{"id":30222,"date":"2026-02-11T18:02:28","date_gmt":"2026-02-11T10:02:28","guid":{"rendered":"http:\/\/139.9.1.231\/?p=30222"},"modified":"2026-02-11T18:02:30","modified_gmt":"2026-02-11T10:02:30","slug":"neural-audio-codecs-how-to-get-audio-into-llms","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2026\/02\/11\/neural-audio-codecs-how-to-get-audio-into-llms\/","title":{"rendered":"\u97f3\u9891 Tokenizer\u7684\u65b9\u6cd5\uff1aMoshi \u56e2\u961f\u5206\u4eab"},"content":{"rendered":"\n<p class=\"has-light-pink-background-color has-background\"><strong><em>\u539f\u6587\u94fe\u63a5\uff1a<a href=\"https:\/\/kyutai.org\/codec-explainer\">https:\/\/kyutai.org\/codec-explainer<\/a><\/em><\/strong><\/p>\n\n\n\n\n\n<figure class=\"wp-block-video\"><video controls src=\"https:\/\/kyutai.org\/assets\/codec-explainer\/codecIntro.mp4\"><\/video><figcaption>\u65b9\u6848\uff1a\u5c06\u8bed\u8a00\u6a21\u578b\u5d4c\u5165\u97f3\u9891\u7f16\u7801\u5668\/\u89e3\u7801\u5668\u5bf9\uff08=\u795e\u7ecf\u97f3\u9891\u7f16\u89e3\u7801\u5668\uff09\u4e2d\uff0c\u4f7f\u5176\u80fd\u591f\u9884\u6d4b\u97f3\u9891\u7684\u540e\u7eed\u5185\u5bb9\u3002<\/figcaption><\/figure>\n\n\n\n<p>\u622a\u81f3 2025 \u5e74 10 \u6708\uff0cSpeech LLM \u7684\u8868\u73b0\u8fd8\u5f88\u7cdf\u7cd5\u3002\u8bb8\u591a LLM \u90fd\u63d0\u4f9b\u4e86\u8bed\u97f3\u63a5\u53e3\uff0c\u4f46\u5b83\u4eec\u7684\u5de5\u4f5c\u6d41\u901a\u5e38\u662f\uff1a\u8bed\u97f3\u8f6c\u6587\u5b57\uff0c\u751f\u6210\u6587\u672c\u7b54\u6848\uff0c\u518d\u901a\u8fc7 text-to-speech (TTS) \u8bfb\u51fa\u6765\u3002\u8fd9\u5728\u5f88\u591a\u573a\u666f\u4e0b\u591f\u7528\u4e86\uff08\u6bd4\u5982&nbsp;Unmute\uff09\uff0c\u4f46\u8fd9\u672c\u8d28\u4e0a\u53ea\u662f\u4e00\u4e2a wrapper\uff0c\u5e76\u975e\u771f\u6b63\u7684\u8bed\u97f3\u7406\u89e3\u3002\u6a21\u578b\u65e0\u6cd5\u611f\u77e5\u4f60\u58f0\u97f3\u91cc\u7684\u6cae\u4e27\u5e76\u5171\u60c5\u5730\u56de\u5e94\uff0c\u65e0\u6cd5\u5728\u56de\u7b54\u4e2d\u5f3a\u8c03\u91cd\u70b9\uff0c\u4e5f\u542c\u4e0d\u51fa\u8bbd\u523a\u3002<\/p>\n\n\n\n<p>\u662f\u7684\uff0c\u73b0\u5728\u786e\u5b9e\u6709\u4e00\u4e9b LLM\uff08\u5982 Gemini\u3001ChatGPT \u7684 Advanced Voice Mode\u3001Qwen\u3001Moshi\uff09\u80fd\u591f\u539f\u751f (natively) \u7406\u89e3\u548c\u751f\u6210\u8bed\u97f3\u3002\u4f46\u5728\u5b9e\u8df5\u4e2d\uff0c\u5b83\u4eec\u8981\u4e48\u4e0d\u591f\u667a\u80fd\uff0c\u8981\u4e48\u5176\u884c\u4e3a\u5c31\u50cf\u4e00\u4e2a text model wrapper\u3002\u4f60\u53ef\u4ee5\u8bd5\u8bd5\u7528\u5f88\u9ad8\u7684\u97f3\u8c03\u95ee\u5b83\u4eec\u4efb\u4f55\u4e00\u4e2a\uff1a\u201c\u6211\u8bf4\u8bdd\u7684\u58f0\u97f3\u662f\u4f4e\u8fd8\u662f\u9ad8\uff1f\u201d\uff0c\u5b83\u4eec\u90fd\u7b54\u4e0d\u4e0a\u6765\u3002<\/p>\n\n\n\n<p>\u663e\u7136\uff0cSpeech LLM \u7684\u53d1\u5c55\u6ede\u540e\u4e8e Text LLM\u3002\u4f46\u4e3a\u4ec0\u4e48\u5462\uff1f\u5bf9\u4e8e\u6587\u672c\uff0c\u6211\u4eec\u51e0\u5e74\u524d\u5c31\u53d1\u73b0\uff0c\u53ea\u8981\u6709\u6d77\u91cf\u6587\u672c\u6570\u636e\u3001\u4e00\u4e2a\u5de8\u5927\u7684 Transformer \u6a21\u578b\u548c\u5927\u91cf\u7684 GPU\uff0c\u5c31\u80fd\u5f97\u5230\u6548\u679c\u60ca\u4eba\u7684\u6587\u672c\u7eed\u5199\u6a21\u578b\u3002\u90a3\u4e3a\u4ec0\u4e48\u6211\u4eec\u4e0d\u80fd\u76f4\u63a5\u628a text \u6362\u6210 audio\uff0c\u7136\u540e\u5f97\u5230\u540c\u6837\u60ca\u4eba\u7684\u8bed\u97f3\u7eed\u5199\u6a21\u578b\u5462\uff1f<\/p>\n\n\n\n<p>\u5148\u5356\u4e2a\u5173\u5b50\uff0c\u5982\u679c\u4f60\u771f\u7684\u5929\u771f\u5730\u8fd9\u4e48\u505a\u4e86\uff0c\u5c31\u4f1a\u5f97\u5230\u4e0b\u9762\u8fd9\u79cd\u7ed3\u679c\uff08\u8b66\u544a\uff0c\u97f3\u91cf\u5f88\u5927\uff09\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-audio\"><audio controls src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/02\/39b30e90-abc8-49cb-8e2e-70647227ff35.wav\"><\/audio><\/figure>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"756\" height=\"100\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/02\/image-9.png\" alt=\"\" class=\"wp-image-30238\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/02\/image-9.png 756w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/02\/image-9-300x40.png 300w\" sizes=\"(max-width: 756px) 100vw, 756px\" \/><\/figure>\n\n\n\n<p>\u63a5\u4e0b\u6765\uff0c\u6211\u4eec\u5c06\u63a2\u8ba8\u4e3a\u4ec0\u4e48 audio \u6bd4 text \u66f4\u96be\u5efa\u6a21\uff0c\u4ee5\u53ca\u5982\u4f55\u5229\u7528&nbsp;<strong>neural audio codec<\/strong>&nbsp;\u6765\u964d\u4f4e\u5efa\u6a21\u96be\u5ea6\u3002\u2014\u2014\u8fd9\u5df2\u662f\u5c06 audio \u8f93\u5165\u548c\u8f93\u51fa LLM \u7684\u5b9e\u9645\u6807\u51c6\u65b9\u6cd5\u3002\u901a\u8fc7 codec\uff0c\u6211\u4eec\u53ef\u4ee5\u5c06\u8fde\u7eed\u7684\u97f3\u9891\u4fe1\u53f7\u8f6c\u6362\u6210\u66f4\u5927\u7c92\u5ea6\u7684\u79bb\u6563 token\uff0c\u7136\u540e\u8bad\u7ec3\u6a21\u578b\u6765\u9884\u6d4b\u8fd9\u4e9b token \u7684\u540e\u7eed\u5e8f\u5217\uff0c\u6700\u540e\u518d\u5c06\u8fd9\u4e9b \u7eed\u5199\u7684token \u89e3\u7801\u8fd8\u539f\u6210 audio\uff1a\u8bf7\u770b\u4e0a\u9762\u7684\u52a8\u753b\u3002<\/p>\n\n\n\n<p>Kyutai \u7684\u540c\u4ec1\u4eec\u5728\u8fd9\u4e2a\u9886\u57df\u505a\u4e86\u5927\u91cf\u5de5\u4f5c\uff0c\u8fd9\u4e5f\u662f\u6211\u9009\u62e9\u8fd9\u4e2a\u4e3b\u9898\u7684\u90e8\u5206\u539f\u56e0\u3002\u6211\u4eec\u5c06\u4ece\u57fa\u7840\u8bb2\u8d77\uff0c\u4e00\u76f4\u8bb2\u5230\u6211\u4eec\u7684 neural audio codec\u2014\u2014<strong><a rel=\"noreferrer noopener\" href=\"https:\/\/huggingface.co\/kyutai\/mimi\" target=\"_blank\">Mimi<\/a><\/strong>\u3002\u5b83\u6700\u521d\u662f\u4e3a <a rel=\"noreferrer noopener\" href=\"https:\/\/arxiv.org\/abs\/2410.00037\" target=\"_blank\"><strong>Moshi<\/strong><\/a> \u5f00\u53d1\u7684\uff0c\u540e\u6765\u88ab\u5176\u4ed6\u6a21\u578b\u6240\u91c7\u7528\uff0c\u5305\u62ec Sesame \u7684 <a rel=\"noreferrer noopener\" href=\"https:\/\/www.sesame.com\/research\/crossing_the_uncanny_valley_of_voice\" target=\"_blank\"><strong>CSM \u6a21\u578b<\/strong><\/a>\u3002<\/p>\n\n\n\n<h3><strong>Text is easy<\/strong><\/h3>\n\n\n\n<p>\u5728\u6587\u672c\u5206\u8bcd\u65b9\u9762\uff0c\u4e1a\u754c\u666e\u904d\u91c7\u7528\u4e00\u79cd\u79f0\u4e3a<strong>\u5b57\u8282\u5bf9\u7f16\u7801<\/strong>\uff08Byte-Pair Encoding, BPE\uff09\u7684\u6280\u672f\uff0c\u5e76\u4e14\u6781\u5c11\u5bf9\u5206\u8bcd\u5668\u8fdb\u884c\u66f4\u6539\u3002\u4ee5 OpenAI \u4e3a\u4f8b\uff0c\u81ea GPT-4o \u4ee5\u6765\u4e00\u76f4\u6cbf\u7528\u540c\u4e00\u5957\u5206\u8bcd\u5668\u2014\u2014\u5982\u679c\u4ee5\u5927\u8bed\u8a00\u6a21\u578b\u7684\u53d1\u5c55\u8282\u594f\u6765\u8861\u91cf\uff0cGPT-4o \u5df2\u53ef\u7b97\u4f5c\u201c\u76f8\u5f53\u4e45\u8fdc\u201d\u7684\u6a21\u578b\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img src=\"https:\/\/kyutai.org\/assets\/codec-explainer\/image.png\" alt=\"\"\/><figcaption>A random text from Wikipedia tokenized via the GPT-4o tokenizer<\/figcaption><\/figure>\n\n\n\n<p>\u5373\u4fbf\u5b8c\u5168\u4e0d\u5bf9\u6587\u672c\u8fdb\u884c\u5206\u8bcd\u3001\u4ec5\u5bf9\u5355\u4e2a\u5b57\u7b26\u8fdb\u884c\u9010\u5b57\u7b26\u9884\u6d4b\uff0c\u4e5f\u80fd\u591f\u53d6\u5f97\u76f8\u5f53\u4e0d\u9519\u7684\u6548\u679c\u3002\u65e9\u671f\u8ba9\u6211\u5bf9\u673a\u5668\u5b66\u4e60\u4ea7\u751f\u6d53\u539a\u5174\u8da3\u7684\u4e00\u7bc7\u6587\u7ae0\uff0c\u662f Andrej Karpathy \u4e8e 2015 \u5e74\u53d1\u8868\u7684\u5173\u4e8e\u5faa\u73af\u795e\u7ecf\u7f51\u7edc\uff08RNN\uff09\u6709\u6548\u6027\u7684\u535a\u5ba2\u3002\u5728\u90a3\u7bc7\u6587\u7ae0\u4e2d\uff0cKarpathy \u4f7f\u7528\u5355\u5757 GPU \u8bad\u7ec3\u4e86\u4e00\u4e2a\u4e09\u5c42 LSTM \u6a21\u578b\uff0c\u4f7f\u5176\u80fd\u591f\u751f\u6210\u7ed3\u6784\u4e0a\u8f83\u4e3a\u5408\u7406\u7684\u4ee3\u7801\u4e0e LaTeX \u6587\u672c\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image is-resized\"><img loading=\"lazy\" src=\"https:\/\/kyutai.org\/assets\/codec-explainer\/rnns-code.png\" alt=\"\" width=\"393\" height=\"494\"\/><\/figure>\n\n\n\n<figure class=\"wp-block-image is-resized\"><img loading=\"lazy\" src=\"https:\/\/kyutai.org\/assets\/codec-explainer\/rnns-latex.png\" alt=\"\" width=\"400\" height=\"456\"\/><\/figure>\n\n\n\n<p>\u8981\u77e5\u9053\uff0c\u8fd9\u53ef\u662f\u5341\u5e74\u524d\u7684\u4e8b\u4e86\uff0c\u90a3\u65f6\u5019\u6211\u4eec\u751a\u81f3\u8fd8\u4e0d\u77e5\u9053 &#8220;attention is all we need&#8221;\u3002\u73b0\u5728\uff0c\u6211\u4eec\u518d\u6765\u5bf9\u6bd4\u4e00\u4e0b Karpathy \u7684\u7ed3\u679c\u548c WaveNet \u7684\u6837\u672c\uff0c\u540e\u8005\u662f DeepMind \u5728\u4e00\u5e74\u540e\u53d1\u5e03\u7684\u6a21\u578b\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-audio\"><audio controls src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/02\/0d33bc9a-b1d3-4f54-b94f-1598aa3aa112.wav\"><\/audio><\/figure>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"135\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/02\/image-10-1024x135.png\" alt=\"\" class=\"wp-image-30244\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/02\/image-10-1024x135.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/02\/image-10-300x40.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/02\/image-10-768x102.