{"id":31718,"date":"2026-09-26T09:11:54","date_gmt":"2026-09-26T01:11:54","guid":{"rendered":"http:\/\/139.9.1.231\/?p=31718"},"modified":"2026-09-26T09:12:08","modified_gmt":"2026-09-26T01:12:08","slug":"qwen-audio-3-0-asr","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2026\/09\/26\/qwen-audio-3-0-asr\/","title":{"rendered":"Qwen-Audio-3.0-ASR\uff1a\u7528 MoE\u3001\u4e0a\u4e0b\u6587\u4e0e\u6307\u4ee4\u63a7\u5236\u628a ASR \u63a8\u5411\u771f\u5b9e\u5de5\u4e1a\u573a\u666f"},"content":{"rendered":"\n

\u8bba\u6587\uff1a<\/em><\/strong>Qwen-Audio-3.0-ASR Technical Report<\/em><\/a><\/p>\n\n\n\n

\u4e00\u3001\u5148\u8bf4\u7ed3\u8bba\uff1a\u8fd9\u4e0d\u662f\u53ea\u8ffd\u6c42 Benchmark \u7684 ASR<\/h2>\n\n\n\n

Qwen-Audio-3.0-ASR \u7684\u6838\u5fc3\u76ee\u6807\uff0c\u662f\u628a\u201c\u542c\u6e05\u695a\u201d\u6269\u5c55\u6210\u201c\u5728\u771f\u5b9e\u4ea7\u54c1\u91cc\u6301\u7eed\u53ef\u7528\u201d\u3002\u8bba\u6587\u628a\u95ee\u9898\u5f52\u7eb3\u4e3a\u56db\u7c7b\uff1a\u591a\u8bed\u8a00\u548c\u5730\u533a\u65b9\u8a00\u3001\u52a8\u6001\u4eba\u540d\/\u54c1\u724c\/\u4e13\u4e1a\u672f\u8bed\u3001\u957f\u8bed\u97f3\u8de8\u53e5\u4e0a\u4e0b\u6587\uff0c\u4ee5\u53ca\u505c\u987f\u3001\u91cd\u590d\u3001\u53e3\u8bef\u548c\u81ea\u6211\u7ea0\u6b63\u7b49\u81ea\u7136\u53e3\u8bed\u73b0\u8c61\u3002<\/p>\n\n\n\n

\u6a21\u578b\u91c7\u7528 Qwen Mixture-of-Experts\uff08MoE\uff09\u8bed\u8a00\u6a21\u578b\u4f5c\u4e3a\u89e3\u7801\u5668\uff0c\u8bad\u7ec3\u6570\u636e\u8fbe\u5230\u6570\u5343\u4e07\u5c0f\u65f6\u91cf\u7ea7\uff0c\u7edf\u4e00\u652f\u6301 30 \u79cd\u8bed\u8a00\u548c 16 \u79cd\u4e2d\u6587\u65b9\u8a00\u53d8\u4f53\u3002\u66f4\u91cd\u8981\u7684\u662f\uff0c\u5b83\u628a\u8bed\u8a00\u6307\u5b9a\u3001\u5386\u53f2\u6587\u672c\u3001P0\/P1 \u5206\u5c42\u70ed\u8bcd\u548c\u6587\u672c\u6da6\u8272\u90fd\u8bbe\u8ba1\u6210\u8f93\u5165\u6307\u4ee4\uff0c\u800c\u4e0d\u662f\u4e3a\u6bcf\u79cd\u4efb\u52a1\u7ef4\u62a4\u4e00\u5957\u72ec\u7acb\u6a21\u578b\u3002<\/p>\n\n\n\n

  • 30 \u79cd\u8bed\u8a00\uff0c\u8986\u76d6\u4e1c\u4e9a\u3001\u4e1c\u5357\u4e9a\u3001\u5357\u4e9a\u3001\u963f\u62c9\u4f2f\u8bed\u533a\u548c\u6b27\u6d32\u591a\u79cd\u8bed\u8a00\u3002<\/li>
  • 16 \u79cd\u4e2d\u6587\u65b9\u8a00\/\u65b9\u8a00\u53d8\u4f53\uff0c\u8986\u76d6 8 \u4e2a\u4e3b\u8981\u65b9\u8a00\u533a\uff0c\u65e2\u652f\u6301\u65b9\u8a00\u8f6c\u5199\uff0c\u4e5f\u652f\u6301\u65b9\u8a00\u5230\u666e\u901a\u8bdd\u7684\u8bed\u97f3\u7ffb\u8bd1\u3002<\/li>
  • \u884c\u4e1a\u5b9e\u4f53\u8bc6\u522b\u3001\u5206\u5c42\u70ed\u8bcd\u3001\u957f\u97f3\u9891\u4e0a\u4e0b\u6587\u548c\u539f\u751f\u5355\u6b21\u89e3\u7801\u6da6\u8272\u3002<\/li>
  • \u72ec\u7acb\u7684 Qwen-Audio-3.0-ASR-Streaming \u6d41\u5f0f\u7248\u672c\uff0c\u4ee5\u53ca\u9762\u5411\u8bed\u97f3\u8f93\u5165\u3001\u4f1a\u8bae\u3001\u91c7\u8bbf\u548c\u5ba2\u670d\u7684 Message ASR \u90e8\u7f72\u5f62\u6001\u3002<\/li><\/ul>\n\n\n\n
    \"\"
    \u7cfb\u7edf\u603b\u89c8\uff1a\u591a\u8bed\u8a00\/\u65b9\u8a00\u3001\u6d41\u5f0f\u3001\u884c\u4e1a\u5b9e\u4f53\u3001\u957f\u4e0a\u4e0b\u6587\u3001\u70ed\u8bcd\u548c\u539f\u751f\u6da6\u8272\uff0c\u4ee5\u53ca\u201c\u6307\u4ee4\u63a7\u5236\u3001\u4e0a\u4e0b\u6587\u611f\u77e5\u3001\u5355\u6b21\u751f\u6210\u3001\u4f4e\u5ef6\u8fdf\u201d\u56db\u4e2a\u7cfb\u7edf\u7ea7\u5c5e\u6027\u3002<\/figcaption><\/figure>\n\n\n\n

    \u4e8c\u3001\u7814\u7a76\u80cc\u666f\uff1a\u6570\u636e\u3001\u6a21\u578b\u548c LLM \u4e09\u6761\u8def\u7ebf\u6b63\u5728\u6c47\u5408<\/h2>\n\n\n\n

    \u8bba\u6587\u628a\u8fd1\u5e74\u7684 ASR \u6f14\u8fdb\u6982\u62ec\u4e3a\u4e09\u4e2a\u4e92\u8865\u65b9\u5411\uff1a<\/p>\n\n\n\n

    • Data Scaling\uff1a<\/strong>Whisper \u7cfb\u5217\u4ece\u7ea6 3,000 \u5c0f\u65f6\u6269\u5927\u5230 68 \u4e07\u5c0f\u65f6\u4ee5\u4e0a\u540e\uff0cWER \u4e0b\u964d\u8d85\u8fc7 20 \u4e2a\u70b9\uff1b\u9886\u5148\u7cfb\u7edf\u7684\u8bad\u7ec3\u89c4\u6a21\u5df2\u7ecf\u8fdb\u5165\u6570\u5343\u4e07\u5c0f\u65f6\u3002<\/li>
    • Model Scaling\uff1a<\/strong>Whisper \u53c2\u6570\u91cf\u4ece\u7ea6 38M \u589e\u52a0\u5230 1.5B\uff0c\u591a\u8bed\u8a00 WER \u7ee7\u7eed\u663e\u8457\u6539\u5584\u3002<\/li>
    • LLM Integration\uff1a<\/strong>Seed-ASR\u3001FireRedAudio \u7b49\u6a21\u578b\u5229\u7528\u8bed\u8a00\u77e5\u8bc6\u3001\u4e16\u754c\u77e5\u8bc6\u548c\u4e0a\u4e0b\u6587\u63a8\u7406\u80fd\u529b\uff0c\u5904\u7406\u540c\u97f3\u8bcd\u3001\u5b9e\u4f53\u6b67\u4e49\u548c\u8bed\u4e49\u4e0d\u5b8c\u6574\u7b49\u7eaf\u58f0\u5b66\u6a21\u578b\u96be\u4ee5\u89e3\u51b3\u7684\u95ee\u9898\u3002<\/li><\/ul>\n\n\n\n

