4e8e\u76ee\u6807\u8bf4\u8bdd\u4eba\u6570\u636e\u96c6\uff0c\u504f\u597d\u504f\u5dee\u548c\u53ef\u5fae\u5206\u5956\u52b1\u90fd\u53ef\u4ee5\u964d\u4f4e WER\uff0c\u800c\u5bf9\u5176\u4ed6\u4e24\u4e2a\u6307\u6807\u7684\u6709\u5bb3\u5f71\u54cd\u5f88\u5c0f\u3002\u4f46\u5bf9\u4e8e SEED \u6d4b\u8bd5\u96c6\uff0c\u57fa\u4e8e DPO \u7684\u5f3a\u5316\u4ec5\u5bf9\u4e2d\u6587\u548c\u82f1\u6587\u5b50\u96c6\u6709\u76ca\uff0c\u800c\u786c\u6837\u672c\u4f1a\u66f4\u5dee\u3002\u539f\u56e0\u53ef\u80fd\u662f\u786c\u6837\u672c\u5305\u542b\u8bb8\u591a\u91cd\u590d\u7684\u5355\u8bcd\u6216\u77ed\u8bed\uff0c\u5728 DPO \u8bad\u7ec3\u671f\u95f4\u53ef\u4ee5\u88ab\u89c6\u4e3a\u88ab\u62d2\u7edd\u7684\u6837\u672c\u3002\u4f46\u662f\uff0c\u53ef\u5fae\u5206\u7684 ASR \u5956\u52b1\u4e0d\u4f1a\u9047\u5230\u8fd9\u4e2a\u95ee\u9898\uff0c\u56e0\u4e3a\u5b83\u53ef\u4ee5\u76f4\u63a5\u901a\u8fc7 ASR \u540e\u9a8c\u4f18\u5316 TTS \u7cfb\u7edf\u3002\u8fd9\u610f\u5473\u7740\u53ef\u5fae\u5206\u7684 ASR \u5956\u52b1\u5728\u57df\u5916\u60c5\u51b5\u4e0b\u5177\u6709\u66f4\u597d\u7684\u6cdb\u5316\u80fd\u529b\u3002\u6700\u540e\uff0c\u6211\u4eec\u53ef\u4ee5\u5c06\u5b83\u4eec\u76f8\u4e92\u7ec4\u5408\u4ee5\u8fdb\u884c\u8fdb\u4e00\u6b65\u6539\u8fdb\u3002<\/p>\n\n\n\n<h3>Conclusion&nbsp;<\/h3>\n\n\n\n<p>\u5728 CosyVoice \u6210\u529f\u7684\u57fa\u7840\u4e0a\uff0c\u672c\u62a5\u544a\u4ecb\u7ecd\u4e86 CosyVoice 2\uff0c\u8fd9\u662f\u4e00\u79cd\u6539\u8fdb\u7684\u6d41\u5f0f\u8bed\u97f3\u5408\u6210\u6a21\u578b\uff0c\u5b83\u5229\u7528\u4e86\u5927\u578b\u8bed\u8a00\u6a21\u578b\u3002\u901a\u8fc7\u5c06\u6d41\u5f0f\u548c\u975e\u6d41\u5f0f\u5408\u6210\u7edf\u4e00\u5728\u4e00\u4e2a\u6846\u67b6\u4e2d\uff0cCosyVoice 2 \u5b9e\u73b0\u4e86\u4eba\u7c7b\u5947\u5076\u6821\u9a8c\u7684\u81ea\u7136\u6027\u3001\u6700\u5c0f\u7684\u54cd\u5e94\u5ef6\u8fdf\u548c\u6d41\u5f0f\u6a21\u5f0f\u4e0b\u51e0\u4e4e\u65e0\u635f\u7684\u5408\u6210\u8d28\u91cf\u3002\u5173\u952e\u521b\u65b0\u5305\u62ec\u7528\u4e8e\u5145\u5206\u5229\u7528\u7801\u672c\u7684\u6709\u9650\u6807\u91cf\u91cf\u5316\u3001\u5305\u542b\u9884\u8bad\u7ec3\u6587\u672cLLMs\uff0c\u4ee5\u53ca\u5f00\u53d1\u5757\u611f\u77e5\u56e0\u679c\u6d41\u5339\u914d\u6a21\u578b\u4ee5\u652f\u6301\u4e0d\u540c\u7684\u5408\u6210\u573a\u666f\u3002\u6b64\u5916\uff0c\u6307\u4ee4 TTS \u80fd\u529b\u7684\u6539\u8fdb\u5141\u8bb8\u901a\u8fc7\u5bf9\u60c5\u611f\u3001\u53e3\u97f3\u3001\u89d2\u8272\u98ce\u683c\u548c\u4eba\u58f0\u7206\u53d1\u7684\u7cbe\u7ec6\u63a7\u5236\uff0c\u751f\u6210\u591a\u529f\u80fd\u548c\u751f\u52a8\u7684\u8bed\u97f3\u3002\u901a\u8fc7\u7cfb\u7edf\u7684\u4fee\u6539\u548c\u4f18\u5316\uff0cCosyVoice 2 \u4e0d\u4ec5\u63d0\u4f9b\u4e86\u5353\u8d8a\u7684\u5408\u6210\u8d28\u91cf\uff0c\u800c\u4e14\u653e\u5bbd\u4e86\u90e8\u7f72\u8981\u6c42\uff0c\u4f7f\u5176\u9002\u7528\u4e8e\u6d41\u5f0f\u548c\u975e\u6d41\u5f0f\u5e94\u7528\u3002\u6211\u4eec\u76f8\u4fe1 CosyVoice 2 \u4ee3\u8868\u4e86\u53ef\u6269\u5c55\u3001\u9ad8\u8d28\u91cf\u548c\u4ea4\u4e92\u5f0f\u6587\u672c\u8f6c\u8bed\u97f3\u5408\u6210\u7684\u91cd\u5927\u8fdb\u6b65\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>CosyVoice 2: Scalable Streaming Speech Synthesis with L &hellip; <a href=\"http:\/\/139.9.1.231\/index.php\/2024\/12\/20\/cosyvoice-2-0\/\" class=\"more-link\">\u7ee7\u7eed\u9605\u8bfb<span class=\"screen-reader-text\">CosyVoice-2.0 \u8bed\u97f3\u751f\u6210\u5927\u6a21\u578b\u5347\u7ea7<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[40,4,9,38,34],"tags":[],"_links":{"self":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/23119"}],"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=23119"}],"version-history":[{"count":193,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/23119\/revisions"}],"predecessor-version":[{"id":27657,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/23119\/revisions\/27657"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=23119"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=23119"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=23119"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}