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/02\/image-10.png 1134w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>\u4ece\u7eaf\u58f0\u5b66\u89d2\u5ea6\u6765\u770b\uff0c\u8fd9\u6bb5\u97f3\u9891\u7684\u542c\u611f\u8d28\u91cf\u8f83\u9ad8\uff0c\u4f46\u5374\u51e0\u4e4e\u65e0\u6cd5\u751f\u6210\u54ea\u6015\u4e00\u4e2a\u6b63\u786e\u7684\u82f1\u6587\u5355\u8bcd\u3002\u5f53\u7136\uff0c\u6211\u4eec\u4e5f\u4e0d\u5e94\u5bf9 WaveNet \u8fc7\u4e8e\u82db\u8d23\u3002Karpathy \u7684 RNN \u6240\u751f\u6210\u7684\u6587\u672c\u6837\u672c\u957f\u5ea6\u4e0d\u8fc7\u6570\u5343\u4e2a\u5b57\u7b26\uff0c\u800c\u8fd9\u6bb5 10 \u79d2\u7684\u97f3\u9891\u5374\u5305\u542b\u7ea6 16 \u4e07\u4e2a\u97f3\u9891\u91c7\u6837\u70b9\uff1b\u5e76\u4e14\uff0cWaveNet \u662f\u4ee5\u9010\u91c7\u6837\u70b9\u9884\u6d4b\uff08sample-by-sample generation\uff09\u7684\u65b9\u5f0f\uff0c\u6781\u5176\u7ec6\u81f4\u5730\u751f\u6210\u6574\u6bb5\u97f3\u9891\u7684\u3002<\/p>\n\n\n\n<figure class=\"wp-block-video\"><video controls src=\"https:\/\/kyutai.org\/assets\/codec-explainer\/wavenet-audio.mp4\"><\/video><figcaption>\u4e00\u79d2\u949f\u7684\u97f3\u9891\u901a\u5e38\u5305\u542b\u6570\u4ee5\u4e07\u8ba1\u7684\u91c7\u6837\u70b9\uff0c\u7136\u800c\u5176\u8bed\u4e49\u5185\u5bb9\u5f80\u5f80\u53ea\u5bf9\u5e94\u5c11\u91cf\u51e0\u4e2a\u8bcd\u8bed\u3002\uff08\u52a8\u753b\u5f15\u81ea WaveNet \u535a\u5ba2\u6587\u7ae0\u3002\uff09<\/figcaption><\/figure>\n\n\n\n<p>\u5728\u5982\u6b64\u957f\u7684\u65f6\u95f4\u5c3a\u5ea6\u4e0a\u7ef4\u6301\u751f\u6210\u5185\u5bb9\u7684\u8fde\u8d2f\u6027\u662f\u6781\u5177\u6311\u6218\u6027\u7684\uff1b\u540c\u65f6\uff0c\u7531\u4e8e\u9700\u8981\u6267\u884c\u6570\u91cf\u5e9e\u5927\u7684\u9010\u6b65\u9884\u6d4b\uff0c\u6a21\u578b\u7684\u8fd0\u884c\u5f00\u9500\u4e5f\u5341\u5206\u9ad8\u6602\u3002<\/p>\n\n\n\n<p>\u56e0\u6b64\uff0c\u4e0e\u5176\u8ba9\u6a21\u578b\u76f4\u63a5\u9010\u91c7\u6837\u70b9\u5730\u8fdb\u884c\u9884\u6d4b\uff0c\u4e0d\u5982\u5148\u8bad\u7ec3\u4e00\u4e2a\u6a21\u578b\uff0c\u5c06\u97f3\u9891\u538b\u7f29\u5230\u66f4\u6613\u5904\u7406\u7684\u8868\u793a\u7a7a\u95f4\u3002\u5177\u4f53\u800c\u8a00\uff0c\u53ef\u4ee5\u5148\u5bf9\u97f3\u9891\u8fdb\u884c\u538b\u7f29\uff0c\u5728\u538b\u7f29\u540e\u7684\u8868\u793a\u4e0a\u5229\u7528\u5927\u8bed\u8a00\u6a21\u578b\u9884\u6d4b\u5176\u540e\u7eed\u5185\u5bb9\uff0c\u518d\u5c06\u9884\u6d4b\u7ed3\u679c\u89e3\u538b\u8fd8\u539f\u4e3a\u97f3\u9891\u4fe1\u53f7\u3002<\/p>\n\n\n\n<h3>Sample by sample<\/h3>\n\n\n\n<p>\u4e0d\u8fc7\uff0c\u5728\u6b64\u4e4b\u524d\uff0c\u6211\u4eec\u5148\u6784\u5efa\u4e00\u4e2a\u57fa\u7ebf\u6a21\u578b\uff0c\u50cf WaveNet \u90a3\u6837\u9010\u91c7\u6837\u70b9\u751f\u6210\u97f3\u9891\u3002\u8fd9\u4e9b\u5b9e\u9a8c\u7684\u4ee3\u7801\u5747\u5df2\u5f00\u6e90\uff0c\u53ef\u5728\u76f8\u5e94\u4ed3\u5e93\u4e2d\u83b7\u53d6\u3002\u6211\u57fa\u4e8e Andrej Karpathy \u7684 nanoGPT \u4ed3\u5e93\u8fdb\u884c\u4e86\u590d\u73b0\u4e0e\u6269\u5c55\uff1b\u8be5\u4ed3\u5e93\u662f GPT-2 \u7684\u4e00\u4e2a\u7b80\u6d01\u5b9e\u73b0\u7248\u672c\u3002<\/p>\n\n\n\n<p>\u4ece\u8bed\u8a00\u6a21\u578b\u7684\u89c6\u89d2\u6765\u770b\uff0c\u6587\u672c\u4e0e\u97f3\u9891\u5728\u5f62\u5f0f\u4e0a\u5e76\u65e0\u672c\u8d28\u533a\u522b\uff1a\u672c\u8d28\u90fd\u662f\u201c\u8f93\u5165 token\uff0c\u8f93\u51fa token\u201d\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u9700\u8981\u505a\u7684\u4ec5\u662f\u5c06\u8fde\u7eed\u7684\u97f3\u9891\u91c7\u6837\u503c\u91cf\u5316\u4e3a\u79bb\u6563\u7684\u53d6\u503c\u533a\u95f4\u3002\u4e0e WaveNet \u7c7b\u4f3c\uff0c\u6211\u4eec\u91c7\u7528 \u03bc-law\uff08\u03bc\u5f8b\uff09\u7b97\u6cd5\uff0c\u5c06\u8fde\u7eed\u632f\u5e45\u6620\u5c04\u5230 256 \u4e2a\u79bb\u6563\u6876\uff08buckets\uff09\uff0c\u5e76\u5c06\u5176\u89c6\u4f5c 256 \u4e2a\u53ef\u80fd\u7684\u79bb\u6563 token\u3002<\/p>\n\n\n\n<p>\u5728\u6b64\u57fa\u7840\u4e0a\uff0c\u6211\u4eec\u4f7f\u7528\u8fd9\u79cd\u9010\u91c7\u6837\u70b9\u91cf\u5316\u540e\u7684\u97f3\u9891 token \u6765\u8bad\u7ec3\u4e00\u4e2a\u8bed\u8a00\u6a21\u578b\u3002\u6570\u636e\u96c6\u65b9\u9762\uff0c\u6cbf\u7528 <a href=\"https:\/\/arxiv.org\/abs\/2209.03143\" target=\"_blank\" rel=\"noreferrer noopener\">AudioLM<\/a>\uff08Neil Zeghidour \u4e0e Eugene Kharitonov\uff09\u7684\u8bbe\u7f6e\uff0c\u91c7\u7528 Libri-Light \u6570\u636e\u96c6\u3002\u8be5\u6570\u636e\u96c6\u7684\u8bad\u7ec3\u96c6\u603b\u65f6\u957f\u7ea6\u4e3a 5 \u4e07\u5c0f\u65f6\uff0c\u4f46\u5728\u672c\u5b9e\u9a8c\u4e2d\u6211\u4eec\u4ec5\u4f7f\u7528\u5176\u4e2d 1000 \u5c0f\u65f6\u7684\u6570\u636e\u3002\u91c7\u7528\u9010\u91c7\u6837\u70b9\u7684\u79bb\u6563\u5316\u65b9\u5f0f\u540e\uff0c\u6700\u7ec8\u5f97\u5230\u7684\u8bad\u7ec3\u6570\u636e\u89c4\u6a21\u7ea6\u4e3a 53GB\u3002<\/p>\n\n\n\n<p>\u6a21\u578b\u65b9\u9762\uff0c\u6211\u4eec\u8bad\u7ec3\u4e86\u4e00\u4e2a\u89c4\u6a21\u76f8\u5bf9\u8f83\u5c0f\u7684 Transformer\uff0c\u603b\u53c2\u6570\u91cf\u4e3a 151.28M\uff0c\u4e0e\u6700\u5c0f\u89c4\u683c\u7684 <a href=\"https:\/\/openai.com\/index\/better-language-models\/\" target=\"_blank\" rel=\"noreferrer noopener\">GPT-2 <\/a>\u6a21\u578b\u5927\u81f4\u76f8\u5f53\u3002\u5f53\u4ece\u8be5\u6a21\u578b\u8fdb\u884c\u91c7\u6837\u751f\u6210\u65f6\uff0c\u5176\u8f93\u51fa\u8868\u73b0\u4e3a\u7c7b\u4f3c\u54bf\u5440\u5b66\u8bed\u822c\u7684\u58f0\u97f3\uff08\u63d0\u793a\uff1a\u97f3\u91cf\u6709\u65f6\u8f83\u5927\uff09\u3002<\/p>\n\n\n\n<figure class=\"wp-block-audio\"><audio controls src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/02\/8e7c2b20-843f-4033-8dcd-3b70534b683d.wav\"><\/audio><\/figure>\n\n\n\n<p>\u6a21\u578b\u5f80\u5f80\u4f1a\u8fdb\u5165\u4e00\u79cd\u201c\u567c\u556a\u6742\u97f3\u6a21\u5f0f\u201d\uff08crackling mode\uff09\uff0c\u5e76\u4e14\u4e00\u65e6\u9677\u5165\u5176\u4e2d\uff0c\u4f3c\u4e4e\u96be\u4ee5\u81ea\u884c\u6062\u590d\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-audio\"><audio controls src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/02\/37518e7e-de7b-4891-b032-23767a2aa046.wav\"><\/audio><\/figure>\n\n\n\n<p>\u6211\u8fd8\u8bad\u7ec3\u4e86\u4e00\u4e2a\u8f83\u5c0f\u7684\u6a21\u578b\uff0c\u5c31\u662f\u4e4b\u524d\u5f00\u5934\u63d0\u5230\u7684\u90a3\u4e2a\u3002\u5b83\u5bb9\u6613\u751f\u6210\u4ee4\u4eba\u4e0d\u9002\u7684\u5c16\u9510\u566a\u58f0\uff08\u97f3\u91cf\u8f83\u5927\uff01\uff09\u3002<\/p>\n\n\n\n<figure class=\"wp-block-audio\"><audio controls src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/02\/9239ef01-1ab6-4617-875d-9cde216d604c.wav\"><\/audio><\/figure>\n\n\n\n<p>\u6b63\u5982\u4f60\u6240\u770b\u5230\u7684\uff0c\u6211\u4eec\u8fd8\u8fdc\u672a\u8fbe\u5230\u901a\u7528\u4eba\u5de5\u667a\u80fd\uff08AGI\uff09\u7684\u6c34\u5e73\u3002\u6a21\u578b\u751f\u6210\u7684\u97f3\u9891\u542c\u8d77\u6765\u50cf\u662f\u8bed\u97f3\uff0c\u4f46\u4f60\u51e0\u4e4e\u542c\u4e0d\u51fa\u4efb\u4f55\u5355\u8bcd\uff0c\u800c\u4e14\u58f0\u97f3\u4e5f\u5728\u4e0d\u65ad\u53d8\u5316\u3002\u8fd9\u4e5f\u4e0d\u8db3\u4e3a\u5947\uff1a\u6a21\u578b\u7684\u4e0a\u4e0b\u6587\u957f\u5ea6\u4e3a 2048\uff0c\u5bf9\u4e8e 16 kHz \u7684\u97f3\u9891\u800c\u8a00\uff0c\u4ec5\u76f8\u5f53\u4e8e 128 \u6beb\u79d2\uff0c\u8fde\u4e00\u4e2a\u5355\u8bcd\u7684\u957f\u5ea6\u90fd\u4e0d\u5230\u3002\u6b64\u5916\uff0c\u8fd9\u4e9b 10 \u79d2\u7684\u97f3\u9891\u6837\u672c\u5728 H100 \u4e0a\u751f\u6210\u8017\u65f6\u7ea6 30 \u5206\u949f\u2014\u2014\u8ddd\u79bb\u5b9e\u65f6\u751f\u6210\u8fd8\u6709\u51e0\u4e2a\u6570\u91cf\u7ea7\u7684\u5dee\u8ddd\u3002<\/p>\n\n\n\n<p>\u56e0\u6b64\uff0c\u6211\u4eec\u9700\u8981\u6784\u5efa\u4e00\u4e2a\u795e\u7ecf\u97f3\u9891\u7f16\u89e3\u7801\u5668\uff08neural audio codec\uff09\u6765\u538b\u7f29\u97f3\u9891\u3002\u601d\u8def\u662f\uff1a\u5982\u679c\u5c06\u91c7\u6837\u7387\u964d\u4f4e 100 \u500d\uff0c\u6a21\u578b\u751f\u6210\u7684\u5185\u5bb9\u4e5f\u6709\u671b\u53d8\u5f97\u201c100 \u500d\u66f4\u8fde\u8d2f\u201d\u3002\u5728\u673a\u5668\u5b66\u4e60\u4e2d\uff0c\u4e00\u4e2a\u7ecf\u5178\u65b9\u6cd5\u662f\u4f7f\u7528<strong>\u81ea\u7f16\u7801\u5668\uff08autoencoder\uff09<\/strong>\uff1a\u8be5\u6a21\u578b\u63a5\u6536\u8f93\u5165\uff0c\u5c06\u5176\u538b\u7f29\u5230\u8f83\u5c0f\u7684\u201c\u6f5c\u5728\u7a7a\u95f4\uff08latent space\uff09\u201d\uff0c\u7136\u540e\u5c1d\u8bd5\u91cd\u6784\u539f\u59cb\u8f93\u5165\u3002<\/p>\n\n\n\n<p>\u5728\u6211\u4eec\u7684\u573a\u666f\u4e0b\uff0c\u9700\u8981\u4e00\u4e2a<strong>\u6f5c\u5728\u8868\u793a\u53ef\u91cf\u5316\uff08quantized\uff09\u7684\u81ea\u7f16\u7801\u5668<\/strong>\uff0c\u8fd9\u6837\u624d\u80fd\u5c06\u6f5c\u5728\u5411\u91cf\u8f93\u5165\u5230\u8bed\u8a00\u6a21\u578b\u4e2d\uff0c\u5e76\u751f\u6210\u540e\u7eed\u5185\u5bb9\u3002<strong>\u5f53\u7136\uff0c\u4e5f\u53ef\u4ee5\u4f7f\u7528\u672a\u91cf\u5316\u7684\u6f5c\u5728\u5411\u91cf\u751f\u6210\u540e\u7eed\u97f3\u9891\uff0c\u4f46\u64cd\u4f5c\u4f1a\u66f4\u590d\u6742\u2014\u2014\u5177\u4f53\u53ef\u53c2\u89c1\u201c<a href=\"https:\/\/kyutai.org\/codec-explainer#further-reading\" target=\"_blank\" rel=\"noreferrer noopener\">\u8fdb\u4e00\u6b65\u9605\u8bfb<\/a>\u201d\u90e8\u5206\u3002<\/strong><\/p>\n\n\n\n<h3>Autoencoders with vector quantization (VQ-VAE)<\/h3>\n\n\n\n<p>\u8bf7\u8010\u5fc3\u4e00\u70b9\uff0c\u56e0\u4e3a\u6211\u4eec\u5c06\u4ece\u97f3\u9891\u9886\u57df\u7ed5\u4e2a\u5f2f\uff1a\u8ba9\u6211\u4eec\u57fa\u4e8e Fashion MNIST \u7684\u56fe\u50cf\u6765\u6784\u5efa\u4e00\u4e2a\u91cf\u5316\u7684\u81ea\u7f16\u7801\u5668\uff08quantized autoencoder\uff09\u3002\u6211\u4eec\u4f1a\u4f7f\u7528\u4e00\u4e2a\u5305\u542b\u524d\u4e09\u4e2a\u7c7b\u522b\u7684\u6570\u636e\u5b50\u96c6\uff1a<strong>T\u6064<\/strong>\u3001<strong>\u88e4\u5b50<\/strong>\u548c<strong>\u5957\u5934\u886b<\/strong>\u3002<\/p>\n\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter\"><img src=\"https:\/\/kyutai.org\/assets\/codec-explainer\/fashion-mnist-3.png\" alt=\"\"\/><\/figure><\/div>\n\n\n\n<p>\u9996\u5148\uff0c\u6211\u4eec\u5148\u8bad\u7ec3\u4e00\u4e2a<strong>\u666e\u901a\u81ea\u7f16\u7801\u5668<\/strong>\uff0c\u5c06\u56fe\u50cf\u7f16\u7801\u5230\u4e8c\u7ef4\u6f5c\u5728\u7a7a\u95f4\uff082D latent space\uff09\u4e2d\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-video\"><video controls src=\"https:\/\/kyutai.org\/assets\/codec-explainer\/vq_images_unquantized_v2.mp4\"><\/video><figcaption>Training a regular autoencoder on Fashion