      \u4f46 Benchmark \u7684\u4f4e CER\/WER \u5e76\u4e0d\u81ea\u52a8\u7b49\u4ef7\u4e8e\u4ea7\u54c1\u4f53\u9a8c\u3002\u4f20\u7edf\u7cfb\u7edf\u5f80\u5f80\u9700\u8981\u6309\u8bed\u8a00\u6216\u5e02\u573a\u90e8\u7f72\u591a\u4e2a checkpoint\uff0c\u70ed\u8bcd\u66f4\u65b0\u8981\u4f9d\u8d56\u5916\u90e8 biasing \u6a21\u5757\uff0c\u53e3\u8bed\u6da6\u8272\u53c8\u8981\u989d\u5916\u4e32\u8054\u4e00\u4e2a LLM\u3002Qwen-Audio-3.0-ASR \u7684\u8bbe\u8ba1\u53d6\u820d\u662f\uff1a\u628a\u8fd9\u4e9b\u63a7\u5236\u4fe1\u53f7\u653e\u8fdb\u540c\u4e00\u4e2a instruction-following ASR \u89e3\u7801\u8fc7\u7a0b\u91cc\u3002<\/p>\n\n\n\n

      \u4e09\u3001\u6a21\u578b\u8bbe\u8ba1\uff1a\u56db\u4e2a\u6a21\u5757\u52a0\u4e00\u4e2a\u7edf\u4e00\u6307\u4ee4\u63a5\u53e3<\/h2>\n\n\n\n

      Qwen-Audio-3.0-ASR \u5ef6\u7eed\u4e0a\u4e00\u4ee3 Fun-ASR \u7684\u6574\u4f53\u62d3\u6251\uff0c\u4f46\u628a\u539f\u6765\u7684 7B Dense LLM Decoder \u6362\u6210 Qwen MoE Backbone\u3002\u4e00\u6b21\u8bf7\u6c42\u9664\u4e86\u97f3\u9891\uff0c\u8fd8\u53ef\u4ee5\u643a\u5e26\u76ee\u6807\u8bed\u8a00\u5217\u8868\u3001\u662f\u5426\u6da6\u8272\u3001\u5386\u53f2\u8f6c\u5199\uff0c\u4ee5\u53ca P0\/P1 \u70ed\u8bcd\u3002<\/p>\n\n\n\n

      • Audio Encoder\uff1a<\/strong>\u591a\u5c42 Transformer Encoder\uff0c\u4f7f\u7528 SenseVoice \u97f3\u9891\u7f16\u7801\u5668\uff0c\u628a\u6ce2\u5f62\u8f6c\u6362\u4e3a\u58f0\u5b66\u8868\u793a\u3002<\/li>
      • Audio Adapter\uff1a<\/strong>\u4e24\u5c42 Transformer Encoder\uff0c\u5c06\u58f0\u5b66\u8868\u793a\u6620\u5c04\u5230 LLM \u53ef\u6d88\u8d39\u7684\u8868\u793a\u7a7a\u95f4\u3002<\/li>
      • CTC Decoder\uff1a<\/strong>\u6302\u5728 Audio Encoder \u4e0a\u65b9\uff0c\u751f\u6210\u7c97\u7c92\u5ea6\u8bc6\u522b\u5047\u8bbe\u3002\u5b83\u4e0d\u8d1f\u8d23\u6700\u7ec8\u8f93\u51fa\uff0c\u800c\u662f\u4e3a\u70ed\u8bcd\u5b9a\u5236\u548c\u957f\u97f3\u9891\u4e0a\u4e0b\u6587\u6ce8\u5165\u63d0\u4f9b\u5019\u9009\u4e0e\u5bf9\u9f50\u4fe1\u53f7\u3002<\/li>
      • LLM Decoder\uff1a<\/strong>Qwen MoE \u89e3\u7801\u5668\uff0c\u7efc\u5408 Audio Representation\u3001CTC Hypothesis\u3001Instruction \u548c Context\uff0c\u81ea\u56de\u5f52\u751f\u6210\u6700\u7ec8\u6587\u672c\u3002<\/li><\/ul>\n\n\n\n

        \u6700\u7ec8\u8f93\u51fa\u53ef\u4ee5\u62bd\u8c61\u4e3a\uff1a<\/p>\n\n\n\n\\(P(y_{1:T}\\mid x,c)=\\prod_{t=1}^{T}P(y_t\\mid y_{<t},x,c)\\)\n\n\n\n

        \u5176\u4e2d \\(x\\) \u662f\u97f3\u9891\uff0c \\(c\\) \u662f\u8bed\u8a00\u3001\u5386\u53f2\u4e0a\u4e0b\u6587\u3001\u70ed\u8bcd\u548c\u6da6\u8272\u7b49\u6761\u4ef6\u3002\u5173\u952e\u53d8\u5316\u4e0d\u5728\u97f3\u9891\u524d\u7aef\uff0c\u800c\u5728\u89e3\u7801\u5668\uff1aMoE \u901a\u8fc7\u7a00\u758f\u4e13\u5bb6\u8def\u7531\u7ef4\u6301\u8f83\u5927\u7684\u603b\u8bed\u8a00\u5bb9\u91cf\uff0c\u540c\u65f6\u53ea\u6fc0\u6d3b\u90e8\u5206\u4e13\u5bb6\uff0c\u964d\u4f4e\u6bcf\u4e2a token \u7684\u5b9e\u9645\u8ba1\u7b97\u6210\u672c\u3002<\/p>\n\n\n\n

        \"\"
        \u201c\u97f3\u9891 + \u6307\u4ee4\u6761\u4ef6 -> CTC \u7c97\u5047\u8bbe\/\u97f3\u9891\u8868\u793a -> Qwen MoE \u89e3\u7801 -> \u6700\u7ec8\u8f6c\u5199\u201d\u7684\u6d41\u7a0b\u3002<\/figcaption><\/figure>\n\n\n\n

        \u56db\u3001\u8bad\u7ec3\u65b9\u6848\uff1a\u4e94\u9636\u6bb5\u9010\u6b65\u5bf9\u9f50\u58f0\u5b66\u3001\u8bed\u8a00\u548c\u4ea7\u54c1\u76ee\u6807<\/h2>\n\n\n\n

        \u8bad\u7ec3\u987a\u5e8f\u4e3a\uff1aAudio Encoder Pretraining -> Audio LLM Pretraining -> LLM Cooldown Adaptation -> Supervised Fine-tuning\uff08SFT\uff09-> Reinforcement Learning\uff08RL\uff09\u3002\u8fd9\u6761\u94fe\u8def\u7684\u542b\u4e49\u662f\uff0c\u5148\u83b7\u5f97\u7a33\u5065\u58f0\u5b66\u8868\u793a\uff0c\u518d\u5bf9\u9f50 Qwen \u7684\u8bed\u8a00\u7a7a\u95f4\uff0c\u968f\u540e\u7528\u9ad8\u8d28\u91cf\u6570\u636e\u6536\u655b\uff0c\u6700\u540e\u76f4\u63a5\u4f18\u5316\u751f\u4ea7\u6307\u6807\u3002<\/p>\n\n\n\n

        4.1 Audio Encoder\uff1aBEST-RQ \u81ea\u76d1\u7763 + AED \u6709\u76d1\u7763<\/h3>\n\n\n\n

        \u7b2c\u4e00\u9636\u6bb5\u4f7f\u7528 BEST-RQ \u5728\u6d77\u91cf\u65e0\u6807\u6ce8\u8bed\u97f3\u4e0a\u8bad\u7ec3\u3002\u968f\u673a\u906e\u853d\u58f0\u5b66\u5e27\uff0c\u76ee\u6807\u662f\u9884\u6d4b\u7531\u56fa\u5b9a\u968f\u673a\u6295\u5f71\u548c\u51bb\u7ed3 codebook \u751f\u6210\u7684\u79bb\u6563\u76ee\u6807\u3002\u90e8\u5206 Transformer \u5c42\u4ece\u9884\u8bad\u7ec3 Qwen3 \u521d\u59cb\u5316\uff0c\u4f46\u5c06\u8bed\u8a00\u6a21\u578b\u4e2d\u7684 causal attention \u6539\u4e3a\u53cc\u5411 attention\uff0c\u4ee5\u9002\u5e94\u8bed\u97f3\u7f16\u7801\u3002\u6570\u636e\u8986\u76d6 AI\u3001\u751f\u7269\u79d1\u6280\u3001\u7535\u5546\u3001\u6559\u80b2\u3001\u5a31\u4e50\u3001\u91d1\u878d\u3001\u51fa\u884c\u7b49\u771f\u5b9e\u573a\u666f\uff0c\u4e24\u9636\u6bb5\u5408\u8ba1\u8fbe\u5230\u6570\u5343\u4e07\u5c0f\u65f6\u3002<\/p>\n\n\n\n

        \u7b2c\u4e8c\u9636\u6bb5\u628a\u7f16\u7801\u5668\u653e\u8fdb attention-based encoder-decoder\uff08AED\uff09\u6a21\u578b\uff0c\u7528\u4e2d\u6587\u548c\u82f1\u6587\u4e3a\u4e3b\u7684\u5e26\u6807\u6ce8\u6216\u4f2a\u6807\u6ce8\u6570\u636e\u7ee7\u7eed\u8bad\u7ec3\u3002\u6570\u636e\u6d41\u7a0b\u5305\u542b VAD\u3001\u591a ASR \u6a21\u578b\u4f2a\u6807\u7b7e\uff08Paraformer-V2\u3001Whisper\u3001SenseVoice\uff09\u548c ITN\u3002\u8bad\u7ec3\u7ed3\u675f\u540e\uff0cAED Decoder \u88ab\u4e22\u5f03\uff0c\u53ea\u4fdd\u7559\u8bad\u7ec3\u597d\u7684 Audio Encoder \u4f5c\u4e3a\u540e\u7eed\u521d\u59cb\u5316\u3002<\/p>\n\n\n\n

        \"\"
        BEST-RQ \u5230 AED \u7684\u4e24\u9636\u6bb5\u97f3\u9891\u7f16\u7801\u5668\u9884\u8bad\u7ec3\u56fe<\/figcaption><\/figure>\n\n\n\n

        4.2 Audio LLM Pretraining\uff1a\u7ea6 2,000 \u4e07\u5c0f\u65f6\u505a\u97f3\u9891-\u6587\u672c\u5bf9\u9f50<\/h3>\n\n\n\n

        Audio Encoder \u5728\u8fd9\u4e00\u9636\u6bb5\u51bb\u7ed3\uff0cAudio Adapter \u548c Qwen MoE Decoder \u8054\u5408\u8bad\u7ec3\u3002\u6570\u636e\u7ea6 2,000 \u4e07\u5c0f\u65f6\uff0c\u4e3b\u8981\u7531 ASR \u6570\u636e\u548c\u4ea4\u9519 text-audio \u6570\u636e\u7ec4\u6210\uff0c\u76ee\u6807\u662f\u6587\u672c token \u7684\u4ea4\u53c9\u71b5\uff1a<\/p>\n\n\n\n\\(\\mathcal{L}_{CE}=-\\sum_{t=1}^{T}\\log P_{\\theta}(y_t\\mid y_{<t},x,c)\\)\n\n\n\n

        \u8fd9\u4e00\u9636\u6bb5\u628a\u58f0\u5b66\u8868\u793a\u5bf9\u9f50\u5230 LLM \u7684\u6587\u672c\u8868\u793a\u7a7a\u95f4\uff0c\u5f62\u6210\u591a\u8bed\u8a00\u77e5\u8bc6\u3001\u4e16\u754c\u77e5\u8bc6\u3001\u540c\u97f3\u8bcd\u548c\u5b9e\u4f53\u6d88\u6b67\u80fd\u529b\uff0c\u4ee5\u53ca\u540e\u7eed\u6307\u4ee4\u8ddf\u968f\u80fd\u529b\u3002<\/p>\n\n\n\n

        4.3 LLM Cooldown\uff1a\u51cf\u5c11\u6570\u91cf\uff0c\u589e\u52a0\u8d28\u91cf\u4e0e\u5e73\u8861<\/h3>\n\n\n\n

        Cooldown \u4f7f\u7528\u7ea6 100 \u4e07\u5c0f\u65f6\u7684\u66f4\u9ad8\u8d28\u91cf\u6570\u636e\uff0c\u5e76\u91cd\u65b0\u5e73\u8861\u8bed\u8a00\u548c\u65b9\u8a00\u5206\u5e03\uff0c\u540c\u65f6\u52a0\u5165 Audio Understanding\u3001Speaker Diarization \u548c SpeechQA\u3002ASR \u6570\u636e\u7ecf VibeVoice \u91cd\u8bc6\u522b\uff0c\u9519\u8bef\u7387\u5927\u4e8e 0.3 \u7684\u6837\u672c\u88ab\u8fc7\u6ee4\u3002Audio Encoder \u7ee7\u7eed\u51bb\u7ed3\uff0c\u53ea\u66f4\u65b0 Adapter \u548c Decoder\u3002\u5b83\u7684\u4f5c\u7528\u4e0d\u662f\u7ee7\u7eed\u5806\u6570\u636e\uff0c\u800c\u662f\u628a\u5927\u89c4\u6a21\u9884\u8bad\u7ec3\u6a21\u578b\u201c\u6536\u655b\u201d\u5230\u66f4\u53ef\u9760\u3001\u66f4\u5747\u8861\u7684\u97f3\u9891\u8bed\u8a00\u4efb\u52a1\u3002<\/p>\n\n\n\n

        4.4 SFT\uff1a\u7edf\u4e00\u6307\u4ee4\u683c\u5f0f\u8986\u76d6 30 \u8bed\u8a00\u300116 \u65b9\u8a00\u548c\u591a\u79cd\u4e1a\u52a1\u6761\u4ef6<\/h3>\n\n\n\n

        SFT \u4f7f\u7528\u6570\u5343\u4e07\u5c0f\u65f6\u8bed\u97f3\uff0c\u5305\u542b\u4eba\u5de5\u6807\u6ce8\u3001\u9ad8\u7f6e\u4fe1\u4f2a\u6807\u7b7e\u3001\u566a\u58f0\u589e\u5f3a\u3001TTS\u3001\u6d41\u5f0f\u6a21\u62df\u3001\u8fdc\u573a\/\u623f\u95f4\u6a21\u62df\u3001code-switch\u3001\u591a\u8bed\u8a00\/\u65b9\u8a00\u548c\u4e1a\u52a1\u573a\u666f\u6570\u636e\u3002Audio Encoder \u4e0e Adapter \u66f4\u65b0\uff0cBase LLM \u51bb\u7ed3\u5e76\u901a\u8fc7 LoRA \u9002\u914d Decoder\uff0c\u8bad\u7ec3\u5e8f\u5217\u6700\u591a 8,192 tokens\u3002<\/p>\n\n\n\n

        • \u6307\u5b9a\u8bed\u8a00\u8f6c\u5199\uff1a<\/strong>\u5355\u8bed\u573a\u666f\u53ea\u8f93\u51fa\u6307\u5b9a\u8bed\u8a00\uff1b\u591a\u8bed\u6df7\u8bf4\u4fdd\u7559\u5404\u8bed\u8a00\u539f\u59cb\u4e66\u5199\u5f62\u5f0f\uff0c\u51cf\u5c11 language leakage\u3002<\/li>
        • \u884c\u4e1a\u5b9e\u4f53\u589e\u5f3a\uff1a<\/strong>\u4ece AI\u3001\u91d1\u878d\u3001\u533b\u7597\u3001IT \u7b49\u9886\u57df\u6316\u6398\u5b9e\u4f53\uff0c\u7ecf\u8fc7\u201c\u5b9e\u4f53 -> \u4e8b\u5b9e\u68c0\u7d22 -> LLM \u9020\u53e5 -> CosyVoice3 \u5408\u6210 -> \u5f53\u524d ASR \u91cd\u8bc6\u522b\u8fc7\u6ee4 -> SFT\u201d\u6d41\u7a0b\uff0c\u5f3a\u5316\u957f\u5c3e\u8bcd\u3002<\/li>
        • \u539f\u751f\u6da6\u8272\uff1a<\/strong>\u7528\u4e8c\u503c\u6307\u4ee4\u63a7\u5236 verbatim \u6216 polished \u8f93\u51fa\uff0c\u5b66\u4e60\u5220\u9664 filler\u3001\u6d88\u9664\u91cd\u590d\u3001\u5904\u7406 self-correction\u3001\u8865\u6807\u70b9\u548c\u89c4\u8303\u683c\u5f0f\u3002<\/li>
        • \u5386\u53f2\u4e0a\u4e0b\u6587\uff1a<\/strong>\u5c06\u540c\u4e00 session \u7684\u524d\u5e8f\u8f6c\u5199\u4f5c\u4e3a context\uff0c\u540c\u65f6\u8bad\u7ec3\u6b63\u786e\u3001\u51b2\u7a81\u3001\u566a\u58f0\u548c\u7a7a context\uff0c\u660e\u786e\u544a\u8bc9\u6a21\u578b\u97f3\u9891\u59cb\u7ec8\u662f\u4e3b\u8981\u8bc1\u636e\u3002<\/li>
        • \u5206\u5c42\u70ed\u8bcd\uff1a<\/strong>P0 \u662f\u9ad8\u7f6e\u4fe1\u3001\u9ad8\u4f18\u5148\u7ea7\u8bcd\uff1bP1 \u662f\u66f4\u5bbd\u7684\u5019\u9009\u6c60\u3002LLM \u6839\u636e reference transcript \u505a\u8bed\u4e49\u5339\u914d\uff0cP0 bias \u66f4\u5f3a\u4f46\u4ecd\u9700\u5f97\u5230\u58f0\u5b66\u548c\u8bed\u4e49\u8bc1\u636e\u652f\u6301\u3002<\/li><\/ul>\n\n\n\n