MNIST<\/figcaption><\/figure>\n\n\n\n<p>\u6bcf\u4e00\u5e27\u663e\u793a\u7684\u662f\u4e00\u4e2a\u8bad\u7ec3\u6279\u6b21\uff08\u6709\u4e9b\u6279\u6b21\u88ab\u7565\u8fc7\uff09\u3002\u5c0f\u56fe\u50cf\u8868\u793a\u81ea\u7f16\u7801\u5668\u5bf9\u8be5\u6279\u6b21\u56fe\u50cf\u7684\u91cd\u6784\u7ed3\u679c\u3002\u6211\u4e3a\u4e09\u7c7b\u56fe\u50cf\u6dfb\u52a0\u4e86\u989c\u8272\u6807\u8bb0\uff08T \u6064\/\u4e0a\u8863 = \u84dd\u8272\u3001\u88e4\u5b50 = \u7eff\u8272\u3001\u5957\u5934\u886b = \u9ec4\u8272\uff09\uff0c\u4f46\u81ea\u7f16\u7801\u5668\u5e76\u6ca1\u6709\u63a5\u6536\u7c7b\u522b\u4fe1\u606f\u4f5c\u4e3a\u8f93\u5165\u2014\u2014\u6f5c\u5728\u7a7a\u95f4\u81ea\u7136\u4f1a\u6839\u636e\u7c7b\u522b\u5f62\u6210\u805a\u7c7b\u3002\u63a5\u4e0b\u6765\uff0c\u6211\u4eec\u653e\u5927\u89c2\u5bdf\u51e0\u5f20\u91cd\u6784\u7ed3\u679c\uff1a<\/p>\n\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter\"><img src=\"https:\/\/kyutai.org\/assets\/codec-explainer\/rvq-without-quantization-v4.png\" alt=\"\"\/><figcaption>Original images (top) and their reconstructed versions (bottom)<\/figcaption><\/figure><\/div>\n\n\n\n<p>\u5982\u4f60\u6240\u89c1\uff0c\u91cd\u6784\u6548\u679c\u5e76\u4e0d\u7406\u60f3\u3002\u56fe\u50cf\u6bd4\u8f83\u6a21\u7cca\uff0c\u800c\u4e14\u524d\u4e24\u5f20\u91cd\u6784\u51e0\u4e4e\u5b8c\u5168\u76f8\u540c\u3002\u4f46\u6211\u4eec\u4f7f\u7528\u7684\u7f51\u7edc\u975e\u5e38\u5c0f\u2014\u2014\u7f16\u7801\u5668\u548c\u89e3\u7801\u5668\u5404\u53ea\u6709\u56db\u5c42\u5168\u8fde\u63a5\u7f51\u7edc\uff0c\u5e76\u4e14\u53ea\u5c06\u6570\u636e\u6295\u5f71\u5230\u4e8c\u7ef4\u7a7a\u95f4\uff0c\u56e0\u6b64\u4e0d\u80fd\u5bf9\u6a21\u578b\u671f\u671b\u8fc7\u9ad8\u3002<\/p>\n\n\n\n<p>\u63a5\u4e0b\u6765\uff0c\u6211\u4eec\u5c06\u5bf9\u8fd9\u4e9b\u6f5c\u5728\u5411\u91cf\u8fdb\u884c\u91cf\u5316\uff08quantization\uff09\uff0c\u65b9\u6cd5\u662f\u901a\u8fc7\u805a\u7c7b\u5b9e\u73b0\u3002\u5927\u81f4\u6b65\u9aa4\u5982\u4e0b\uff1a<\/p>\n\n\n\n<ol><li>\u7c7b\u4f3c <strong>k-means<\/strong>\uff1a\u7ef4\u62a4\u4e00\u4e2a\u7c07\u4e2d\u5fc3\uff08cluster center\uff09\u7684\u4f4d\u7f6e\u5217\u8868\u3002<\/li><li>\u7c07\u4e2d\u5fc3\u521d\u59cb\u5316\u4e3a\u968f\u673a\u4f4d\u7f6e\u3002<\/li><li>\u5bf9\u6bcf\u4e2a\u8bad\u7ec3\u6279\u6b21\uff1a<ul><li>\u67e5\u770b\u6bcf\u4e2a\u6f5c\u5728\u5411\u91cf\u5c5e\u4e8e\u54ea\u4e2a\u7c07\uff08assignment\uff09\uff0c\u6ce8\u610f\u6211\u4eec <strong>\u4e0d\u4fee\u6539\u6f5c\u5728\u5411\u91cf<\/strong>\uff0c\u4ec5\u8fdb\u884c\u5206\u914d\u3002<\/li><li>\u5c06\u6bcf\u4e2a\u7c07\u4e2d\u5fc3\u5411\u5176\u6240\u5c5e\u6f5c\u5728\u5411\u91cf\u7684\u5e73\u5747\u4f4d\u7f6e\u8f7b\u5fae\u79fb\u52a8\uff08nudge\uff09\u3002<\/li><\/ul><\/li><li>\u5982\u679c\u67d0\u4e2a\u7c07\u4e2d\u5fc3\u957f\u65f6\u95f4\u672a\u88ab\u5206\u914d\u5230\u4efb\u4f55\u6f5c\u5728\u5411\u91cf\uff0c\u5219\u5c06\u5176\u91cd\u65b0\u201c\u4f20\u9001\u201d\u5230\u5f53\u524d\u6279\u6b21\u7684\u67d0\u4e2a\u968f\u673a\u6f5c\u5728\u5411\u91cf\u4e0a\uff0c\u4ee5\u907f\u514d\u7c07\u4e2d\u5fc3\u9677\u5165\u5c40\u90e8\u505c\u6ede\u3002<\/li><\/ol>\n\n\n\n<figure class=\"wp-block-video\"><video controls src=\"https:\/\/kyutai.org\/assets\/codec-explainer\/vq_images_unquantized_with_clustering_v2.mp4\"><\/video><figcaption>Quantizing by fitting a clustering on top of the autoencoder<\/figcaption><\/figure>\n\n\n\n<p>\u4f60\u53ef\u4ee5\u770b\u5230\uff0c\u968f\u7740\u8bad\u7ec3\u7684\u8fdb\u884c\uff0c\u7c07\u4e2d\u5fc3\u7684\u91cd\u6784\u6548\u679c\u9010\u6e10\u88ab\u4f18\u5316\u548c\u7ec6\u5316\u3002<\/p>\n\n\n\n<p>\u63a5\u4e0b\u6765\uff0c\u6211\u4eec\u5e0c\u671b\u8ba9\u7f16\u7801\u5668\uff08encoder\uff09\u548c\u89e3\u7801\u5668\uff08decoder\uff09\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u66f4\u597d\u5730\u5904\u7406\u91cf\u5316\u540e\u7684\u6f5c\u5728\u5411\u91cf\uff08quantized embeddings\uff09\u3002\u76ee\u524d\uff0c\u6211\u4eec\u53ea\u662f\u5c06\u805a\u7c7b\u64cd\u4f5c\u53e0\u52a0\u5728\u4e00\u4e2a<strong>\u5bf9\u91cf\u5316\u201c\u672a\u77e5\u201d\u7684\u81ea\u7f16\u7801\u5668<\/strong>\u4e4b\u4e0a\u2014\u2014\u4e5f\u5c31\u662f\u8bf4\uff0c\u6a21\u578b\u5728\u8bad\u7ec3\u65f6\u5e76\u6ca1\u6709\u610f\u8bc6\u5230\u6f5c\u5728\u5411\u91cf\u4f1a\u88ab\u91cf\u5316\u3002\u6211\u4eec\u5e0c\u671b\u81ea\u7f16\u7801\u5668\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d<strong>\u9002\u5e94\u91cf\u5316\u64cd\u4f5c<\/strong>\uff0c\u4ece\u800c\u751f\u6210\u66f4\u6613\u91cd\u6784\u7684\u91cf\u5316\u8868\u793a\u3002\u76ee\u524d\u7684\u505a\u6cd5\u662f\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>x = get_batch()\nz = encoder(x)\nx_reconstructed = decoder(z)\nloss = reconstruction_loss(x, x_reconstructed)<\/code><\/pre>\n\n\n\n<p>\u4e0e\u5176\u5c06\u672a\u91cf\u5316\u7684\u6f5c\u5728\u5411\u91cf\u76f4\u63a5\u8f93\u5165\u89e3\u7801\u5668\uff0c\u6211\u4eec\u5148\u5c06\u5176\u6620\u5c04\u5230<strong>\u6700\u8fd1\u7684\u7c07\u4e2d\u5fc3<\/strong>\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>x = get_batch()\nz = encoder(x)\n\nz_quantized = to_nearest_cluster(z)     # \ud83d\udc48\nx_reconstructed = decoder(z_quantized)  # \ud83d\udc48\n\nloss = reconstruction_loss(x, x_reconstructed)<\/code><\/pre>\n\n\n\n<p>\u8fd9\u91cc\u6709\u4e00\u4e2a\u95ee\u9898\uff1a\u5982\u679c\u76f4\u63a5\u8fd9\u6837\u505a\uff0c\u81ea\u7f16\u7801\u5668\u5c31\u65e0\u6cd5\u7ee7\u7eed\u8bad\u7ec3\u4e86\u3002\u539f\u56e0\u662f<strong>\u91cf\u5316\u64cd\u4f5c\u4e0d\u53ef\u5fae<\/strong>\uff0c\u4e5f\u5c31\u662f\u8bf4\uff0c\u635f\u5931\u51fd\u6570\u7684\u68af\u5ea6\u65e0\u6cd5\u4f20\u56de\u7f16\u7801\u5668\u7684\u6743\u91cd\u3002\u672c\u8d28\u4e0a\uff0c\u6a21\u578b\u65e0\u6cd5\u56de\u7b54\u8fd9\u4e2a\u95ee\u9898\uff1a\u201c\u5982\u679c\u6211\u60f3\u8ba9\u635f\u5931\u51cf\u5c11\u4e00\u70b9\uff0c\u5e94\u8be5\u6cbf\u54ea\u4e2a\u65b9\u5411\u8c03\u6574\u7f16\u7801\u5668\u7684\u6743\u91cd\uff1f\u201d<\/p>\n\n\n\n<p>\u89e3\u51b3\u65b9\u6cd5\u5f88\u5de7\u5999\uff1a<strong>\u5047\u88c5\u8fd9\u4e2a\u95ee\u9898\u4e0d\u5b58\u5728<\/strong>\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u6211\u4eec\u5c06\u91cf\u5316\u540e\u7684\u6f5c\u5728\u5411\u91cf z<em><sub>quantized<\/sub><\/em> \u770b\u4f5c\u662f\u539f\u5411\u91cf z \u52a0\u4e0a\u4e00\u4e2a<strong>\u4efb\u610f\u5411\u91cf\uff0c\u4f46\u4e0d\u5f71\u54cd\u68af\u5ea6<\/strong>\u3002\u8fd9\u6837\uff0c  z<em><sub>quantized<\/sub><\/em> \u7684\u68af\u5ea6\u5c31\u7b49\u540c\u4e8e z \u7684\u68af\u5ea6\u3002\u8fd9\u5c31\u662f\u6240\u8c13\u7684 <strong>straight-through gradient estimator\uff08\u76f4\u901a\u68af\u5ea6\u4f30\u8ba1\u5668\uff09<\/strong> \u7684\u539f\u7406\u3002<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>x = get_batch()\nz = encoder(x)\n\nresidual = z - to_nearest_cluster(z)\n# .detach() means \"forget that this needs a gradient\"\nz_quantized = z - residual.detach()\nx_reconstructed = decoder(z_quantized)\n\nloss = reconstruction_loss(x, x_reconstructed)<\/code><\/pre>\n\n\n\n<p>\u5728\u524d\u5411\u4f20\u64ad\uff08forward pass\uff09\u4e2d\uff0c z<em><sub>quantized<\/sub><\/em> \u200b \u7684\u53d6\u503c\u4e0e\u4e4b\u524d\u76f8\u540c\uff0c\u4f46\u5173\u952e\u662f\uff1az \u7684\u68af\u5ea6\u73b0\u5728\u88ab\u8bbe\u7f6e\u4e3a\u7b49\u540c\u4e8e  z<em><sub>quantized<\/sub><\/em> \u200b \u200b \u7684\u68af\u5ea6\uff0c\u800c\u4e0d\u662f\u56e0\u4e3a\u4e0d\u53ef\u5fae\u7684 <code>to_nearest_cluster(z)<\/code> \u64cd\u4f5c\u800c\u4e3a 0\u3002<\/p>\n\n\n\n<p>\u8fd9\u79cd\u201c\u5047\u88c5\u201d\u505a\u6cd5\u662f\u6709\u4ee3\u4ef7\u7684\uff1a\u5728\u8bad\u7ec3\u65f6\uff0c\u7f16\u7801\u5668\u7684\u6743\u91cd\u4f1a\u6839\u636e\u91cd\u6784\u635f\u5931\u8fdb\u884c\u66f4\u65b0\uff0c\u4f46\u66f4\u65b0\u7684\u65b9\u5411\u662f<strong>\u5047\u8bbe\u91cf\u5316\u4e0d\u5b58\u5728\u7684\u65b9\u5411<\/strong>\uff0c\u56e0\u6b64\u4e0d\u4e00\u5b9a\u662f\u6700\u4f18\u7684\u68af\u5ea6\u65b9\u5411\u3002\u4f46\u53ea\u8981\u6f5c\u5728\u5411\u91cf\u5927\u81f4\u4fdd\u6301\u5728\u5404\u81ea\u7c07\u4e2d\u5fc3\u9644\u8fd1\uff0c\u8fd9\u4e2a\u68af\u5ea6\u65b9\u5411\u4ecd\u7136\u662f\u201c\u57fa\u672c\u6b63\u786e\u201d\u7684\u3002<\/p>\n\n\n\n<p>\u4e3a\u4e86\u8ba9\u7f16\u7801\u5668\u751f\u6210\u66f4\u5bb9\u6613\u91cf\u5316\u7684\u6f5c\u5728\u5411\u91cf\uff0c\u6211\u4eec\u53ef\u4ee5\u5f15\u5165<strong>\u627f\u8bfa\u635f\u5931\uff08commitment loss\uff09<\/strong>\uff1a<strong>\u5bf9\u6bcf\u4e2a\u6f5c\u5728\u5411\u91cf\u6839\u636e\u5176\u8ddd\u79bb\u7c07\u4e2d\u5fc3\u7684\u8fdc\u8fd1\u65bd\u52a0\u60e9\u7f5a<\/strong>\u3002\u8fd9\u4e00\u635f\u5931\u7684\u68af\u5ea6\u4f1a\u5c06\u6f5c\u5728\u5411\u91cf\u63a8\u5411\u5bf9\u5e94\u7684\u7c07\u4e2d\u5fc3\uff0c\u4ece\u800c\u63d0\u9ad8\u91cf\u5316\u53cb\u597d\u6027\u3002<\/p>\n\n\n\n<p>\u901a\u8fc7\u5728\u8bad\u7ec3\u65f6\u8fdb\u884c\u91cf\u5316\u5e76\u52a0\u5165\u627f\u8bfa\u635f\u5931\uff0c\u6a21\u578b\u4e0d\u518d\u53ea\u662f\u5355\u7eaf\u5728\u5d4c\u5165\u4e0a\u505a\u805a\u7c7b\uff0c\u800c\u662f<strong>\u81ea\u7f16\u7801\u5668\u672c\u8eab\u88ab\u8bad\u7ec3\u6210\u5bf9\u91cf\u5316\u53cb\u597d<\/strong>\uff0c\u4ece\u800c\u5728\u540e\u7eed\u751f\u6210\u548c\u538b\u7f29\u4e2d\u8868\u73b0\u66f4\u597d\u3002<\/p>\n\n\n\n<figure class=\"wp-block-video\"><video controls src=\"https:\/\/kyutai.org\/assets\/codec-explainer\/vq_images_balanced_v2.mp4\"><\/video><figcaption>An autoencoder trained explicitly to be easy to quantize<\/figcaption><\/figure>\n\n\n\n<p>\u4f60\u4f1a\u6ce8\u610f\u5230\uff0c\u8bad\u7ec3\u52a8\u6001\u53d1\u751f\u4e86\u53d8\u5316\uff1a\u52a0\u5165\u627f\u8bfa\u635f\u5931\uff08commitment