          \u5206\u5c42\u70ed\u8bcd\u53ef\u4ee5\u5f62\u5f0f\u5316\u4e3a\uff1a\u5bf9 P0 \u5019\u9009\u4f7f\u7528\u66f4\u9ad8\u7684\u6761\u4ef6\u6743\u91cd\uff0c\u5bf9 P1 \u4f7f\u7528\u8f83\u5f31\u6743\u91cd\uff1b\u76ee\u6807\u4e0d\u662f\u628a\u5019\u9009\u8bcd\u786c\u585e\u8fdb\u7ed3\u679c\uff0c\u800c\u662f\u5728\u53ec\u56de\u957f\u5c3e\u8bcd\u548c\u6291\u5236\u9519\u8bef\u63d2\u5165\u4e4b\u95f4\u53d6\u5f97\u5e73\u8861\u3002<\/p>\n\n\n\n

          4.5 RL\uff1aFunVerl-ASR + GRPO\uff0c\u76f4\u63a5\u4f18\u5316 ASR \u4ea7\u54c1\u6307\u6807<\/h3>\n\n\n\n

          RL \u6570\u636e\u96c6\u5f3a\u8c03 compact\u3001high-quality\u3001hard cases\uff1a\u6a21\u578b\u4e0e Whisper\/FireRed-ASR\/SenseVoice \u4e0d\u4e00\u81f4\u7684\u6837\u672c\u3001\u8d85\u8fc7 20 \u79d2\u7684\u957f\u8bed\u97f3\u3001\u5e7b\u89c9\u6837\u672c\u3001\u5173\u952e\u8bcd\/\u70ed\u8bcd\u6837\u672c\uff0c\u4ee5\u53ca\u7528\u4e8e\u9632\u6b62\u707e\u96be\u6027\u9057\u5fd8\u7684\u666e\u901a ASR \u6570\u636e\u3002<\/p>\n\n\n\n

          FunVerl-ASR \u4e3a\u97f3\u9891\u5927\u6a21\u578b\u6269\u5c55\u4e86\u666e\u901a\u6587\u672c RL \u6846\u67b6\uff0c\u5305\u542b\u56db\u4e2a\u5e76\u884c\u6a21\u5757\uff1a<\/p>\n\n\n\n

          • Audio Processor\uff1a\u6ce2\u5f62\u8f6c Mel \u7279\u5f81\uff0c\u6307\u4ee4\u6587\u672c\u8f6c token\u3002<\/li>
          • vLLM Rollout Engine\uff1a\u51bb\u7ed3 Audio Encoder\/Adapter \u751f\u6210\u97f3\u9891 embedding\uff0c\u4e0e\u6587\u672c embedding \u5408\u5e76\u540e\u7531 MoE \u751f\u6210\u591a\u4e2a\u5019\u9009\u3002<\/li>
          • Reward Module\uff1a\u652f\u6301 CER\u3001\u5173\u952e\u8bcd\u53ec\u56de\u3001\u70ed\u8bcd\u547d\u4e2d\u3001\u4e0a\u4e0b\u6587\u5b9e\u4f53\u5339\u914d\u3001\u5e7b\u89c9\u68c0\u6d4b\u3001\u8bed\u8a00\u4e00\u81f4\u6027\u548c LLM Judge\u3002<\/li>
          • Megatron Policy Trainer\uff1a\u4f7f\u7528 tensor\/pipeline\/expert parallel\uff0c\u66f4\u65b0\u540e\u901a\u8fc7 NCCL \u5c06 Decoder \u6743\u91cd\u540c\u6b65\u56de rollout worker\u3002<\/li><\/ul>\n\n\n\n

            \u5176\u5de5\u7a0b\u521b\u65b0\u662f fully asynchronous loop\uff1arollout \u6301\u7eed\u751f\u6210\u8f68\u8ff9\u5e76\u653e\u5165\u6d88\u606f\u961f\u5217\uff0ctrainer \u72ec\u7acb\u6d88\u8d39\u5e76\u66f4\u65b0\u7b56\u7565\uff1b\u5f53\u8f68\u8ff9\u5bf9\u5e94\u7684\u65e7\u7b56\u7565\u8fc7\u4e8e\u9648\u65e7\u65f6\u76f4\u63a5\u4e22\u5f03\u3002\u9ed8\u8ba4\u6bcf\u4e2a prompt \u751f\u6210 8 \u4e2a hypothesis\uff0c\u6bcf\u6b21 actor update \u4f7f\u7528 32 \u4e2a prompt\uff0c\u5728 32 \u5f20 A100\uff08rollout \u548c\u8bad\u7ec3\u5404 16 \u5f20\uff09\u4e0a\uff0c\u4e00\u5929\u5185\u5b8c\u6210\u5b8c\u6574 RL\u3002<\/p>\n\n\n\n

            GRPO \u4e0d\u9700\u8981\u5355\u72ec\u8bad\u7ec3 value network\u3002\u7ec4\u5185 advantage \u4e3a\uff1a<\/p>\n\n\n\n\\(\\hat{A}_{i,t}=\\frac{R_i-\\operatorname{mean}(\\{R_j\\}_{j=1}^{G})}{\\operatorname{std}(\\{R_j\\}_{j=1}^{G})}\\)\n\n\n\n

            \u7b56\u7565\u76ee\u6807\u4f7f\u7528 clipped objective \u548c KL \u60e9\u7f5a\uff1a<\/p>\n\n\n\n\\(\\mathcal{L}_{GRPO}(\\theta)=\\frac{1}{G}\\sum_{i=1}^{G}\\frac{1}{|o_i|}\\sum_t\\min\\left(r_{i,t}(\\theta)\\hat{A}_{i,t},\\operatorname{clip}(r_{i,t}(\\theta),1-\\epsilon,1+\\epsilon)\\hat{A}_{i,t}\\right)-\\beta D_{KL}(\\pi_{\\theta}\\parallel\\pi_{ref})\\)\n\n\n\n

            \u5f53 reward \u53ea\u6709 WER \u65f6\uff0cGRPO \u4e0e ASR \u4e2d\u7684 MWER \u8bad\u7ec3\u63a5\u8fd1\uff1b\u8bba\u6587\u8fdb\u4e00\u6b65\u4f7f\u7528\u7ec4\u5408 reward\uff1a<\/p>\n\n\n\n

            • \\(R_1=1-WER\\)\uff1a\u603b\u4f53\u8bc6\u522b\u51c6\u786e\u7387\u3002<\/li>
            • \u5173\u952e\u8bcd\u51c6\u786e\u7387\u4e0e\u53ec\u56de\u7387\uff1a\u65e2\u5956\u52b1\u547d\u4e2d\uff0c\u4e5f\u60e9\u7f5a\u9519\u8bef\u63d2\u5165\u3002<\/li>
            • \u70ed\u8bcd reward\uff1a\u5206\u522b\u8bc4\u4ef7 P0\/P1\uff0c\u5e76\u4f7f\u7528\u4e0d\u5bf9\u79f0\u6b63\u8d1f\u6743\u91cd\u3002<\/li>
            • \u4e0a\u4e0b\u6587 reward\uff1a\u5956\u52b1\u547d\u4e2d\u5386\u53f2\u5b9e\u4f53\uff0c\u60e9\u7f5a\u9519\u8bef\u590d\u5236\u3002<\/li>
            • \u5e7b\u89c9\u6291\u5236\uff1a\u6309\u5e7b\u89c9 span \u957f\u5ea6\u60e9\u7f5a\u3002<\/li><\/ul>\n\n\n\n
              \"\"
              FunVerl-ASR \u7684\u5f02\u6b65 RL \u67b6\u6784\u56fe\uff0c\u7a81\u51fa rollout\u3001reward\u3001policy trainer \u548c NCCL \u6743\u91cd\u540c\u6b65<\/figcaption><\/figure>\n\n\n\n

              \u4e94\u3001\u90e8\u7f72\u5f62\u6001\uff1a\u4e00\u4e2a\u6a21\u578b\uff0c\u4e09\u79cd\u4ea7\u54c1\u80fd\u529b<\/h2>\n\n\n\n