loss\uff09\u4e3a\u6f5c\u5728\u5411\u91cf\u589e\u52a0\u4e86\u4e00\u5b9a\u7684\u201c\u7ea6\u675f\u529b\uff08stiffness\uff09\u201d\uff0c\u4f7f\u5b83\u4eec\u4e0d\u518d\u50cf\u4e4b\u524d\u90a3\u6837\u81ea\u7531\u79fb\u52a8\u3002<\/p>\n\n\n\n<p>\u4e0b\u9762\u662f\u4f7f\u7528<strong>\u91cf\u5316\u6f5c\u5728\u5411\u91cf<\/strong>\u8fdb\u884c\u91cd\u6784\u7684\u6548\u679c\uff1a<\/p>\n\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter\"><img src=\"https:\/\/kyutai.org\/assets\/codec-explainer\/rvq-1-level-v4.png\" alt=\"\"\/><\/figure><\/div>\n\n\n\n<p>\u6ce8\u610f\u524d\u4e24\u5f20\u56fe\u88ab\u91cd\u6784\u6210\u4e86\u5b8c\u5168\u76f8\u540c\u7684\u56fe\u50cf\u3002\u8fd9\u53ea\u662f\u56e0\u4e3a\u5b83\u4eec\u7684\u6f5c\u5728\u5411\u91cf\u88ab\u5206\u914d\u5230\u540c\u4e00\u4e2a\u7c07\uff0c\u56e0\u6b64\u91cf\u5316\u540e\u53d6\u5230\u4e86\u76f8\u540c\u7684\u503c\u3002<\/p>\n\n\n\n<p>\u8fd9\u91cc\u63cf\u8ff0\u7684\u6a21\u578b\u88ab\u79f0\u4e3a <strong>VQ-VAE\uff08\u5411\u91cf\u91cf\u5316\u53d8\u5206\u81ea\u7f16\u7801\u5668\uff09<\/strong>\u3002\u5176\u4e2d\u7684\u201c\u53d8\u5206\uff08variational\uff09\u201d\u4e00\u8bcd\u5728\u8fd9\u91cc\u5df2\u7ecf\u6ca1\u6709\u5b9e\u9645\u610f\u4e49\uff0c\u53ea\u662f\u5386\u53f2\u9057\u7559\u7684\u547d\u540d\u3002<\/p>\n\n\n\n<h3>Residual vector quantization<\/h3>\n\n\n\n<p>\u4e3a\u4e86\u63d0\u9ad8\u91cd\u6784\u7684\u4fdd\u771f\u5ea6\uff0c\u6211\u4eec\u53ef\u4ee5\u7b80\u5355\u5730\u589e\u52a0\u7c07\u4e2d\u5fc3\u7684\u6570\u91cf\u3002\u4f46\u5982\u679c\u7c07\u4e2d\u5fc3\u8fc7\u591a\uff0c\u8ba1\u7b97\u548c\u5185\u5b58\u5f00\u9500\u4f1a\u53d8\u5f97\u975e\u5e38\u6602\u8d35\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u91c7\u7528\u4e00\u4e2a\u5de7\u5999\u7684\u505a\u6cd5\uff1a\u5982\u679c\u5e0c\u671b\u6f5c\u5728\u5411\u91cf\u6709 2^20\uff08\u7ea6 100 \u4e07\uff09\u79cd\u53ef\u80fd\u503c\uff0c\u6211\u4eec\u4e0d\u4f1a\u76f4\u63a5\u521b\u5efa   2^20  \u4e2a\u7c07\u3002\u76f8\u53cd\uff0c\u6211\u4eec\u4f7f\u7528\u4e24\u4e2a\u72ec\u7acb\u7684\u91cf\u5316\u5668\uff08quantizer\uff09\uff0c\u6bcf\u4e2a\u91cf\u5316\u5668\u6709 2^10=1024 \u4e2a\u7c07\uff0c\u7136\u540e\u5c06\u5b83\u4eec\u7684\u7ed3\u679c\u7ec4\u5408\u8d77\u6765\u3002\u8fd9\u6837\uff0c\u6bcf\u4e2a\u6f5c\u5728\u5411\u91cf\u5c31\u91cf\u5316\u4e3a\u4e24\u4e2a\u6574\u6570\u7684\u5143\u7ec4\uff08\u6bcf\u4e2a\u5728 0\u20131023 \u4e4b\u95f4\uff09\uff0c\u603b\u5171\u6709 2^20 \u79cd\u53ef\u80fd\u7ec4\u5408\u3002<\/p>\n\n\n\n<p>\u90a3\u4e48\u5177\u4f53\u600e\u4e48\u505a\u5462\uff1f\u56de\u60f3\u4e00\u4e0b\u6211\u4eec\u5728\u76f4\u901a\u68af\u5ea6\u4f30\u8ba1\u5668\uff08straight-through estimator\uff09\u4e2d\u4f7f\u7528\u7684\u6b8b\u5dee\u53d8\u91cf\uff1a<em>residual=z\u2212to_nearest_cluster(z)<\/em><\/p>\n\n\n\n<p>\u5b83\u8868\u793a\u5728\u91cf\u5316\u5230\u6700\u8fd1\u7c07\u4e2d\u5fc3\u65f6\uff0c\u539f\u5411\u91cf <em>z <\/em>\u4e2d\u672a\u88ab\u6355\u6349\u5230\u7684\u90e8\u5206\u3002<\/p>\n\n\n\n<p>\u5bf9\u4e8e\u6279\u6b21\u4e2d\u7684\u6bcf\u4e2a\u6f5c\u5728\u5411\u91cf\uff0c\u6211\u4eec\u90fd\u6709\u5bf9\u5e94\u7684\u6b8b\u5dee\u5411\u91cf\u3002\u89e3\u51b3\u65b9\u6848\u5f88\u81ea\u7136\uff1a<strong>\u7528\u4e0e\u539f\u59cb\u6f5c\u5728\u5411\u91cf\u76f8\u540c\u7684\u65b9\u6cd5\u5bf9\u6b8b\u5dee\u5411\u91cf\u8fdb\u884c\u91cf\u5316<\/strong>\uff0c\u901a\u8fc7\u8bad\u7ec3\u53e6\u4e00\u4e2a\u5411\u91cf\u91cf\u5316\u5668\u5b9e\u73b0\u3002<\/p>\n\n\n\n<p>\u8fd9\u4e00\u6b21\uff0c\u7531\u4e8e\u6211\u4eec\u9700\u8981\u7ec4\u5408\u4e24\u4e2a\u91cf\u5316\u5668\uff0c\u5355\u4e2a\u91cf\u5316\u5668\u7684\u4e8c\u7ef4\u7c07\u4f4d\u7f6e\u5e76\u4e0d\u518d\u5bf9\u5e94\u56fe\u50cf\uff0c\u56e0\u6b64\u6211\u4eec\u5c06\u5176\u53ef\u89c6\u5316\u4e3a<strong>\u70b9\u7684\u5206\u5e03<\/strong>\u5373\u53ef\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-video\"><video controls src=\"https:\/\/kyutai.org\/assets\/codec-explainer\/rvq_fmnist.mp4\"><\/video><figcaption>\u8fd9\u662f\u4e8c\u7ea7\u91cf\u5316\uff08two-level quantization\uff09<strong>\u7684\u601d\u8def\uff1a\u5728\u7b2c\u4e00\u7ea7\u91cf\u5316\u5668\u91cf\u5316\u540e\u7684<\/strong>\u6b8b\u5dee\uff08residuals\uff0c\u4e5f\u5c31\u662f\u7b2c\u4e00\u7ea7\u91cf\u5316\u5668\u7684\u8bef\u5dee\uff09\u4e0a\uff0c\u518d\u8bad\u7ec3\u4e00\u4e2a\u91cf\u5316\u5668\u8fdb\u884c\u8fdb\u4e00\u6b65\u91cf\u5316\uff0c\u4ece\u800c\u66f4\u7cbe\u7ec6\u5730\u8868\u793a\u6f5c\u5728\u5411\u91cf\u3002<\/figcaption><\/figure>\n\n\n\n<p>\u8fd9\u6837\uff0c\u6bcf\u5f20\u56fe\u50cf\u5c31\u53ef\u4ee5\u7528<strong>\u6f5c\u5728\u5411\u91cf\u6240\u5728\u7c07\u7684\u7d22\u5f15<\/strong>\u548c<strong>\u6b8b\u5dee\u7c07\u7684\u7d22\u5f15<\/strong>\u6765\u8868\u793a\u3002\u63a5\u4e0b\u6765\uff0c\u6211\u4eec\u7528\u8fd9\u4e2a\u4e8c\u7ea7\u91cf\u5316\u5668\u5c1d\u8bd5\u91cd\u6784\u51e0\u5f20\u56fe\u50cf\uff1a<\/p>\n\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter is-resized\"><img loading=\"lazy\" src=\"https:\/\/kyutai.org\/assets\/codec-explainer\/rvq-2-level-v4.png\" alt=\"\" width=\"490\" height=\"369\"\/><figcaption>\u539f\u59cb\u56fe\u50cf\uff08\u9876\u90e8\uff09\u3001\u4e00\u7ea7\u91cf\u5316\u91cd\u6784\uff08\u4e2d\u90e8\uff09\u3001\u4e8c\u7ea7\u91cf\u5316\u91cd\u6784\uff08\u5e95\u90e8\uff09\u3002\u8fd9\u4e9b\u56fe\u50cf\u5728\u4e8c\u7ea7\u91cf\u5316\u4e0b\u5206\u522b\u88ab\u7f16\u7801\u4e3a\u7d22\u5f15\u5bf9\uff1a(4, 3)\u3001(4, 5)\u3001(16, 21) \u548c (30, 3)\u3002<\/figcaption><\/figure><\/div>\n\n\n\n<p>\u524d\u4e24\u5f20\u56fe\u50cf\u7684\u91cd\u6784\u4ecd\u7136\u76f8\u4f3c\uff0c\u4f46\u4e0d\u518d\u5b8c\u5168\u76f8\u540c\uff1a\u7b2c\u4e00\u5f20\u56fe\u88ab\u7f16\u7801\u4e3a (4, 3)\uff0c\u7b2c\u4e8c\u5f20\u56fe\u4e3a (4, 5)\u3002\u6362\u53e5\u8bdd\u8bf4\uff0c\u5b83\u4eec\u5728<strong>\u7b2c\u4e00\u7ea7\u91cf\u5316\u5668<\/strong>\u4e0a\u4f7f\u7528\u76f8\u540c\u7684 token\uff0c\u4f46\u5728\u6b8b\u5dee\u7684\u91cf\u5316\u4e0a\u6709\u6240\u4e0d\u540c\u3002\u7531\u4e8e\u5dee\u5f02\u8f83\u4e3a\u5fae\u5c0f\uff0c\u4e0b\u9762\u662f<strong>\u4e00\u7ea7\u91cf\u5316\u4e0e\u4e8c\u7ea7\u91cf\u5316\u91cd\u6784\u6548\u679c\u7684\u5bf9\u6bd4<\/strong>\uff1a<\/p>\n\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter\"><img src=\"https:\/\/kyutai.org\/assets\/codec-explainer\/rvq-2-level-diff-v3.png\" alt=\"\"\/><\/figure><\/div>\n\n\n\n<p>\u6211\u60f3\u5f3a\u8c03\u7684\u662f\uff0c\u7b2c\u4e8c\u7ea7\u91cf\u5316\u4f5c\u7528\u5728\u6f5c\u5728\u5411\u91cf\uff08embedding\uff09\u4e0a\uff0c\u800c\u4e0d\u662f\u76f4\u63a5\u4fee\u6539\u8f93\u51fa\u50cf\u7d20\u3002\u8fd9\u4e00\u70b9\u53ef\u4ee5\u4ece\u6700\u5de6\u548c\u6700\u53f3\u7684\u56fe\u50cf\u770b\u51fa\uff0c\u5b83\u4eec\u5206\u522b\u7f16\u7801\u4e3a (4, 3) \u548c (30, 3)\u3002\u4e5f\u5c31\u662f\u8bf4\uff0c\u5b83\u4eec\u4f7f\u7528\u4e86\u76f8\u540c\u7684\u6b8b\u5dee\u7f16\u7801 3\uff0c\u4f46\u5bf9\u4e24\u5f20\u91cd\u6784\u56fe\u50cf\u7684\u5f71\u54cd\u5374\u4e0d\u540c\u3002<\/p>\n\n\n\n<p>\u663e\u7136\uff0c\u91cd\u6784\u6548\u679c\u4ecd\u7136\u4e0d\u591f\u7cbe\u786e\u3002\u6f5c\u5728\u5411\u91cf\u672a\u91cf\u5316\u65f6\u7684\u91cd\u6784\u8d28\u91cf\u624d\u662f\u4e0a\u9650\uff0c\u56e0\u6b64\u5982\u679c\u81ea\u7f16\u7801\u5668\u672c\u8eab\u6027\u80fd\u4e0d\u597d\uff08\u6211\u4eec\u7684\u5c31\u662f\u5982\u6b64\uff09\uff0c\u6539\u8fdb\u91cf\u5316\u65b9\u6cd5\u4e5f\u65e0\u6cd5\u5e26\u6765\u663e\u8457\u63d0\u5347\u3002<\/p>\n\n\n\n<p>\u5230\u8fd9\u91cc\u6211\u4eec\u5148\u6682\u505c\uff0c\u4f46\u8fd9\u4e2a\u601d\u8def\u7684\u81ea\u7136\u5ef6\u4f38\u662f<strong>\u8d85\u8fc7\u4e24\u7ea7\u91cf\u5316<\/strong>\uff1a\u5bf9\u4e8c\u7ea7\u91cd\u6784\u7684\u6b8b\u5dee\u7ee7\u7eed\u91cf\u5316\uff0c\u4f9d\u6b21\u6211\u4eec\u5728\u8fd9\u91cc\u5c31\u5148\u505c\u4e0b\u3002\u4e0d\u8fc7\uff0c\u8fd9\u4e2a\u601d\u8def\u7684\u4e00\u4e2a\u81ea\u7136\u6269\u5c55\u662f\u5f15\u5165\u591a\u4e8e\u4e24\u7ea7\u7684\u91cf\u5316\u3002\u53ea\u9700\u8981\u5bf9 two-level reconstruction \u7684 residual \u518d\u6b21\u8fdb\u884c\u91cf\u5316\uff0c\u5982\u6b64\u53cd\u590d\u5373\u53ef\u3002\u8fd9\u4e2a\u5e7f\u4e49\u7684 Residual Vector Quantization \u7b97\u6cd5\u5f62\u5f0f\u5982\u4e0b\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>def rvq_quantize(z):\n    residual = z\n    codes = &#091;]\n\n    for level in range(levels):\n        quantized, cluster_i = to_nearest_cluster(level, residual)\n        residual -= quantized\n        codes.append(cluster_i)\n\n    return codes<\/code><\/pre>\n\n\n\n<p>\u6b8b\u5dee\u5411\u91cf\u91cf\u5316\uff08Residual Vector Quantization, RVQ\uff09\u6700\u65e9\u5728 <strong>SoundStream<\/strong> \u4e2d\u5e94\u7528\u4e8e\u795e\u7ecf\u97f3\u9891\u7f16\u89e3\u7801\u5668\uff0c\u4f46\u8fd9\u4e00\u601d\u60f3\u5176\u5b9e\u65e9\u5728 1980 \u5e74\u4ee3\u5c31\u5df2\u51fa\u73b0\u3002<\/p>\n\n\n\n<h3>Now let\u2019s tokenize audio<\/h3>\n\n\n\n<p>\u5c06 RVQ \u5e94\u7528\u4e8e\u97f3\u9891\u662f\u76f8\u5f53\u76f4\u63a5\u7684\u3002\u4f5c\u4e3a\u6211\u4eec\u7684\u81ea\u7f16\u7801\u5668\uff08autoencoder\uff09\uff0c\u6211\u4eec\u5c06\u4f7f\u7528\u4e00\u4e2a\u7c7b\u4f3c\u4e8e Jukebox \u6240\u4f7f\u7528\u7684\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\uff08CNN\uff09\u3002\u8fd9\u91cc\u7684\u67b6\u6784\u7ec6\u8282\u5e76\u4e0d\u592a\u91cd\u8981\u3002\u91cd\u8981\u7684\u662f\uff0c\u8fd9\u662f\u4e00\u4e2a\u80fd\u63a5\u6536\u00a0<code>t<\/code>\u00a0\u4e2a\u91c7\u6837\u70b9\u7684\u97f3\u9891\uff0c\u5e76\u5c06\u5176\u8f6c\u6362\u4e3a\u5f62\u72b6\u4e3a\u00a0<code>(t\/128, 32)<\/code>\u00a0\u7684\u5411\u91cf\u7684\u7f51\u7edc\u3002\u6362\u53e5\u8bdd\u8bf4\uff0c\u5b83\u4ee5 128 \u500d\u7684\u7cfb\u6570\u8fdb\u884c\u964d\u91c7\u6837\uff08downsamples\uff09\uff0c\u5e76\u4e3a\u6211\u4eec\u63d0\u4f9b 32 \u7ef4\u7684\u6d6e\u70b9\u6570\u8868\u793a\u3002\u7136\u540e\uff0c\u89e3\u7801\u5668\uff08decoder\uff09\u63a5\u6536\u8fd9\u4e2a\u00a0<code>(t\/128, 