              5.1 \u591a\u8bed\u8a00\u4e0e\u65b9\u8a00\uff1a\u4ece\u591a checkpoint \u53d8\u6210\u8fd0\u884c\u65f6\u6307\u4ee4<\/h3>\n\n\n\n

              \u4e0a\u4e00\u4ee3 Fun-ASR \u9700\u8981\u4e2d\u6587\u82f1\u6587\u4e3b\u6a21\u578b\u548c\u72ec\u7acb\u7684 Fun-ASR-ML \u591a\u8bed\u8a00 checkpoint\u3002Qwen-Audio-3.0-ASR \u6539\u6210\u4e00\u4e2a\u6a21\u578b\u52a0 runtime language instruction\uff0c\u4f8b\u5982 Chinese+English+Japanese\u3001Chinese+French\u3001English+Indonesian\u3002\u652f\u6301\u8bed\u8a00\u5305\u62ec\u4e2d\u6587\u3001\u65e5\u8bed\u3001\u97e9\u8bed\u3001\u8d8a\u5357\u8bed\u3001\u6cf0\u8bed\u3001\u5370\u5c3c\u8bed\u3001\u9a6c\u6765\u8bed\u3001\u83f2\u5f8b\u5bbe\u8bed\u3001\u5370\u5730\u8bed\u3001\u963f\u62c9\u4f2f\u8bed\uff0c\u4ee5\u53ca\u82f1\u8bed\u3001\u6cd5\u8bed\u3001\u5fb7\u8bed\u3001\u897f\u73ed\u7259\u8bed\u3001\u8461\u8404\u7259\u8bed\u3001\u4fc4\u8bed\u3001\u610f\u5927\u5229\u8bed\u3001\u8377\u5170\u8bed\u3001\u745e\u5178\u8bed\u3001\u4e39\u9ea6\u8bed\u3001\u82ac\u5170\u8bed\u3001\u5e0c\u814a\u8bed\u3001\u6ce2\u5170\u8bed\u3001\u6377\u514b\u8bed\u3001\u5308\u7259\u5229\u8bed\u3001\u7f57\u9a6c\u5c3c\u4e9a\u8bed\u3001\u4fdd\u52a0\u5229\u4e9a\u8bed\u3001\u514b\u7f57\u5730\u4e9a\u8bed\u3001\u65af\u6d1b\u4f10\u514b\u8bed\u548c\u632a\u5a01\u8bed\u3002<\/p>\n\n\n\n

              5.2 Streaming\uff1a\u5148\u5feb\u540e\u51c6\u7684\u6df7\u5408\u5f0f\u89e3\u7801<\/h3>\n\n\n\n

              \u6d41\u5f0f\u6a21\u578b\u5148\u7528\u77ed chunk \u548c\u6709\u9650 right context \u8fde\u7eed\u8f93\u51fa\u4e34\u65f6\u6587\u672c\uff1b\u7aef\u70b9\u68c0\u6d4b\u540e\uff0c\u518d\u7528\u5b8c\u6574\u4e0a\u4e0b\u6587\u5927\u6a21\u578b\u91cd\u7b97\u5e76\u5237\u65b0\u6700\u7ec8\u6587\u672c\u3002chunk size \u4e0e right context \u53ef\u4ee5\u72ec\u7acb\u8c03\u6574\uff1a\u4e0a\u4e0b\u6587\u8d8a\u77ed\uff0c\u54cd\u5e94\u8d8a\u5feb\uff1b\u4e0a\u4e0b\u6587\u8d8a\u957f\uff0c\u51c6\u786e\u7387\u8d8a\u9ad8\u3002\u8bba\u6587\u5728 H100 \u4e0a\u6d4b\u91cf first-token latency \u548c text-display latency\uff0c\u4f4e\u5ef6\u8fdf\u914d\u7f6e\u4e0b\u524d\u8005\u4f4e\u4e8e 200 ms\uff0c\u540e\u8005\u4f4e\u4e8e 300 ms\u3002<\/p>\n\n\n\n

              5.3 Message ASR\uff1a\u9762\u5411\u4f1a\u8bae\u3001\u91c7\u8bbf\u548c\u5ba2\u670d\u7684\u7ec4\u5408\u5f0f\u63a5\u53e3<\/h3>\n\n\n\n

              Message ASR \u4e0d\u662f\u53e6\u4e00\u4e2a\u6a21\u578b\uff0c\u800c\u662f\u628a streaming audio\u3001\u5386\u53f2\u4e0a\u4e0b\u6587\u3001P0\/P1 \u70ed\u8bcd\u3001native polishing \u548c right context \u7ec4\u5408\u6210\u4e00\u4e2a\u90e8\u7f72 profile\u3002\u5b83\u5728\u8bf4\u8bdd\u8fc7\u7a0b\u4e2d\u63d0\u4f9b partial result\uff0c\u7ed3\u675f\u540e\u7528\u5b8c\u6574 utterance \u5237\u65b0\u7ed3\u679c\uff1b\u4e0a\u4e0b\u6587\u5e2e\u52a9\u7ef4\u6301\u4e3b\u9898\u548c\u5b9e\u4f53\u4e00\u81f4\u6027\uff0c\u70ed\u8bcd\u5e2e\u52a9\u8bc6\u522b\u4eba\u540d\u3001\u4ea7\u54c1\u548c\u6280\u672f\u672f\u8bed\uff0c\u4f46\u97f3\u9891\u8bc1\u636e\u59cb\u7ec8\u4f18\u5148\u4e8e context \u548c hotword\u3002<\/p>\n\n\n\n

              \"\"
              Message ASR \u5de5\u4f5c\u6d41\u56fe: \u201c\u8bc6\u522b\u3001\u4e0a\u4e0b\u6587\u3001\u6da6\u8272\u548c\u6d41\u5f0f\u5ef6\u8fdf\u5728\u540c\u4e00\u63a5\u53e3\u5185\u5b8c\u6210\u201d<\/figcaption><\/figure>\n\n\n\n

              \u516d\u3001\u516c\u5f00\u57fa\u51c6\uff1a\u5728\u7edf\u4e00\u8bc4\u6d4b\u4e0b\uff0c\u4e2d\u6587\u3001\u82f1\u6587\u548c\u591a\u8bed\u8a00\u90fd\u5177\u7ade\u4e89\u529b<\/h2>\n\n\n\n

              \u8bba\u6587\u540c\u65f6\u4f7f\u7528\u516c\u5f00\u6570\u636e\u96c6\u548c\u5185\u90e8\u5de5\u4e1a\u6d4b\u8bd5\u96c6\u3002\u516c\u5f00\u96c6\u5305\u62ec AISHELL-1\u3001AISHELL-2\u3001FLEURS\u3001LibriSpeech\u3001WeNetSpeech\u3001GigaSpeechBench \u548c Common Voice 15\u3002\u4e2d\u6587\u4f7f\u7528 CER\uff0c\u82f1\u6587\u4f7f\u7528 WER\uff1b\u6570\u503c\u8d8a\u4f4e\u8d8a\u597d\u3002<\/p>\n\n\n\n

              \u4f5c\u8005\u5c06\u4e2d\u6587\u82f1\u6587\u7ed3\u679c\u5206\u4e3a\u4e24\u7ec4\uff1aPanel A \u662f\u5404\u6a21\u578b\u5b98\u65b9\u62a5\u544a\u7ed3\u679c\uff0cPanel B \u662f\u6240\u6709\u7cfb\u7edf\u901a\u8fc7\u540c\u4e00\u4e2a API \u8bc4\u6d4b\u6d41\u7a0b\u5f97\u5230\u7684\u7ed3\u679c\u3002Panel A \u53ef\u80fd\u5b58\u5728\u9884\u5904\u7406\u3001\u53c2\u8003\u6587\u672c\u548c\u6253\u5206\u7ea6\u5b9a\u5dee\u5f02\uff0c\u56e0\u6b64\u771f\u6b63\u9002\u5408\u6a2a\u5411\u6bd4\u8f83\u7684\u662f Panel B\u3002<\/p>\n\n\n\n

              \u6d4b\u8bd5\u96c6<\/th>\u6307\u6807<\/th>GPT-4o Transcribe<\/th>Azure<\/th>Doubao-ASR<\/th>Fun-ASR-Flash<\/th>Tencent Hy-ASR-3.0-preview<\/th>Qwen-Audio-3.0-ASR<\/th><\/tr><\/thead>
              AISHELL-1<\/td>CER<\/td>3.34<\/td>1.60<\/td>1.54<\/td>0.76<\/td>0.78<\/td>1.03<\/td><\/tr>
              AISHELL-2<\/td>CER<\/td>4.47<\/td>3.24<\/td>2.66<\/td>1.93<\/td>2.15<\/td>2.02<\/td><\/tr>
              FLEURS-zh<\/td>CER<\/td>6.14<\/td>5.51<\/td>8.18<\/td>5.52<\/td>2.38<\/td>2.08<\/td><\/tr>
              FLEURS-en<\/td>WER<\/td>5.79<\/td>5.47<\/td>6.50<\/td>4.95<\/td>5.45<\/td>4.17<\/td><\/tr>
              LibriSpeech-clean<\/td>WER<\/td>1.58<\/td>1.90<\/td>3.36<\/td>1.35<\/td>1.60<\/td>1.19<\/td><\/tr>
              LibriSpeech-other<\/td>WER<\/td>3.87<\/td>3.54<\/td>4.46<\/td>2.51<\/td>2.68<\/td>2.24<\/td><\/tr>
              WeNetSpeech-net<\/td>CER<\/td>12.27<\/td>4.58<\/td>4.81<\/td>3.83<\/td>4.16<\/td>4.31<\/td><\/tr><\/tbody><\/table>
              \u8868 1\uff1a\u7edf\u4e00 API \u8bc4\u6d4b\u7ed3\u679c\uff0c\u5355\u4f4d\u4e3a\u767e\u5206\u6bd4\uff0c\u8d8a\u4f4e\u8d8a\u597d\u3002<\/figcaption><\/figure>\n\n\n\n

              \u5728 Panel B \u4e2d\uff0cQwen-Audio-3.0-ASR \u5728 7 \u4e2a\u6d4b\u8bd5\u96c6\u4e2d\u7684 4 \u4e2a\u53d6\u5f97\u6700\u4f4e\u9519\u8bef\u7387\uff1aFLEURS-zh 2.08%\u3001FLEURS-en 4.17%\u3001LibriSpeech-clean 1.19%\u3001LibriSpeech-other 2.24%\u3002\u76f8\u5bf9\u540c\u4e00 pipeline \u4e0b\u6700\u5f3a\u7ade\u4e89\u7cfb\u7edf\uff0c\u5206\u522b\u964d\u4f4e 0.30\u30010.78\u30010.16 \u548c 0.27 \u4e2a\u767e\u5206\u70b9\uff1bAISHELL-2 \u4e5f\u53ea\u843d\u540e\u6700\u4f73\u7ed3\u679c 0.09 \u4e2a\u767e\u5206\u70b9\u3002<\/p>\n\n\n\n

              6.1 \u591a\u8bed\u8a00\u7ed3\u679c\uff1aGigaSpeechBench \u4e0e Common Voice 15 \u7684\u5b8f\u5e73\u5747\u9886\u5148<\/h3>\n\n\n\n
              \u6570\u636e\u96c6<\/th>Qwen-Audio-3.0-ASR \u5b8f\u5e73\u5747<\/th>\u5bf9\u6bd4\u7ed3\u8bba<\/th><\/tr><\/thead>
              GigaSpeechBench<\/td>22.50%<\/td>\u5728\u5b8c\u6574\u8986\u76d6\u7684\u7cfb\u7edf\u4e2d\u6700\u4f4e\uff1b7 \u79cd\u8bed\u8a00\u4e2d 5 \u79cd\u6700\u4f4e<\/td><\/tr>
              Common Voice 15<\/td>4.57%<\/td>\u5b8f\u5e73\u5747\u6700\u4f4e\uff1b7 \u79cd\u8bed\u8a00\u4e2d 6 \u79cd\u6700\u4f4e<\/td><\/tr>
              FLEURS<\/td>4.94%<\/td>9 \u79cd\u8bed\u8a00\u4e2d 4 \u79cd\u6700\u4f4e\uff0c\u6574\u4f53\u5177\u6709\u7ade\u4e89\u529b<\/td><\/tr><\/tbody><\/table>
              \u8868 2\uff1a\u591a\u8bed\u8a00\u5b8f\u5e73\u5747\u7ed3\u679c\u3002\u5b8f\u5e73\u5747\u53ea\u7edf\u8ba1\u8986\u76d6\u5bf9\u5e94\u5168\u90e8\u8bed\u8a00\u7684\u7cfb\u7edf\u3002<\/figcaption><\/figure>\n\n\n\n

              \u8fd9\u7ec4\u7ed3\u679c\u4f53\u73b0\u4e86\u8bba\u6587\u7684\u4e3b\u8981\u5e73\u8861\uff1a\u6a21\u578b\u4e0d\u662f\u5728\u6bcf\u4e00\u79cd\u8bed\u8a00\u3001\u6bcf\u4e00\u4e2a\u6d4b\u8bd5\u96c6\u4e0a\u90fd\u7edd\u5bf9\u7b2c\u4e00\uff0c\u4f46\u5728\u8de8\u8bed\u8a00\u8986\u76d6\u9762\u548c\u5b8f\u5e73\u5747\u7a33\u5b9a\u6027\u4e0a\u8868\u73b0\u5f88\u5f3a\uff0c\u5c24\u5176\u9002\u5408\u4e00\u4e2a\u6a21\u578b\u670d\u52a1\u591a\u4e2a\u5e02\u573a\u7684\u573a\u666f\u3002<\/p>\n\n\n\n

              \u4e03\u3001\u5185\u90e8\u5de5\u4e1a\u8bc4\u6d4b\uff1a\u8bba\u6587\u771f\u6b63\u60f3\u8bc1\u660e\u7684\u56db\u9879\u80fd\u529b<\/h2>\n\n\n\n

              7.1 16 \u79cd\u4e2d\u6587\u65b9\u8a00\uff1aCER \u4e0e\u4efb\u52a1\u4e00\u81f4\u6027\u540c\u65f6\u8bc4\u4f30<\/h3>\n\n\n\n

              \u65b9\u8a00\u6d4b\u8bd5\u8986\u76d6 16 \u79cd\u65b9\u8a00\u3002\u56db\u5ddd\u3001\u5c71\u897f\u3001\u6cb3\u5357\u3001\u6d4e\u5357\u3001\u7ca4\u8bed\u3001\u9655\u897f\u3001\u9752\u5c9b\u91c7\u7528 ASR \u8bbe\u7f6e\uff0c\u8981\u6c42\u4fdd\u7559\u76ee\u6807\u65b9\u8a00\uff1b\u4e0a\u6d77\u3001\u5357\u660c\u3001\u5b81\u6ce2\u3001\u5ba2\u5bb6\u3001\u676d\u5dde\u3001\u6e29\u5dde\u3001\u6e56\u5357\u3001\u798f\u5efa\u3001\u82cf\u5dde\u91c7\u7528 AST \u8bbe\u7f6e\uff0c\u8981\u6c42\u7ffb\u8bd1\u4e3a\u81ea\u7136\u666e\u901a\u8bdd\u3002<\/p>\n\n\n\n

              \u4e00\u81f4\u6027\u7531 Qwen3.7-Max \u6309 0-10 \u5206\u8bc4\u4ef7\uff0c\u5206\u6570\u81f3\u5c11\u4e3a 6 \u624d\u7b97\u4e00\u81f4\u3002\u8ba1\u7b97\u516c\u5f0f\u4e3a\uff1a<\/p>\n\n\n\n\\(Consistency=\\frac{100}{N}\\sum_{i=1}^{N}\\mathbf{1}[s_i\\ge 6]\\)\n\n\n\n

              Qwen-Audio-3.0-ASR \u5728 16 \u4e2a\u65b9\u8a00\u5b50\u96c6\u4e2d\u7684 11 \u4e2a\u53d6\u5f97\u6700\u4f4e CER\uff0c\u5b8f\u5e73\u5747 CER \u4e3a 9.40%\uff0c\u5bf9\u6bd4 Doubao-ASR \u7684 16.77% \u548c Tencent Hy-ASR-3.0-preview \u7684 20.40%\u3002\u5728\u6e29\u5dde\u8bdd\u4e0a\uff0c\u6a21\u578b CER \u4e3a 14.22%\uff0c\u800c\u4e24\u4e2a\u5bf9\u6bd4\u7cfb\u7edf\u5206\u522b\u4e3a 64.63% \u548c 64.66%\uff0c\u5dee\u8ddd\u5c24\u5176\u660e\u663e\u3002<\/p>\n\n\n\n

              \"\"
              16 \u65b9\u8a00 CER<\/figcaption><\/figure>\n\n\n\n
              \"\"
              \u4efb\u52a1\u4e00\u81f4\u6027\u7387<\/figcaption><\/figure>\n\n\n\n

              7.2 \u884c\u4e1a\u5b9e\u4f53\uff1a15 \u4e2a\u9886\u57df\u5168\u90e8\u53d6\u5f97\u6700\u9ad8\u53ec\u56de<\/h3>\n\n\n\n