32)<\/code>\u00a0\u7684\u5d4c\u5165\uff08embeddings\uff09\uff0c\u5e76\u5c06\u5b83\u4eec\u89e3\u7801\u56de\u00a0<code>t<\/code>\u00a0\u4e2a\u91c7\u6837\u70b9\u3002<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>audio = get_batch()               # shape: &#091;B, T]\nz = encoder(audio)                # shape: &#091;B, T\/128, 32]\naudio_reconstructed = decoder(z)  # shape: &#091;B, T]<\/code><\/pre>\n\n\n\n<p>\u548c\u4e4b\u524d\u4e00\u6837\uff0c\u6211\u4eec\u5c06\u5728\u7f16\u7801\u5668\u4e4b\u540e\u6dfb\u52a0\u4e00\u4e2a RVQ\u3002\u4e0e\u5904\u7406\u56fe\u50cf\u7684\u552f\u4e00\u533a\u522b\u662f\uff0c\u5bf9\u4e8e\u6bcf\u4e2a\u97f3\u9891\u6837\u672c\uff0c\u6211\u4eec\u6709\u00a0<code>t\/128<\/code>\u00a0\u4e2a\u5d4c\u5165\u5411\u91cf\uff0c\u800c\u4e0d\u4ec5\u4ec5\u662f\u50cf\u56fe\u50cf\u90a3\u6837\u53ea\u6709\u4e00\u4e2a\u3002\u6211\u4eec\u53ea\u9700\u72ec\u7acb\u5730\u5bf9\u8fd9\u4e9b\u5411\u91cf\u8fdb\u884c\u91cf\u5316\uff08\u5373\u4f7f\u7f16\u7801\u5668\u201c\u770b\u5230\u201d\u7684\u97f3\u9891\u8303\u56f4\u6bd4\u5355\u4e2a\u5411\u91cf\u6240\u5bf9\u5e94\u7684\u8303\u56f4\u8981\u5e7f\uff09\u3002\u5728\u8bad\u7ec3\u671f\u95f4\uff0c\u6211\u4eec\u8fd8\u6709\u4e00\u4e2a\u6279\u6b21\u7ef4\u5ea6\uff08batch dimension\uff09\uff0c\u6240\u4ee5\u6211\u4eec\u7684\u6a21\u578b\u73b0\u5728\u770b\u8d77\u6765\u662f\u8fd9\u6837\u7684\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>audio = get_batch()                         # &#091;B, T]\nz = encoder(audio)                          # &#091;B, T\/128, 32]\n\n# Combine the batch and time dimensions\nz = rearrange(                              # &#091;B*T\/128, 32]\n    z, \"b t_emb d -&gt; (b t_emb) d\"\n)\n\ncodes = rvq_quantize(z)           # integers, &#091;B*T\/128, levels]\nz_quantized = codes_to_embeddings(codes)    # &#091;B*T\/128, 32]\nz_quantized = rearrange(                    # &#091;B, T\/128, 32]\n    z_quantized, \"(b t_emb) d -&gt; b t_emb d\"\n)\n\naudio_reconstructed = decoder(z_quantized)  # &#091;B, T]<\/code><\/pre>\n\n\n\n<figure class=\"wp-block-video\"><video controls src=\"https:\/\/kyutai.org\/assets\/codec-explainer\/codecWithRvq.mp4\"><\/video><\/figure>\n\n\n\n<p>\u5728\u6211\u4eec\u8bad\u7ec3\u7b2c\u4e00\u4e2a\u795e\u7ecf\u97f3\u9891\u7f16\u89e3\u7801\u5668\uff08neural audio codec\uff09\u4e4b\u524d\uff0c\u6700\u540e\u7f3a\u5c11\u7684\u4e00\u5757\u662f\u635f\u5931\u51fd\u6570\uff08loss function\uff09\u3002\u5173\u4e8e\u9009\u62e9\u54ea\u4e00\u4e2a\u635f\u5931\u51fd\u6570\uff0c\u6211\u4eec\u53ef\u4ee5\u6df1\u5165\u63a2\u8ba8\u4e00\u6574\u5957\u590d\u6742\u7406\u8bba\uff0c\u4f46\u6211\u4eec\u5c06\u907f\u5f00\u5b83\uff0c\u53ea\u4f7f\u7528\u4e00\u4e2a\u975e\u5e38\u7b80\u5355\u7684\u3002\u6211\u4eec\u4f1a\u8ba1\u7b97\u539f\u59cb\u97f3\u9891\u548c\u91cd\u5efa\u97f3\u9891\u7684\u5bf9\u6570\u632f\u5e45\u8c31\u56fe\uff08log amplitude spectrogram\uff09\uff0c\u7136\u540e\u53d6\u5b83\u4eec\u7684\u5dee\u503c\u3002\u8fd9\u4e2a\u5dee\u503c\u7684\u5747\u65b9\u503c\u5c31\u662f\u635f\u5931\u3002<\/p>\n\n\n\n<p>\u4e3a\u4e86\u8ba9\u6a21\u578b\u66f4\u96be\u5bf9\u8fd9\u4e2a\u635f\u5931\u51fd\u6570\u8fc7\u62df\u5408\uff0c\u6211\u4eec\u4f7f\u7528\u4e09\u79cd\u4e0d\u540c\u7684\u77ed\u65f6\u5085\u91cc\u53f6\u53d8\u6362\uff08short-time Fourier transform\uff09\u53c2\u6570\u6765\u8ba1\u7b97\u8c31\u56fe\uff0c\u5e76\u5c06\u6211\u4eec\u7684\u635f\u5931\u8bbe\u4e3a\u8fd9\u4e09\u4e2a\u5b50\u635f\u5931\u7684\u5e73\u5747\u503c\u3002\u8fd9\u88ab\u79f0\u4e3a<strong>\u591a\u5c3a\u5ea6\u9891\u8c31\u635f\u5931\uff08multi-scale spectral loss\uff09<\/strong>\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img src=\"https:\/\/kyutai.org\/assets\/codec-explainer\/image%202.png\" alt=\"\"\/><\/figure>\n\n\n\n<p>\u6700\u540e\uff0c\u8ba9\u6211\u4eec\u6765\u8bad\u7ec3\u4e00\u4e9b\u7f16\u89e3\u7801\u5668\uff08codecs\uff09\u5427\uff01\u6211\u4eec\u5c06\u89c2\u5bdf\u6539\u53d8 RVQ \u7684\u5c42\u7ea7\uff08levels\uff09\u6570\u91cf\u5982\u4f55\u5f71\u54cd\u91cd\u5efa\u8d28\u91cf\u3002\u6b63\u5982\u6211\u4eec\u6240\u9884\u671f\u7684\uff0c\u589e\u52a0\u5c42\u7ea7\u6570\u91cf\u6709\u52a9\u4e8e\u964d\u4f4e\u9891\u8c31\u635f\u5931\uff08spectral loss\uff09\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img src=\"https:\/\/kyutai.org\/assets\/codec-explainer\/image%203.png\" alt=\"\"\/><\/figure>\n\n\n\n<p>\u8ba9\u6211\u4eec\u542c\u542c\u8fd9\u4e9b\u7f16\u89e3\u7801\u5668\u542c\u8d77\u6765\u600e\u4e48\u6837\u3002\u6211\u4eec\u5c06\u4f7f\u7528\u8fd9\u4e09\u4e2a\u7f16\u89e3\u7801\u5668\u6765\u91cd\u5efa\u6765\u81ea Expresso \u6570\u636e\u96c6\u7684\u8fd9\u6bb5\u97f3\u9891\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-audio\"><audio controls src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/02\/dd51ff47-895e-4241-bbed-5b39158a5a30.wav\"><\/audio><figcaption>\u539f\u59cb\u97f3\u9891<\/figcaption><\/figure>\n\n\n\n<p>\u91cd\u5efa\u7ed3\u679c\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-audio\"><audio controls src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/02\/3332cbff-95d6-4e95-a29f-fa23f6bd1e1a.wav\"><\/audio><figcaption>4 RVQ levels<font class=\"notranslate immersive-translate-target-wrapper\" lang=\"zh-CN\"><font class=\"notranslate\" data-immersive-translate-translation-element-mark=\"1\">\u00a0<\/font><\/font><\/figcaption><\/figure>\n\n\n\n<figure class=\"wp-block-audio\"><audio controls src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/02\/4720a151-0f23-4587-952f-a3cb9feee2b4.wav\"><\/audio><figcaption> 8 RVQ levels <\/figcaption><\/figure>\n\n\n\n<figure class=\"wp-block-audio\"><audio controls src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/02\/899a76a5-8318-4440-acfd-0f67c7dd99a1.wav\"><\/audio><figcaption>12RVQ levels<\/figcaption><\/figure>\n\n\n\n<p>\u663e\u7136\uff0c\u968f\u7740\u589e\u52a0\u66f4\u591a\u7684\u6b8b\u5dee\u5411\u91cf\u91cf\u5316\uff08RVQ\uff09\u7ea7\u6570\uff0c\u97f3\u9891\u8d28\u91cf\u9010\u6e10\u63d0\u5347\u3002<\/p>\n\n\n\n<p>\u5373\u4fbf\u4f7f\u7528 16 \u7ea7\u91cf\u5316\uff0c\u4ecd\u7136\u4f1a\u51fa\u73b0\u4e00\u4e9b\u567c\u556a\u6742\u97f3\uff0c\u97f3\u9891\u542c\u8d77\u6765\u6709\u4e9b\u95f7\uff0c\u5e76\u4f34\u968f\u6301\u7eed\u7684\u9ad8\u9891\u566a\u58f0\u3002\u540e\u7eed\u6211\u4eec\u4f1a\u8ba8\u8bba\u8fdb\u4e00\u6b65\u6539\u8fdb\u7f16\u89e3\u7801\u5668\u7684\u65b9\u6cd5\uff0c\u4f46\u51fa\u4e8e\u6f14\u793a\u76ee\u7684\uff0c\u76ee\u524d\u7684\u6548\u679c\u5df2\u7ecf\u8db3\u591f\u3002<\/p>\n\n\n\n<h3><strong>\u4e3a\u4ec0\u4e48\u8981\u5173\u5fc3\u97f3\u9891<\/strong><\/h3>\n\n\n\n<p>\u6240\u4ee5\u73b0\u5728\u6211\u4eec\u6709\u4e86\u4e00\u4e2a <strong>\u795e\u7ecf\u97f3\u9891\u7f16\u89e3\u7801\u5668<\/strong>\uff1a\u6211\u4eec\u53ef\u4ee5\u628a\u97f3\u9891\u8f6c\u6362\u6210\u9002\u5408 LLM \u7684 token\uff0c\u7136\u540e\u518d\u8fd8\u539f\u56de\u97f3\u9891\u3002\u8fd9\u91cc\u7684 \u201cCodec\u201d \u672c\u8d28\u4e0a\u5c31\u662f\u97f3\u9891\u7684 <strong>\u5206\u8bcd\u5668<\/strong>\uff08tokenizer\uff09\uff0c\u4f46\u6211\u4eec\u7528 \u201ccodec\u201d \u8fd9\u4e2a\u8bcd\uff0c\u662f\u56e0\u4e3a\u5b83\u5728\u7ecf\u5178\u538b\u7f29\u683c\u5f0f\uff08\u6bd4\u5982 MP3\uff09\u91cc\u5df2\u7ecf\u88ab\u4f7f\u7528\u3002\u6211\u4f1a\u628a codec \u548c tokenizer \u4ea4\u66ff\u4f7f\u7528\u3002<\/p>\n\n\n\n<p>\u56de\u5230\u6211\u4eec\u6700\u521d\u60f3\u505a\u7684\u4e8b\u60c5\uff1a<strong>\u5efa\u6a21\u97f3\u9891<\/strong>\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u6211\u4eec\u8981\u505a\u4e00\u4e2a\u6a21\u578b\uff0c\u5b83\u53ef\u4ee5\u63a5\u53d7\u4e00\u6bb5\u97f3\u9891\u524d\u7f00\uff0c\u7136\u540e\u751f\u6210\u4e00\u4e2a\u5408\u7406\u7684\u7eed\u6bb5\u3002<\/p>\n\n\n\n<p>\u63d0\u9192\u4e00\u4e0b\uff0c\u6211\u4eec\u7684\u76ee\u6807\u662f\u8bad\u7ec3\u4f18\u79c0\u7684\u97f3\u9891 LLM\uff0c\u4f7f\u6a21\u578b\u80fd\u591f <strong>\u539f\u751f\u7406\u89e3\u5e76\u751f\u6210\u8bed\u97f3<\/strong>\uff0c\u7406\u89e3\u60c5\u7eea\u3001\u91cd\u97f3\u7b49\u7279\u5f81\u3002\u5b83\u4eec\u8fd8\u53ef\u4ee5\u8fdb\u4e00\u6b65\u5fae\u8c03\u6210 <strong>\u6587\u672c\u8f6c\u8bed\u97f3\u3001\u8bed\u97f3\u8f6c\u6587\u672c\u3001\u7ffb\u8bd1\u6a21\u578b<\/strong> \u7b49\u3002<\/p>\n\n\n\n<p>\u65e2\u7136\u4f60\u5df2\u7ecf\u76f8\u4fe1\u97f3\u9891 LLM \u662f\u901a\u5411 AGI \u7684\u8def\u5f84\uff0c\u90a3\u6211\u4eec\u5c31\u5f00\u59cb\u8bad\u7ec3\u51e0\u4e2a\u6a21\u578b\u5427\u3002<\/p>\n\n\n\n<p>\u5728\u6570\u636e\u96c6\u65b9\u9762\uff0c\u6211\u4eec\u5c06\u4f7f\u7528 <strong>Libri-Light<\/strong>\uff0c\u5c31\u50cf\u4e4b\u524d\u8bad\u7ec3\u9010\u6837\u672c\u6a21\u578b\u65f6\u7528\u7684\u90a3\u6837\u3002\u8fd9\u4e00\u6b21\u6211\u4eec\u4f1a\u4f7f\u7528 <strong>10000 \u5c0f\u65f6\u97f3\u9891<\/strong>\uff0c\u800c\u4e0d\u662f\u4e4b\u524d\u7684 1000 \u5c0f\u65f6\u3002\u8fd9\u4e2a\u6570\u636e\u96c6\u662f <strong>\u516c\u5171\u9886\u57df\u6709\u58f0\u4e66<\/strong>\uff0c\u6240\u4ee5\u5982\u679c\u6211\u4eec\u8bad\u7ec3\u51fa\u4e86\u4e00\u4e2a\u4e0d\u9519\u7684\u6a21\u578b\uff0c\u4e5f\u8bb8\u80fd\u751f\u6210\u66f4\u591a\u6545\u4e8b\uff08\u4e0d\u8fc7\u4e0d\u8981\u62b1\u592a\u5927\u5e0c\u671b\uff09\u3002\u6211\u4eec\u552f\u4e00\u9700\u8981\u505a\u7684\uff0c\u5c31\u662f\u628a\u97f3\u9891\u6570\u636e\u96c6\u8f6c\u6362\u6210 <strong>\u79bb\u6563 token \u5e8f\u5217<\/strong>\uff0c\u4ee5\u4fbf\u8f93\u5165\u5230 LLM \u4e2d\u3002<\/p>\n\n\n\n<h3><strong>\u5904\u7406\u591a\u4e2a\u5c42\u7ea7\uff08levels\uff09<\/strong><\/h3>\n\n\n\n<p>\u6211\u4eec\u5c06\u4f7f\u7528\u6211\u4eec\u7684 8 \u5c42 RVQ codec \u6765\u5b9e\u73b0\u8fd9\u4e00\u70b9\u3002\u5bf9\u4e8e\u4e00\u4e2a\u6709 t \u4e2a\u91c7\u6837\u70b9\u7684\u97f3\u9891\uff0c\u6211\u4eec\u5c06\u5f97\u5230\u4e00\u4e2a\u5f62\u72b6\u4e3a (t\/128, 8) \u7684 token \u6570\u7ec4\u3002\u4f46\u73b0\u5728\u6709\u4e00\u4e2a\u95ee\u9898\uff1a\u5982\u4f55\u5904\u7406\u5728\u6bcf\u4e2a\u65f6\u95f4\u6b65\uff08time step\uff09\u4e0d\u662f\u4e00\u4e2a\u800c\u662f 8 \u4e2a token \u7684\u60c5\u51b5\uff1f\u5728\u6587\u672c LLM \u4e2d\u6211\u4eec\u65e0\u9700\u5904\u7406\u8fd9\u4e2a\u95ee\u9898\uff0c\u56e0\u4e3a\u6211\u4eec\u53ea\u6709\u4e00\u4e2a token \u5e8f\u5217\u3002<\/p>\n\n\n\n<p>\u6211\u4eec\u5c06\u91c7\u53d6\u6700\u7b80\u5355\u7684\u65b9\u6cd5\uff0c\u76f4\u63a5\u5c06\u8be5\u6570\u7ec4\u5c55\u5e73\uff08flatten\uff09\u6210\u4e00\u4e2a\u5f62\u72b6\u4e3a (t\/128 * 8) \u7684\u4e00\u7ef4\u6570\u7ec4\uff0c\u5e76\u8ba9\u6211\u4eec\u7684 LLM \u5728\u4e0d\u540c\u7684\u65f6\u95f4\u6b65\u4e2d\u9884\u6d4b\u8fd9\u516b\u4e2a\u5c42\u7ea7\u3002<\/p>\n\n\n\n<figure class=\"wp-block-video\"><video controls src=\"https:\/\/kyutai.org\/assets\/codec-explainer\/flattenRvq.mp4\"><\/video><figcaption>lattening a three-level RVQ to allow it to be fed into a language mode<\/figcaption><\/figure>\n\n\n\n<p>\u8fd9\u6837\u505a\u7684\u4e00\u5927\u7f3a\u70b9\u662f\u6211\u4eec\u635f\u5931\u4e86\u4e00\u90e8\u5206\u65f6\u95f4\u538b\u7f29\uff08temporal compression\uff09\u80fd\u529b\u3002\u6211\u4eec\u5c06\u97f3\u9891\u964d\u91c7\u6837\u4e86 128 \u500d\uff0c\u4f46\u73b0\u5728\u901a\u8fc7\u5c55\u5e73\u5c42\u7ea7\uff0c\u53c8\u5c06\u5176\u201c\u81a8\u80c0\u201d\u4e86 8 \u500d\u3002\u8fd9\u4f7f\u5f97\u63a8\u7406\uff08inference\uff09\u6548\u7387\u964d\u4f4e\uff0c\u5e76\u4e14\u53ef\u80fd\u5bfc\u81f4\u8d28\u91cf\u4e0b\u964d\uff0c\u56e0\u4e3a\u6709\u6548\u4e0a\u4e0b\u6587\u5927\u5c0f\uff08context size\uff09\u51cf\u5c0f\u4e86\u3002\u6211\u4eec\u5c06\u4f7f\u7528 8 \u5c42\u7684 RVQ codec \u800c\u4e0d\u662f 16 \u5c42\u7684\uff0c\u4ee5\u907f\u514d\u8ba9\u538b\u7f29\u60c5\u51b5\u53d8\u5f97\u66f4\u7cdf\u3002<\/p>\n\n\n\n<p>\u4f60\u4e5f\u53ef\u4ee5\u4e00\u6b21\u6027\u9884\u6d4b\u5355\u4e2a\u65f6\u95f4\u6b65\u7684\u6240\u6709 RVQ \u5c42\u7ea7\uff08\u201c\u5e76\u884c\u6a21\u5f0f\u201d\uff0cparallel pattern\uff09\uff0c\u4f46\u8fd9\u4e5f\u4f1a\u8ba9\u6a21\u578b\u66f4\u96be\u5904\u7406\uff0c\u56e0\u4e3a\u5b83\u5fc5\u987b\u4e00\u6b21\u6027\u51b3\u5b9a\u6240\u6709\u5c42\u7ea7\u3002\u4eba\u4eec\u8fd8\u5c1d\u8bd5\u4e86\u8bb8\u591a\u5176\u4ed6\u65b9\u6848\u6765\u5e73\u8861\u538b\u7f29\u4e0e\u8d28\u91cf\u3002\u4ee5\u4e0b\u662f MusicGen \u4e2d\u5c1d\u8bd5\u8fc7\u7684\u51e0\u79cd\u65b9\u6848\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img src=\"https:\/\/kyutai.org\/assets\/codec-explainer\/image%204.png\" alt=\"\"\/><\/figure>\n\n\n\n<p>\u6709\u8da3\u7684\u662f\uff0c\u622a\u81f3 2025 \u5e74\uff0c\u8fd8\u6ca1\u6709\u4e00\u4e2a\u201c\u80dc\u51fa\u201d\u7684\u7edf\u4e00\u89e3\u51b3\u65b9\u6848\uff1a\u6bcf\u7bc7\u8bba\u6587\u7684\u505a\u6cd5\u90fd\u4e0d\u540c\uff0c\u800c\u4e14\u8fd9\u4e9b\u65b9\u6848\u53ef\u80fd\u53d8\u5f97\u76f8\u5f53\u590d\u6742\u3002\u770b\u770b\u8fd9\u4e2a\u6765\u81ea <strong>MiMo-Audio<\/strong> \u7684\u56fe\u8868\u5c31\u77e5\u9053\u4e86\uff0c\u8fd9\u662f\u4e00\u4e2a\u5728 2025 \u5e74 9 \u6708\u53d1\u5e03\u7684\u6a21\u578b\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img src=\"https:\/\/kyutai.org\/assets\/codec-explainer\/image%205.png\" alt=\"\"\/><figcaption>\u5904\u7406 <strong>\u591a\u4e2a RVQ \u7ea7\u522b<\/strong> \u7684\u65b9\u6cd5\u53ef\u80fd\u4f1a\u76f8\u5f53\u590d\u6742\u3002<\/figcaption><\/figure>\n\n\n\n<h3>Finally, let&#8217;s train<\/h3>\n\n\n\n<p>\u7ec8\u4e8e\u5230\u4e86\u8bad\u7ec3\u4e00\u4e2a\u5c01\u88c5\u4e86 codec \u7684\u8bed\u8a00\u6a21\u578b\u7684\u65f6\u5019\u4e86\uff01\u6b63\u5982\u6211\u6240\u63d0\u5230\u7684\uff0c\u6211\u4eec\u7684\u4ee3\u7801\u57fa\u4e8e Andrej Karpathy \u7528\u4e8e\u8bad\u7ec3\u6587\u672c LLM \u7684 nanoGPT \u4ee3\u7801\u5e93\u3002\u6211\u4eec\u53ea\u9700\u8981\u4fee\u6539\u5b83\u4ee5\u63a5\u53d7\u97f3\u9891\u4f5c\u4e3a\u8f93\u5165\u3002\u4f46\u8fd9\u5f88\u7b80\u5355\uff0c\u56e0\u4e3a LLM \u5e76\u4e0d\u5173\u5fc3\u4f60\u8f93\u5165\u7684\u662f\u54ea\u79cd token\u2014\u2014\u5bf9\u5b83\u6765\u8bf4\u90fd\u53ea\u662f\u6570\u5b57\u800c\u5df2\u3002\u4e00\u65e6\u6211\u4eec\u5c06\u6570\u636e\u96c6\u6807\u8bb0\u5316\uff08tokenized\uff09\u5e76\u5c06\u5176\u5c55\u5e73\uff08flattened\uff09\u4e3a\u4e00\u7ef4\u5e8f\u5217\uff0c\u6211\u4eec\u5c31\u53ef\u4ee5\u5f00\u59cb\u4e86\u3002\u4ee5\u8fd9\u79cd\u65b9\u5f0f\u6807\u8bb0\u5316\u540e\uff0c\u6211\u4eec 10000 \u5c0f\u65f6\u7684\u97f3\u9891\u5360\u7528\u4e86 134 GB \u7684\u7a7a\u95f4\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u5c06\u8fd9\u4e48\u591a\u6570\u636e\u5b58\u50a8\u4e3a\u672a\u538b\u7f29\u7684\u97f3\u9891\u5c06\u9700\u8981\u8d85\u8fc7 1 TB\u3002<\/p>\n\n\n\n<p>\u6211\u4eec\u5c06\u4f7f\u7528\u4e0e\u9010\u6837\u672c\uff08sample-by-sample\uff09\u6a21\u578b\u5b8c\u5168\u76f8\u540c\u7684\u6a21\u578b\u67b6\u6784\u548c\u8d85\u53c2\u6570\uff08hyperparameters\uff09\uff1a\u552f\u4e00\u7684\u533a\u522b\u5728\u4e8e\u6807\u8bb0\u5316\uff08tokenization\uff09\u65b9\u5f0f\u3002\u6211\u4eec\u7684\u6570\u636e\u96c6\u4e5f\u5927\u4e86 10 \u500d\uff0c\u4f46\u9010\u6837\u672c\u6a21\u578b\u751a\u81f3\u8fde 1000 \u5c0f\u65f6\u7684\u6570\u636e\u96c6\u90fd\u65e0\u6cd5\u5bb9\u7eb3\uff0c\u6240\u4ee5\u66f4\u591a\u7684\u6570\u636e\u4e5f\u6551\u4e0d\u4e86\u5b83\u3002<\/p>\n\n\n\n<p>\u6211\u7528 8 \u4e2a H100 \u663e\u5361\u8bad\u7ec3\u4e86\u8fd9\u4e2a\u6a21\u578b\u5927\u7ea6 5 \u5929\u3002\u4e3a\u4e86\u5f97\u5230\u4e00\u4e9b\u6837\u672c\uff0c\u6211\u51b3\u5b9a\u7528 Michael Field \u7684\u8bd7\u300a\u4e03\u6708\u300b\u4e2d\u7684\u4e24\u884c Libri-Light \u6717\u8bfb\u6837\u672c\u6765\u63d0\u793a\uff08prompt\uff09\u6a21\u578b\u3002\uff08\u5728\u505a\u8fd9\u4e2a\u9879\u76ee\u65f6\u6211\u4e86\u89e3\u5230\uff0cMichael Field \u662f Katherine Harris \u548c Edith Emma Cooper \u7684\u7b14\u540d\u3002\uff09\u8ba9\u6211\u4eec\u770b\u770b\u80fd\u4ece\u6211\u4eec\u7684\u6a21\u578b\u4e2d\u5f97\u5230\u4ec0\u4e48\u6837\u7684\u8bd7\u6b4c\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-audio\"><audio controls src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/02\/39c2073f-6417-4826-93ab-b525b0e55f49.wav\"><\/audio><\/figure>\n\n\n\n<p>\u53ef\u4ee5\u770b\u5230\u4e00\u4e9b\u201c\u751f\u547d\u7684\u8ff9\u8c61\u201d\uff0c\u4f46\u6211\u4eec\u8fd8\u6ca1\u6709\u4e00\u4e2a\u771f\u6b63\u7684\u201c\u8bd7\u4eba\u201d\u3002\u542c\u8d77\u6765\u5c31\u50cf\u6709\u4eba\u5728 <strong>\u5e18\u5e55\u540e\u9762\u8bf4\u8bdd<\/strong>\uff1a\u4f60\u65e0\u6cd5\u5b8c\u5168\u542c\u6e05\u5b83\u5728\u8bf4\u4ec0\u4e48\uff0c\u4f46 <strong>\u8bed\u8c03\u662f\u5b58\u5728\u7684<\/strong>\u2014\u2014\u542c\u8d77\u6765\u50cf\u6709\u4eba\u5728\u6717\u8bfb\u4e66\u672c\uff0c\u800c\u8fd9\u6b63\u662f\u6a21\u578b\u8bad\u7ec3\u65f6\u7684\u5185\u5bb9\u3002<\/p>\n\n\n\n<p>\u5b83\u8fd8\u80fd\u4fdd\u6301 <strong>\u8fde\u8d2f\u7684\u58f0\u97f3<\/strong>\uff0c\u76f4\u5230\u6700\u540e\u51e0\u79d2\u624d\u5207\u6362\u5230\u53e6\u4e00\u4e2a\u58f0\u97f3\u3002\u8fd9\u4e5f\u4e0e\u8bad\u7ec3\u6570\u636e\u4e00\u81f4\uff1a\u6211\u4eec\u4ece\u6240\u6709\u6709\u58f0\u4e66\u4e2d <strong>\u526a\u5207\u7247\u6bb5\u5e76\u6df7\u5408\u5728\u4e00\u8d77<\/strong> \u6765\u91c7\u6837\u8bad\u7ec3\u6570\u636e\uff0c\u6240\u4ee5\u6a21\u578b\u786e\u5b9e\u4f1a\u9047\u5230 <strong>\u4e0d\u540c\u8bf4\u8bdd\u4eba\u4e4b\u95f4\u7684\u754c\u9650<\/strong>\u3002<\/p>\n\n\n\n<h3><strong>\u4e00\u4e2a codec \u80fd\u5e26\u6211\u4eec\u8d70\u591a\u8fdc\uff1f<\/strong><\/h3>\n\n\n\n<p>\u6211\u4eec\u7684 codec \u662f\u6709\u610f\u8bbe\u8ba1\u5f97\u975e\u5e38\u7b80\u5355\u7684\uff0c\u8fd9\u4e5f\u89e3\u91ca\u4e86\u4e3a\u4ec0\u4e48\u7ed3\u679c\u4e0d\u5c3d\u5982\u4eba\u610f\u2014\u2014\u4f46\u5728\u8fc7\u53bb\u56db\u5e74\u91cc\uff0c\u5173\u4e8e\u795e\u7ecf\u97f3\u9891\u7f16\u89e3\u7801\u5668\u7684\u7814\u7a76\u5df2\u7ecf\u76f8\u5f53\u4e30\u5bcc\uff0c\u6211\u4eec\u53ef\u4ee5\u52a0\u4ee5\u5229\u7528\u3002\u6211\u4eec\u4e0d\u4f1a\u5728\u8fd9\u91cc\u5b9e\u73b0\u6240\u6709\u7684\u6539\u8fdb\uff0c\u800c\u662f\u770b\u770b\u5f53\u6211\u4eec\u4f7f\u7528 Mimi \u4f5c\u4e3a\u5206\u8bcd\u5668\uff08tokenizer\uff09\u65f6\u4f1a\u53d1\u751f\u4ec0\u4e48\u3002<\/p>\n\n\n\n<p>Mimi \u662f Kyutai \u4e3a\u6211\u4eec\u7684\u97f3\u9891\u8bed\u8a00\u6a21\u578b Moshi \u6784\u5efa\u7684\u4e00\u6b3e\u73b0\u4ee3\u795e\u7ecf\u97f3\u9891\u7f16\u89e3\u7801\u5668\u3002\u6b64\u540e\uff0c\u5b83\u4e5f\u88ab\u7528\u4f5c\u5176\u4ed6\u6a21\u578b\u7684\u5206\u8bcd\u5668\uff0c\u5982 Sesame CSM\u3001VoXtream \u548c LFM2-Audio\u3002<\/p>\n\n\n\n<p>\u4e0d\u51fa\u6240\u6599\uff0cMimi \u542c\u8d77\u6765\u6bd4\u6211\u4eec\u4e4b\u524d\u8bad\u7ec3\u7684\u81ea\u5236 codec \u597d\u5f97\u591a\u3002<\/p>\n\n\n\n<p>Mimi \u6ca1\u6709\u4f7f\u7528\u591a\u5c3a\u5ea6\u9891\u8c31\u635f\u5931\uff08multi-scale spectral loss\uff09\uff0c\u800c\u662f\u4f7f\u7528\u4e86\u50cf GAN \u4e00\u6837\u7684\u5bf9\u6297\u6027\u635f\u5931\uff08adversarial loss\uff09\u3002\u6709\u4e00\u4e2a\u5224\u522b\u5668\u7f51\u7edc\uff08discriminator network\uff09\u8bd5\u56fe\u5c06\u97f3\u9891\u5206\u7c7b\u4e3a\u539f\u59cb\u7684\u6216\u7531 codec \u91cd\u5efa\u7684\uff0c\u800c codec \u7684\u76ee\u6807\u5c31\u662f\u9a97\u8fc7\u8fd9\u4e2a\u5224\u522b\u5668\u3002<\/p>\n\n\n\n<p>Mimi \u589e\u52a0\u7684\u53e6\u4e00\u4e2a\u6539\u8fdb\u662f\u4f7f\u7528\u00a0<strong>RVQ dropout<\/strong>\uff1a\u5b83\u4f7f\u7528 32 \u4e2a RVQ \u5c42\u7ea7\uff0c\u4f46\u5728\u8bad\u7ec3\u671f\u95f4\uff0c\u91cd\u5efa\u6709\u65f6\u4f1a\u968f\u673a\u622a\u65ad\u5230\u8f83\u5c11\u7684\u5c42\u7ea7\u6570\u3002\u8fd9\u4f7f\u5f97\u6211\u4eec\u53ef\u4ee5\u5728\u63a8\u7406\u65f6\u4ee5\u8f83\u5c11\u7684 RVQ \u5c42\u7ea7\u8fd0\u884c