              \u5185\u90e8\u884c\u4e1a\u6d4b\u8bd5\u8986\u76d6 15 \u4e2a\u9886\u57df\uff0c\u5305\u62ec AI\u3001\u533b\u7597\u3001\u519c\u4e1a\u3001\u5236\u9020\u3001\u6587\u5316\u4f53\u80b2\u5a92\u4f53\u3001\u8f6f\u4ef6\u4e0e IT \u7b49\u3002Qwen-Audio-3.0-ASR \u5728\u6240\u6709\u9886\u57df\u7684 entity recall \u90fd\u9ad8\u4e8e Doubao-ASR \u548c Tencent Hy-ASR-3.0-preview\uff0c\u8f6f\u4ef6\u4e0e IT \u7684\u63d0\u5347\u6700\u660e\u663e\uff0c\u56e0\u4e3a\u5176\u4e2d\u5305\u542b\u5927\u91cf\u4f4e\u9891\u3001\u5feb\u901f\u53d8\u5316\u7684\u8f6f\u4ef6\u5e93\u3001\u6846\u67b6\u548c API \u540d\u79f0\u3002<\/p>\n\n\n\n

              \u8fd9\u91cc\u9700\u8981\u4fdd\u6301\u4e25\u8c28\uff1a\u8bba\u6587\u8ba4\u4e3a\u5b9e\u4f53\u6316\u6398\u548c\u6570\u636e\u5408\u6210 pipeline \u5f88\u53ef\u80fd\u662f\u4e3b\u8981\u8d21\u732e\u4e4b\u4e00\uff0c\u4f46\u4e0d\u80fd\u636e\u6b64\u8bc1\u660e\u6240\u6709\u63d0\u5347\u90fd\u7531\u8be5\u6a21\u5757\u5355\u72ec\u5e26\u6765\uff0c\u56e0\u4e3a MoE \u89e3\u7801\u5668\u3001\u66f4\u591a\u6570\u636e\u548c RL \u4e5f\u53ef\u80fd\u5171\u540c\u8d21\u732e\u3002<\/p>\n\n\n\n

              \"\"
              15 \u4e2a\u884c\u4e1a\u7684\u5b9e\u4f53\u53ec\u56de\u96f7\u8fbe\u56fe<\/figcaption><\/figure>\n\n\n\n

              7.3 \u5206\u5c42\u70ed\u8bcd\uff1aP0 \u53ec\u56de\u63a5\u8fd1 99%\uff0cP1 \u6709\u610f\u4fdd\u6301\u66f4\u5f31 bias<\/h3>\n\n\n\n
              \u7c7b\u522b<\/th>Qwen \u65e0\u70ed\u8bcd<\/th>Qwen + \u70ed\u8bcd<\/th>\u8bf4\u660e<\/th><\/tr><\/thead>
              Person names P0<\/td>62.12%<\/td>99.43%<\/td>\u63d0\u5347 37.31 \u4e2a\u767e\u5206\u70b9<\/td><\/tr>
              Subject terms P0<\/td>69.36%<\/td>99.42%<\/td>\u9ad8\u4f18\u5148\u7ea7\u8bcd\u5f3a\u589e\u5f3a<\/td><\/tr>
              Trending buzzwords P0<\/td>63.08%<\/td>99.46%<\/td>\u52a8\u6001\u70ed\u8bcd\u6548\u679c\u660e\u663e<\/td><\/tr>
              AI entities P0<\/td>88.22%<\/td>99.39%<\/td>\u63a5\u8fd1\u9971\u548c\u53ec\u56de<\/td><\/tr>
              Brand names P0<\/td>83.83%<\/td>97.95%<\/td>\u5b9e\u4f53\u7c7b\u522b\u7a33\u5b9a\u63d0\u5347<\/td><\/tr>
              Product entities P0<\/td>72.87%<\/td>99.07%<\/td>\u957f\u5c3e\u4ea7\u54c1\u540d\u53d7\u76ca\u660e\u663e<\/td><\/tr>
              Subject terms P1<\/td>88.46%<\/td>95.31%<\/td>\u8f83\u5f31 bias\uff0c\u907f\u514d\u8bef\u63d2\u8bcd<\/td><\/tr>
              Trending buzzwords P1<\/td>74.00%<\/td>88.34%<\/td>\u4fdd\u5b88\u589e\u5f3a<\/td><\/tr><\/tbody><\/table>
              \u8868 3\uff1aQwen-Audio-3.0-ASR \u4e0d\u540c\u70ed\u8bcd\u5c42\u7ea7\u7684\u53ec\u56de\u7387\u3002<\/figcaption><\/figure>\n\n\n\n

              7.4 \u957f\u97f3\u9891\u4e0a\u4e0b\u6587\uff1a\u628a\u524d\u6587\u53d8\u6210 running context cache<\/h3>\n\n\n\n

              \u5f88\u591a\u9519\u8bef\u4e0d\u662f\u5f53\u524d\u7247\u6bb5\u7684\u58f0\u5b66\u4fe1\u606f\u4e0d\u8db3\uff0c\u800c\u662f\u540c\u97f3\u5b9e\u4f53\u3001\u6587\u4ef6\u540d\u3001\u6280\u672f\u7f29\u5199\u6216\u4eba\u540d\u9700\u8981\u8de8\u53e5\u8bb0\u5fc6\u3002\u6a21\u578b\u628a\u540c\u4e00 session \u4e2d\u5df2\u7ecf\u8bc6\u522b\u7684\u524d\u5e8f\u6587\u672c\u4f5c\u4e3a\u4e0a\u4e0b\u6587\uff0c\u4e0d\u8981\u6c42\u4eba\u5de5\u7ef4\u62a4\u5b8c\u6574\u70ed\u8bcd\u8868\u3002\u4f8b\u5982\u524d\u6587\u5df2\u7ecf\u786e\u8ba4\u201cstyle.xsl\u201d\u3001TCC\u3001unigram\u3001softmax\u3001debug \u6216\u67d0\u4e2a\u4eba\u540d\uff0c\u540e\u7eed\u518d\u6b21\u51fa\u73b0\u65f6\u53ef\u4ee5\u6291\u5236\u58f0\u5b66\u4e0a\u76f8\u8fd1\u4f46\u9519\u8bef\u7684\u66ff\u4ee3\u8bcd\u3002<\/p>\n\n\n\n

              \u524d\u6587\u7ebf\u7d22<\/th>\u53c2\u8003<\/th>\u65e0\u957f\u4e0a\u4e0b\u6587<\/th>\u6709\u957f\u4e0a\u4e0b\u6587<\/th><\/tr><\/thead>
              xsltproc style.xsl input.xml<\/td>style.xsl data.xml > out.html<\/td>xsl data.xml > out.html<\/td>style.xsl data.xml > out.html<\/td><\/tr>
              \u201c\u8fd9\u53eb\u6cd5\u6761\u7ade\u5408\u201d<\/td>\u6cd5\u6761\u7ade\u5408<\/td>\u5212\u6389\u5373\u53ef<\/td>\u6cd5\u6761\u7ade\u5408<\/td><\/tr>
              unigram \u5df2\u51fa\u73b0\u4e24\u6b21<\/td>\u6211\u4eec\u79f0\u4e4b\u4e3a unigram<\/td>unique<\/td>unigram<\/td><\/tr>
              TCC \u5df2\u88ab\u53cd\u590d\u786e\u8ba4<\/td>our TCC will be…<\/td>our GGC will be…<\/td>our TCC will be…<\/td><\/tr>
              softmax \u5df2\u51fa\u73b0<\/td>softmax \u5f62\u5f0f<\/td>softmap \u5f62\u5f0f<\/td>softmax \u5f62\u5f0f<\/td><\/tr><\/tbody><\/table>
              \u8868 4\uff1a\u957f\u97f3\u9891\u4e0a\u4e0b\u6587\u5e26\u6765\u7684\u4ee3\u8868\u6027\u7ea0\u9519\u3002<\/figcaption><\/figure>\n\n\n\n

              7.5 \u539f\u751f\u5355\u6b21\u6da6\u8272\uff1a\u53ef\u8bfb\u6027\u63a5\u8fd1\u4e24\u9636\u6bb5\u7ea7\u8054<\/h3>\n\n\n\n