Mimi\uff0c\u5e76\u4e14\u4ecd\u7136\u83b7\u5f97\u4e0d\u9519\u7684\u7ed3\u679c\uff0c\u56e0\u4e3a\u5b83\u4e0d\u4f9d\u8d56\u4e8e\u6240\u6709\u5c42\u7ea7\u7684\u5b58\u5728\u3002<strong>\u800c\u5bf9\u4e8e\u6211\u4eec\u7684\u81ea\u5236 codec\uff0c\u6211\u4eec\u5fc5\u987b\u5206\u5f00\u8bad\u7ec3\u3002<\/strong><\/p>\n\n\n\n<p>\u8ba9\u6211\u4eec\u542c\u542c\u7528 Mimi \u91cd\u5efa\u7684\u793a\u4f8b\u97f3\u9891\uff1a<\/p>\n\n\n\n<p>Original:<\/p>\n\n\n\n<figure class=\"wp-block-audio\"><audio controls src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/02\/7b51bf9a-7903-4411-a72a-45f0fa7b3202.wav\"><\/audio><\/figure>\n\n\n\n<p>\u91cd\u5efa\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-audio\"><audio controls src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/02\/5bcf643e-4cca-493e-960a-a5d59dfbc0ca.wav\"><\/audio><figcaption>4 RVQ levels<\/figcaption><\/figure>\n\n\n\n<figure class=\"wp-block-audio\"><audio controls src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/02\/22805c4b-31d9-4e7a-aadb-e79265b3aaf0.wav\"><\/audio><figcaption>16  RVQ levels <\/figcaption><\/figure>\n\n\n\n<figure class=\"wp-block-audio\"><audio controls src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/02\/8f993035-3d42-4bc6-8b88-7e22508525d4.wav\"><\/audio><figcaption> 32 RVQ levels  <\/figcaption><\/figure>\n\n\n\n<p>\u5c31\u6211\u4eec\u7684\u76ee\u7684\u800c\u8a00\uff0c\u5c42\u7ea7\u8f83\u5c11\u7684\u53d8\u4f53\u53ef\u80fd\u66f4\u5bb9\u6613\u5efa\u6a21\uff0c\u56e0\u4e3a\u5b83\u538b\u7f29\u7a0b\u5ea6\u66f4\u9ad8\u3002\u8ba9\u6211\u4eec\u7528 8 \u5c42\u548c 32 \u5c42\u7684 Mimi \u6765\u8bad\u7ec3\u6a21\u578b\uff0c\u5e76\u6bd4\u8f83\u7ed3\u679c\u3002<\/p>\n\n\n\n<p>\u6211\u8bad\u7ec3\u4e86\u548c\u4e4b\u524d\u5b8c\u5168\u76f8\u540c\u7684\u6a21\u578b\u67b6\u6784\uff0c\u552f\u4e00\u6539\u53d8\u7684\u662f\u5206\u8bcd\u5668\u3002\u6570\u636e\u96c6\u4ecd\u7136\u662f\u6765\u81ea Libri-Light \u7684 10000 \u5c0f\u65f6\u97f3\u9891\uff0c\u5c31\u50cf\u6211\u4eec\u4f7f\u7528\u7b80\u5355 codec \u65f6\u4e00\u6837\u3002Mimi \u7684\u91c7\u6837\u7387\u662f 24 kHz\uff0c\u4f46 Libri-Light \u4f7f\u7528\u7684\u662f 16 kHz\uff0c\u8fd9\u9650\u5236\u4e86\u58f0\u97f3\u7684\u6700\u9ad8\u54c1\u8d28\uff0c\u56e0\u4e3a\u6211\u4eec\u4e22\u5931\u4e86\u97f3\u9891\u7684\u66f4\u9ad8\u9891\u7387\u90e8\u5206\u3002<\/p>\n\n\n\n<p>Mimi \u5bf9\u97f3\u9891\u7684\u964d\u91c7\u6837\uff08downsample\uff09\u4e5f\u66f4\u6fc0\u8fdb\uff1a\u5b83\u7684\u5e27\u7387\u662f\u6bcf\u79d2 12.5 \u5e27\uff0c\u800c\u6211\u4eec\u7684 codec \u662f\u6bcf\u79d2 125 \u5e27\u2014\u2014\u9ad8\u4e86 10 \u500d\uff01\u8fd9\u610f\u5473\u7740\u6570\u636e\u96c6\u5728\u78c1\u76d8\u4e0a\u7684\u4f53\u79ef\u4e5f\u66f4\u5c0f\u3002\u7528\u6211\u4eec\u7684 codec\uff0c\u5b83\u5360\u4e86 134 GB\uff0c\u4f46\u7528 Mimi\uff0c\u201c\u4ec5\u4ec5\u201d\u662f 54 GB\u3002<\/p>\n\n\n\n<p>\u8fd9\u662f\u4e00\u9996\u7528\u5728 Mimi \u6807\u8bb0\u5316\u6570\u636e\u4e0a\u8bad\u7ec3\u7684\u6a21\u578b\u751f\u6210\u7684\u8bd7\u3002\u6211\u548c\u4e4b\u524d\u4e00\u6837\uff0c\u7528\u8bd7\u4e2d\u7684\u4e24\u884c\u6765\u63d0\u793a\u5b83\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-audio\"><audio controls src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/02\/876fd200-a098-4b5c-b6ca-91129818666f.wav\"><\/audio><\/figure>\n\n\n\n<p>\u8fd9\u662f\u6211\u5c3d\u529b\u5c1d\u8bd5\u7684\u8f6c\u5f55\uff1a<\/p>\n\n\n\n<blockquote class=\"wp-block-quote\"><p><em>When the grass is gone<br>And corn still grassy;<\/em><br>Illness worried in the fur<br>this and pelan in stones<br>during the turan\u2019s ciscerey<br>headforths nepet Paul Twain.<br>He&nbsp;<em>sees<\/em>&nbsp;zin in them.<\/p><\/blockquote>\n\n\n\n<p>\u5bf9\u6211\u6765\u8bf4\u6709\u70b9\u592a\u8d85\u73b0\u5b9e\u4e3b\u4e49\u4e86\uff0c\u4f46\u4e5f\u8bb8\u5218\u6613\u65af\u00b7\u5361\u7f57\u5c14\u4f1a\u559c\u6b22\u3002<\/p>\n\n\n\n<h3><strong>\u8bed\u4e49 token (Semantic tokens)<\/strong><\/h3>\n\n\n\n<p>\u6211\u5f97\u5766\u767d\u4e00\u4ef6\u4e8b\uff1a\u6211\u521a\u624d\u5bf9\u4f60\u6492\u8c0e\u4e86\u3002\u4f46\u53ea\u662f\u4e00\u70b9\u70b9\uff0c\u800c\u4e14\u662f\u4e3a\u4e86\u6559\u5b66\u76ee\u7684\u3002\u5b9e\u9645\u4e0a\uff0c\u4e0a\u9762\u7684\u6a21\u578b\u662f\u5728\u4e00\u4e2a 31 \u5c42\u7684 Mimi \u97f3\u9891\u4e0a\u8bad\u7ec3\u7684\uff0c\u6211\u7701\u7565\u4e86\u7b2c\u4e00\u5c42\uff0c\u4e5f\u5c31\u662f\u5305\u542b \u201c<strong>semantic token<\/strong>\u201d \u7684\u90a3\u4e00\u5c42\u3002<\/p>\n\n\n\n<p>\u8fd9\u4e2a token \u7684\u4f5c\u7528\u662f\u8868\u793a\u97f3\u9891\u7684\u8bed\u4e49\u4fe1\u606f\uff0c\u800c\u4e0d\u4e00\u5b9a\u6709\u52a9\u4e8e\u91cd\u5efa\u3002\u6211\u4e0d\u4f1a\u6df1\u5165\u63a2\u8ba8\u5b83\u4eec\u7684\u5de5\u4f5c\u539f\u7406\uff0c\u4f46\u7b80\u5355\u6765\u8bf4\uff0cMimi \u7684 semantic tokens \u662f\u4ece WavLM \u4e2d\u63d0\u70bc\u51fa\u6765\u7684\uff0c\u4f60\u53ef\u4ee5\u628a\u5b83\u770b\u4f5c\u662f\u8bed\u97f3\u9886\u57df\u7684 BERT\u3002<\/p>\n\n\n\n<p>\u4e3a\u4e86\u611f\u53d7 semantic tokens \u7f16\u7801\u4e86\u4ec0\u4e48\u4fe1\u606f\uff0c\u8ba9\u6211\u4eec\u4ee5\u8fd9\u4e2a\u793a\u4f8b\u97f3\u9891\u4e3a\u4f8b\uff0c\u5c06\u5176\u901a\u8fc7 Mimi \u5904\u7406\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-audio\"><audio controls src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/02\/c0d7c134-bc79-4220-a309-55738bdd6d34.wav\"><\/audio><\/figure>\n\n\n\n<p>\u73b0\u5728\uff0c\u8ba9\u6211\u4eec\u8bad\u7ec3\u4e00\u4e2a\u57fa\u4e8e\u5b8c\u6574 Mimi\uff08\u5305\u62ec semantic tokens\uff09\u7684\u8bed\u8a00\u6a21\u578b\u3002\u6211\u4eec\u5c06\u4ee5\u4e00\u79cd\u7279\u6b8a\u7684\u65b9\u5f0f\u8fd0\u884c\u6a21\u578b\uff1a\u4fdd\u7559\u539f\u59cb\u97f3\u9891\u7684 semantic tokens\uff0c\u4f46\u4e22\u5f03\u5176\u4ed6\u6240\u6709 token\uff0c\u7136\u540e\u8ba9\u6a21\u578b\u6765\u9884\u6d4b\u5b83\u4eec\u3002\u8fd9\u610f\u5473\u7740\u6765\u81ea semantic tokens \u7684\u4fe1\u606f\u662f\u56fa\u5b9a\u7684\uff08\u201cteacher-forced\u201d\uff09\uff0c\u4f46\u6a21\u578b\u53ef\u4ee5\u6839\u636e\u5b83\u8ba4\u4e3a\u5408\u7406\u7684\u5ef6\u7eed\u81ea\u7531\u51b3\u5b9a\u5176\u4ed6 token\u3002<\/p>\n\n\n\n<figure class=\"wp-block-video\"><video controls src=\"https:\/\/kyutai.org\/assets\/codec-explainer\/regenerateWithSemantic.mp4\"><\/video><figcaption>\u901a\u8fc7\u56fa\u5b9a semantic tokens \u5e76\u8ba9\u6a21\u578b\u91cd\u65b0\u751f\u6210\u5176\u4f59\u90e8\u5206\uff0c\u6211\u4eec\u53ef\u4ee5\u4e86\u89e3 semantic tokens \u4e2d\u5305\u542b\u4e86\u54ea\u4e9b\u4fe1\u606f\u3002<\/figcaption><\/figure>\n\n\n\n<p>\u542c\u542c\u6211\u4eec\u7528\u8fd9\u79cd\u65b9\u5f0f\u5f97\u5230\u7684\u4e24\u4e2a\u4e0d\u540c\u7684\u91cd\u5efa\u7248\u672c\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-audio\"><audio controls src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/02\/ecc390a3-a9fb-4d75-a7c5-ea6bf9a6aaa5.wav\"><\/audio><\/figure>\n\n\n\n<figure class=\"wp-block-audio\"><audio controls src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/02\/9b0fe838-d373-4263-91f2-eda66d46a20a.wav\"><\/audio><\/figure>\n\n\n\n<p>\u58f0\u97f3\u5b8c\u5168\u4e0d\u540c\uff0c\u4f46\u8bf4\u7684\u5185\u5bb9\u662f\u4e00\u6837\u7684\uff01\u8fd9\u610f\u5473\u7740 semantic tokens \u7f16\u7801\u4e86\u8bf4\u8bdd\u8005\u5728\u8bf4\u4ec0\u4e48\uff0c\u4f46\u4e0e\u55d3\u97f3\u65e0\u5173\u3002\u8fd9\u5f88\u6709\u7528\uff0c\u56e0\u4e3a\u5b83\u5e2e\u52a9\u6a21\u578b\u4e13\u6ce8\u4e8e\u00a0<strong>\u8bf4\u4ec0\u4e48<\/strong>\uff0c\u800c\u4e0d\u662f\u00a0<strong>\u600e\u4e48\u8bf4<\/strong>\u3002\u5728\u8fd9\u65b9\u9762\uff0c\u5b83\u4eec\u66f4\u63a5\u8fd1\u4e8e\u6587\u672c token\uff0c\u56e0\u4e3a\u6587\u672c token \u4e5f\u4e0d\u5305\u542b\u5173\u4e8e\u55d3\u97f3\u3001\u8bed\u8c03\u3001\u65f6\u95f4\u6216\u60c5\u611f\u7684\u4fe1\u606f\u3002<\/p>\n\n\n\n<h3><strong>\u8ba9\u8bd7\u6b4c\u66f4\u5177\u8bed\u4e49<\/strong><\/h3>\n\n\n\n<p>\u73b0\u5728\uff0c\u8ba9\u6211\u4eec\u7528\u5728\u5305\u542b\u8bed\u4e49\u7684 Mimi \u4e0a\u8bad\u7ec3\u7684\u6a21\u578b\u6765\u5b8c\u6210\u8fd9\u9996\u8bd7\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-audio\"><audio controls src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/02\/9a1b3a3e-e69c-4c65-b0a5-5cea10882fa3.wav\"><\/audio><\/figure>\n\n\n\n<p><em>When grass is gone<br>and corn still grassy;<\/em><br><em>from the man was nothing moan.<br>The low death and heart<br>She came\u00a0fyde\u00a0wood.<br>A finteriest, a fall,<br>all them<\/em>.<\/p>\n\n\n\n<p>\u5b83\u4ecd\u7136\u4f1a\u7f16\u9020\u8bcd\u6c47\uff0c\u53e5\u5b50\u4e5f\u4e0d\u592a\u8fde\u8d2f\uff0c\u4f46\u5f88\u660e\u663e\uff0c\u771f\u5b9e\u5355\u8bcd\u7684\u6bd4\u4f8b\u9ad8\u4e86\u5f88\u591a\uff1b\u6a21\u578b\u53d8\u5f97\u201c\u66f4\u5177\u8bed\u4e49\u201d\u4e86\u3002\u58f0\u97f3\u8d28\u91cf\u548c\u4e4b\u524d\u4e00\u6837\uff0c\u8fd9\u4e5f\u7b26\u5408\u6211\u4eec\u7684\u9884\u671f\u3002<\/p>\n\n\n\n<p>\u8ba9\u6211\u4eec\u542c\u7b2c\u4e8c\u9996\u8bd7\uff1a<br><\/p>\n\n\n\n<figure class=\"wp-block-audio\"><audio controls src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/02\/a6d472b2-a9f3-448e-92e2-8b8d7b5d3db9.wav\"><\/audio><\/figure>\n\n\n\n<blockquote class=\"wp-block-quote\"><p><em>When grass is gone<br>and corn still grassy;<\/em><br>hope won and she<br>who is just a night in Tatan<br>in doe ock-ohm?<br>the whom?<\/p><\/blockquote>\n\n\n\n<p>\u786e\u5b9e\uff0cthe whom?<\/p>\n\n\n\n<h3><strong>\u8bed\u4e49\u4e0e\u58f0\u5b66\u7684\u6743\u8861 (Semantic\u2013acoustic tradeoff)<\/strong><\/h3>\n\n\n\n<p>\u6211\u4eec\u53ef\u4ee5 <strong>\u727a\u7272\u4e00\u4e9b\u58f0\u5b66\u8d28\u91cf<\/strong> \u6765\u63d0\u5347\u8bed\u4e49\u6548\u679c\uff0c\u901a\u8fc7 <strong>\u51cf\u5c11 RVQ \u7ea7\u522b\u7684\u6570\u91cf<\/strong>\u3002\u6211\u4eec\u9009\u62e9 8 \u7ea7\u3002\u8fd9\u6837\u4e00\u6765\uff0c\u6211\u4eec\u83b7\u5f97\u4e86 <strong>\u66f4\u9ad8\u7684\u97f3\u9891\u538b\u7f29\u7387<\/strong>\uff0c\u540c\u65f6\u635f\u5931\u4e2d <strong>\u8bed\u4e49 token \u5360\u6bd4\u4e5f\u76f8\u5e94\u63d0\u9ad8<\/strong>\uff0c\u56e0\u4e3a\u73b0\u5728\u662f 1\/8 \u7684 token\uff0c\u800c\u4e0d\u662f\u4e4b\u524d\u7684 1\/32\u3002<\/p>\n\n\n\n<p>\u6211\u5bf9\u8fd9\u4e2a\u6a21\u578b\u7684\u7b2c\u4e00\u5370\u8c61\u4e4b\u4e00\u662f\uff0c\u5b83\u5b66\u4f1a\u4e86 <strong>\u8bb0\u5fc6 Librivox \u7684\u7248\u6743\u58f0\u660e<\/strong>\uff0c\u6240\u4ee5\u6709\u65f6\u5b83\u4f1a\u751f\u6210\u7c7b\u4f3c\u8fd9\u6837\u7684\u5185\u5bb9\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-audio\"><audio controls src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/02\/644e7d75-3609-4f44-87cb-2321565985b2.wav\"><\/audio><\/figure>\n\n\n\n<p><\/p>\n\n\n\n<blockquote class=\"wp-block-quote\"><p>Chapter 6 of The Founday, by R. Auclair.<br>This is a Librivox recording. All Librivox recordings are in the public domain. For information, or to volunteer, please visit librivox.org.<br>Reading by: Kelvert<\/p><\/blockquote>\n\n\n\n<p>\u91cd\u590d\u8bad\u7ec3\u6570\u636e\u901a\u5e38\u4e0d\u662f\u4f60\u60f3\u8981\u7684\uff0c\u4f46\u5728\u6211\u4eec\u7684\u6848\u4f8b\u4e2d\uff0c\u8fd9\u662f\u4e00\u4e2a\u6781\u597d\u7684\u751f\u547d\u8ff9\u8c61\uff0c\u56e0\u4e3a\u4e4b\u524d\u7684\u6a21\u578b\u751a\u81f3\u8fde\u8fd9\u4e2a\u90fd\u505a\u4e0d\u5230\u3002\u5b83\u8fd8\u7f16\u9020\u4e86\u4e66\u540d\u3001\u4f5c\u8005\u548c\u6717\u8bfb\u8005\uff0c\u6240\u4ee5\u8fd9\u91cc\u4ecd\u7136\u6709\u521b\u65b0\u6027\u3002<\/p>\n\n\n\n<p>\u73b0\u5728\uff0c\u8ba9\u6211\u4eec\u5c1d\u8bd5\u521b\u4f5c\u66f4\u591a\u7684\u8bd7\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-audio\"><audio controls src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/02\/d78beed3-665a-4dcc-a1cf-011962f34f2d.wav\"><\/audio><\/figure>\n\n\n\n<blockquote class=\"wp-block-quote\"><p><em>When grass is gone<br>and corn still grassy;<\/em><br>When so we could say<br>that in fairy interesting wife<br>who lay there and gone<br>that save the rosy light of life<br>Jay Dien, the antique mollity<br>and a mollity the beast of gray failed summon<\/p><p>end of poem.<\/p><p>This recording is in the public domain.<\/p><p>[different voice]<br>So we have formed a float that sent in would rattle down. The piece of opportunity reading and assimila\u2014<\/p><p><\/p><\/blockquote>\n\n\n\n<p>\u8fd9\u592a\u68d2\u4e86\u3002\u6709\u51e0\u4e2a\u8ff9\u8c61\u8868\u660e\u8fd9\u4e2a\u6a21\u578b\u6bd4\u4e4b\u524d\u7684\u66f4\u597d\u3002\u6211\u559c\u6b22\u5b83\u7f16\u9020\u4e86\u201cmollity\u201d\u8fd9\u4e2a\u8bcd\uff0c\u7136\u540e\u5728\u4e0b\u4e00\u884c\u91cd\u590d\u5b83\u3002\u800c\u4e14\uff0c\u5b83\u610f\u8bc6\u5230\u81ea\u5df1\u6b63\u5728\u80cc\u8bf5\u4e00\u9996\u8bd7\uff0c\u5e76\u5728\u8be5\u90e8\u5206\u7ed3\u5c3e\u52a0\u4e0a\u4e86 \u201cend of poem\u201d\u3002\u7136\u540e\u5b83\u8ba4\u4e3a\u8fd9\u662f\u7ae0\u8282\/\u90e8\u5206\u7684\u7ed3\u5c3e\uff0c\u5e76\u4ee5\u201cThis recording is in the public domain.\u201d\u7684\u58f0\u660e\u7ed3\u675f\u3002\u4e4b\u540e\uff0c\u5b83\u6362\u4e86\u4e2a\u58f0\u97f3\u7ee7\u7eed\u8bf4\u8bdd\u3002\u8fd9\u662f\u5408\u7406\u7684\uff0c\u56e0\u4e3a\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u6765\u81ea\u4e0d\u540c\u6709\u58f0\u8bfb\u7269\u7684\u7247\u6bb5\u53ea\u662f\u88ab\u968f\u673a\u6253\u4e71\u5e76\u8fde\u63a5\u5728\u4e00\u8d77\uff0c\u6240\u4ee5\u5728\u8fd9\u91cc\u6a21\u578b\u6a21\u62df\u4e86\u4e00\u4e2a\u7247\u6bb5\u8fb9\u754c\u3002<\/p>\n\n\n\n<p>\u5982\u679c\u6211\u4eec\u7ed9 semantic tokens \u7684\u635f\u5931\u8d4b\u4e88\u6bd4\u58f0\u5b66 tokens \u66f4\u9ad8\u7684\u6743\u91cd\uff0c\u53ef\u80fd\u4f1a\u5f97\u5230\u66f4\u597d\u7684\u7ed3\u679c\uff0c\u8ba9\u6a21\u578b\u66f4\u5173\u6ce8\u610f\u4e49\u800c\u975e\u58f0\u97f3\u2014\u2014\u4e8b\u5b9e\u4e0a\uff0c<strong>Moshi \u4f7f\u7528\u4e86\u9ad8\u8fbe 100 \u500d\u7684 semantic loss<\/strong>\u00a0\uff01\u4f46\u6211\u4eec\u603b\u5f97\u6709\u4e2a\u7ec8\u70b9\u3002<\/p>\n\n\n\n<h3>Conclusion<\/h3>\n\n\n\n<p>\u6211\u4eec\u6210\u529f\u5730\u4f7f\u7528\u795e\u7ecf\u97f3\u9891\u7f16\u89e3\u7801\u5668\u5236\u4f5c\u4e86\u4e00\u4e2a\u80fd\u751f\u6210\u67d0\u79cd\u7a0b\u5ea6\u4e0a\u8fde\u8d2f\u8bed\u97f3\u7684\u97f3\u9891\u8bed\u8a00\u6a21\u578b\u3002\u663e\u7136\uff0c\u8fd9\u8fd8\u4e0d\u662f 2025 \u5e74\u7684\u9876\u5c16\u6c34\u5e73\uff08\u6211\u4eec\u5728\u8fd9\u91cc\u4e5f\u5e76\u975e\u8ffd\u6c42\u4e8e\u6b64\uff09\uff0c\u4f46\u8bf7\u8bb0\u4f4f\uff0c\u4f7f\u7528\u5b8c\u5168\u76f8\u540c\u7684\u6a21\u578b\uff0c\u82e5\u4e0d\u91c7\u7528\u795e\u7ecf\u97f3\u9891\u7f16\u89e3\u7801\u5668\uff0c\u6211\u4eec\u5f97\u5230\u7684\u662f\u7c7b\u4f3c\u4e8e\u5f00\u5934\u7684\u97f3\u9891\u3002<\/p>\n\n\n\n<p>\u5f53\u7136\uff0c\u8981\u8d76\u4e0a\u6587\u672c\u6a21\u578b\u8fd8\u6709\u5f88\u957f\u7684\u8def\u8981\u8d70\uff01\u76ee\u524d\uff0c\u8bed\u97f3\u7406\u89e3\u548c\u63a8\u7406\u80fd\u529b\u4e4b\u95f4\u4f3c\u4e4e\u5b58\u5728\u4e00\u79cd\u6743\u8861\u3002\u5728\u6587\u7ae0\u5f00\u5934\u6211\u63d0\u5230\uff0c\u90a3\u4e9b\u539f\u751f\u652f\u6301\u8bed\u97f3\u7684\u6a21\u578b\uff08Gemini\u3001ChatGPT \u7684\u9ad8\u7ea7\u8bed\u97f3\u6a21\u5f0f\u3001Qwen\u3001Moshi\uff09\u90fd\u65e0\u6cd5\u5224\u65ad\u4f60\u662f\u5728\u7528\u9ad8\u97f3\u8fd8\u662f\u4f4e\u97f3\u8bf4\u8bdd\uff0c\u5c3d\u7ba1\u5b83\u4eec\u88ab\u8bad\u7ec3\u6765\u539f\u751f\u7406\u89e3\u97f3\u9891\u3002\u8fd9\u53ef\u80fd\u662f\u56e0\u4e3a\u5b83\u4eec\u5728\u5927\u91cf\u4f7f\u7528\u6587\u672c\u5230\u8bed\u97f3\u6280\u672f\u5408\u6210\u7684\u6570\u636e\u4e0a\u8fdb\u884c\u8bad\u7ec3\uff0c\u6216\u56e0\u4e3a\u7406\u89e3\u58f0\u97f3\u7684\u97f3\u8c03\uff08\u663e\u7136\uff09\u5e76\u4e0d\u80fd\u5e2e\u52a9\u6a21\u578b\u505a\u51fa\u66f4\u51c6\u786e\u7684\u9884\u6d4b\u3002<\/p>\n\n\n\n<p>Kyutai \u66fe\u5c1d\u8bd5\u7528 Moshi\uff08demo\uff0c\u8bba\u6587\uff09\u521b\u5efa\u4e00\u4e2a\u57fa\u4e8e\u97f3\u9891\u8bed\u8a00\u6a21\u578b\u7684\u8bed\u97f3\u804a\u5929\u5e94\u7528\uff0c\u5e76\u4e8e 2024 \u5e74 7 \u6708\u53d1\u5e03\u3002Moshi \u53ef\u80fd\u4e0d\u662f\u4f60\u4f1a\u9009\u62e9\u5e2e\u4f60\u505a\u4f5c\u4e1a\u7684 AI\uff0c\u4f46\u8bf7\u5bf9\u5b83\u5bbd\u5bb9\u4e00\u4e9b\uff1a\u5b83\u662f\u7b2c\u4e00\u4e2a\u7aef\u5230\u7aef\u7684\u8bed\u97f3 AI\uff0c\u751a\u81f3\u6bd4 OpenAI \u7684\u9ad8\u7ea7\u8bed\u97f3\u6a21\u5f0f\u53d1\u5e03\u5f97\u8fd8\u8981\u65e9\u3002<\/p>\n\n\n\n<p>Moshi \u4e3a\u81ea\u5df1\u548c\u7528\u6237\u5e76\u884c\u5730\u6a21\u62df\u4e86\u4e00\u4e2a\u201c\u5185\u5fc3\u72ec\u767d\u201d\u7684\u6587\u672c\u6d41\u548c\u97f3\u9891\u6d41\u3002\u6587\u672c\u6d41\u5e2e\u52a9\u5b83\u89c4\u5212\u8981\u8bf4\u4ec0\u4e48\uff0c\u800c\u6d88\u878d\u7814\u7a76\uff08ablations\uff09\u8868\u660e\uff0c\u6587\u672c\u6d41\u5bf9\u6a21\u578b\u7684\u5e2e\u52a9\u5de8\u5927\u3002\u540c\u65f6\uff0c\u8fd9\u4e5f\u6709\u70b9\u53ef\u60b2\uff1a\u5927\u90e8\u5206\u7684\u63a8\u7406\u4f3c\u4e4e\u90fd\u88ab\u59d4\u6258\u7ed9\u4e86\u6587\u672c\u6d41\uff0c\u800c\u97f3\u9891\u6d41\u53ea\u662f\u7528\u6765\u63d0\u4f9b\u96c6\u6210\u7684\u8bed\u97f3\u5230\u6587\u672c\u548c\u6587\u672c\u5230\u8bed\u97f3\u529f\u80fd\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img src=\"https:\/\/kyutai.org\/assets\/codec-explainer\/moshi-figure-1.png\" alt=\"\"\/><figcaption><a rel=\"noreferrer noopener\" href=\"https:\/\/arxiv.org\/abs\/2410.00037\" target=\"_blank\">Moshi<\/a>\u00a0models two audio streams and a text stream in parallel<\/figcaption><\/figure>\n\n\n\n<p>\u8fd9\u4e0d\u4ec5\u4ec5\u662f Moshi \u7684\u95ee\u9898\uff1a\u6b63\u5982\u201c\u6211\u662f\u5728\u7528\u9ad8\u97f3\u8bf4\u8bdd\u5417\u201d\u7684\u5b9e\u9a8c\u6240\u793a\uff0c\u8fd9\u79cd\u5bf9\u6587\u672c\u800c\u975e\u97f3\u9891\u7684\u8fc7\u5ea6\u4f9d\u8d56\u662f\u6240\u6709\u97f3\u9891 LLM \u7684\u4e00\u4e2a\u95ee\u9898\u3002\u5c3d\u7ba1\u4e3b\u6d41\u7684\u5efa\u6a21\u65b9\u6cd5\u4e0e Moshi \u6709\u6240\u4e0d\u540c\uff1a\u5b83\u4eec\u662f\u4ea4\u9519\u5904\u7406\u6587\u672c\u548c\u97f3\u9891 token\uff0c\u800c\u4e0d\u662f\u5728\u5e76\u884c\u6d41\u4e2d\u5efa\u6a21\u3002<\/p>\n\n\n\n<p>\u5728 Moshi \u53d1\u5e03\u4e00\u5e74\u591a\u540e\uff0c\u97f3\u9891\u6a21\u578b\u4ecd\u7136\u843d\u540e\u4e8e\u6587\u672c LLM\u3002\u4f46\u4e3a\u4ec0\u4e48\u5462\uff1f\u5bf9\u6211\u6765\u8bf4\uff0c\u8fd9\u4e2a\u795e\u79d8\u4e14\u672a\u89e3\u7684\u201c\u6a21\u6001\u9e3f\u6c9f\u201d\uff08modality gap\uff09\u4f7f\u5f97\u97f3\u9891\u673a\u5668\u5b66\u4e60\u6210\u4e3a\u4e00\u4e2a\u4ee4\u4eba\u5174\u594b\u7684\u7814\u7a76\u9886\u57df\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u539f\u6587\u94fe\u63a5\uff1ahttps:\/\/kyutai.org\/codec-explainer \u622a\u81f3 2025 \u5e74 10 \u6708\uff0c &hellip; <a href=\"http:\/\/139.9.1.231\/index.php\/2026\/02\/11\/neural-audio-codecs-how-to-get-audio-into-llms\/\" class=\"more-link\">\u7ee7\u7eed\u9605\u8bfb<span class=\"screen-reader-text\">\u97f3\u9891 Tokenizer\u7684\u65b9\u6cd5\uff1aMoshi \u56e2\u961f\u5206\u4eab<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[4,9,38,43,34],"tags":[],"_links":{"self":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/30222"}],"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=30222"}],"version-history":[{"count":136,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/30222\/revisions"}],"predecessor-version":[{"id":30385,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/30222\/revisions\/30385"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=30222"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=30222"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=30222"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}