              \u4f20\u7edf\u505a\u6cd5\u662f ASR \u5148\u8f93\u51fa\u539f\u59cb\u8f6c\u5199\uff0c\u518d\u8c03\u7528\u4e00\u4e2a LLM \u6e05\u7406 filler\u3001\u91cd\u590d\u548c\u81ea\u6211\u7ea0\u6b63\u3002Qwen-Audio-3.0-ASR \u7528\u4e00\u4e2a polishing flag \u5728\u540c\u4e00\u6b21\u81ea\u56de\u5f52\u89e3\u7801\u4e2d\u9009\u62e9 verbatim \u6216 polished \u8f93\u51fa\u3002<\/p>\n\n\n\n

              \u5185\u90e8 5 \u5206\u5236\u8bc4\u6d4b\u4e2d\uff0c\u539f\u59cb\u8f93\u51fa\u53ef\u8bfb\u6027\u4e3a 2.53\uff0c\u539f\u751f\u6da6\u8272\u63d0\u5347\u5230 3.44\uff1b\u5fe0\u5b9e\u5ea6\u4e3a 3.44\uff0c\u4e0e\u4e24\u6b21\u8c03\u7528\u7684 ASR + Qwen3.6-Plus \u7ea7\u8054\u65b9\u6848 3.47 \u63a5\u8fd1\uff0c\u4f46\u53ea\u9700\u8981\u4e00\u6b21\u63a8\u7406\u3002\u4ee3\u8868\u6027\u884c\u4e3a\u5305\u62ec\u5220\u9664 um\/uh\u3001\u6d88\u9664 the the the\u3001\u89e3\u6790 false start\u3001\u8865\u5145\u6807\u70b9\u5e76\u89c4\u8303\u7f29\u5199\u3002<\/p>\n\n\n\n

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

              7.6 \u6d41\u5f0f\u5ef6\u8fdf-\u51c6\u786e\u7387\uff1a400 ms \u662f\u4e00\u4e2a\u53ef\u7528\u6298\u4e2d\u70b9<\/h3>\n\n\n\n

              \u968f\u7740 algorithmic latency \u589e\u52a0\uff0c\u4e2d\u6587 CER \u548c\u82f1\u6587 WER \u6301\u7eed\u4e0b\u964d\uff1b\u975e\u6d41\u5f0f\u6a21\u5f0f\u51c6\u786e\u7387\u6700\u4f4e\u3002\u8bba\u6587\u7ed9\u51fa 400 ms \u5ef6\u8fdf\u70b9\u7684\u7ed3\u679c\uff1a\u4e2d\u6587 CER 9.24%\u3001\u82f1\u6587 WER 13.38%\uff1b\u975e\u6d41\u5f0f\u5206\u522b\u4e3a 8.03% \u548c 11.78%\uff0c\u76f8\u5bf9\u9519\u8bef\u7387\u589e\u52a0\u7ea6 15.1% \u548c 13.6%\u3002\u8fd9\u8bf4\u660e\u7cfb\u7edf\u53ef\u4ee5\u6309\u4ea7\u54c1\u9700\u6c42\u5728\u54cd\u5e94\u901f\u5ea6\u548c\u51c6\u786e\u7387\u4e4b\u95f4\u9009 operating point\u3002<\/p>\n\n\n\n

              \"\"
              \u914d\u5ef6\u8fdf-\u51c6\u786e\u7387\u66f2\u7ebf\uff0c\u6d4b\u8bd5\u786c\u4ef6\u4e3a NVIDIA H100<\/figcaption><\/figure>\n\n\n\n

              \u516b\u3001\u521b\u65b0\u70b9\u603b\u7ed3\uff1a\u771f\u6b63\u6709\u4ef7\u503c\u7684\u4e0d\u662f\u5355\u4e2a\u6a21\u5757\uff0c\u800c\u662f\u7ec4\u5408\u65b9\u5f0f<\/h2>\n\n\n\n
              1. MoE LLM-ASR\uff1a<\/strong>\u4ee5 Qwen MoE \u66ff\u6362 Dense Decoder\uff0c\u5728\u589e\u52a0\u8bed\u8a00\/\u4e16\u754c\u77e5\u8bc6\u5bb9\u91cf\u7684\u540c\u65f6\u63a7\u5236 token \u7ea7\u8ba1\u7b97\u3002<\/li>
              2. \u7edf\u4e00 instruction-controlled ASR\uff1a<\/strong>\u8bed\u8a00\u3001\u5386\u53f2\u4e0a\u4e0b\u6587\u3001\u70ed\u8bcd\u548c\u6da6\u8272\u4e0d\u662f\u5916\u6302\uff0c\u800c\u662f\u540c\u4e00\u89e3\u7801\u63a5\u53e3\u4e2d\u7684\u6761\u4ef6\u3002<\/li>
              3. \u5206\u5c42 P0\/P1 \u70ed\u8bcd\uff1a<\/strong>\u7528\u4f18\u5148\u7ea7\u548c\u8bc1\u636e\u7ea6\u675f\u540c\u65f6\u4f18\u5316\u957f\u5c3e\u53ec\u56de\u548c\u9519\u8bef\u63d2\u5165\u3002<\/li>
              4. \u539f\u751f\u5355\u6b21\u6da6\u8272\uff1a<\/strong>\u628a\u201c\u8bc6\u522b + \u53ef\u8bfb\u5316\u201d\u653e\u5230\u4e00\u6b21\u751f\u6210\u4e2d\uff0c\u51cf\u5c11\u540e\u5904\u7406\u5ef6\u8fdf\u3002<\/li>
              5. \u4e0a\u4e0b\u6587\u7f13\u5b58\u5f0f\u957f\u97f3\u9891\u5efa\u6a21\uff1a<\/strong>\u5229\u7528\u6b64\u524d\u5df2\u786e\u8ba4\u7684\u6587\u672c\u7ef4\u6301\u8de8\u53e5\u5b9e\u4f53\u4e00\u81f4\u6027\uff0c\u4e0d\u4f9d\u8d56\u4eba\u5de5\u70ed\u8bcd\u7ef4\u62a4\u3002<\/li>
              6. FunVerl-ASR \u5f02\u6b65 RL\uff1a<\/strong>\u9488\u5bf9\u97f3\u9891\u5927\u6a21\u578b\u8865\u9f50 rollout\u3001\u5956\u52b1\u548c\u8bad\u7ec3\u7684\u5f02\u6b65\u7cfb\u7edf\uff0c\u5e76\u7528\u590d\u5408 reward \u4f18\u5316\u771f\u5b9e\u4e1a\u52a1\u76ee\u6807\u3002<\/li>
              7. \u6d41\u5f0f\u4e0e\u5168\u4e0a\u4e0b\u6587\u7ed3\u5408\uff1a<\/strong>\u5b9e\u65f6\u7ed9\u53cd\u9988\uff0c\u7ed3\u675f\u540e\u518d\u7528\u5b8c\u6574\u97f3\u9891\u5237\u65b0\u6700\u7ec8\u6587\u672c\u3002<\/li><\/ol>\n\n\n\n

                \u4e5d\u3001\u5982\u4f55\u7406\u89e3\u8bba\u6587\u7684\u5b9e\u9a8c\u4f18\u52bf<\/h2>\n\n\n\n

                \u4ece\u516c\u5f00 Benchmark \u770b\uff0cQwen-Audio-3.0-ASR \u7684\u4f18\u52bf\u662f\u6574\u4f53\u5747\u8861\uff1a\u4e2d\u6587\u548c\u82f1\u6587\u5404\u6709\u591a\u4e2a\u6d4b\u8bd5\u96c6\u9886\u5148\uff0c\u591a\u8bed\u8a00\u5b8f\u5e73\u5747\u5728 GigaSpeechBench \u548c Common Voice 15 \u4e0a\u9886\u5148\u3002\u4ece\u5de5\u4e1a\u96c6\u770b\uff0c\u6700\u660e\u663e\u7684\u63d0\u5347\u96c6\u4e2d\u5728\u901a\u7528 CER\/WER \u96be\u4ee5\u8868\u8fbe\u7684\u90e8\u5206\uff1a\u65b9\u8a00\u4e00\u81f4\u6027\u3001\u884c\u4e1a\u5b9e\u4f53\u3001\u70ed\u8bcd\u3001\u8de8\u53e5\u4e0a\u4e0b\u6587\u548c\u53ef\u8bfb\u6027\u3002<\/p>\n\n\n\n

                \u66f4\u51c6\u786e\u5730\u8bf4\uff0c\u8bba\u6587\u8bc1\u660e\u7684\u4e0d\u662f\u201c\u6bcf\u4e2a\u6570\u636e\u96c6\u90fd\u7b2c\u4e00\u201d\uff0c\u800c\u662f\u201c\u4e00\u4e2a\u7edf\u4e00\u6a21\u578b\u53ef\u4ee5\u540c\u65f6\u8986\u76d6\u591a\u79cd\u751f\u4ea7\u7ea6\u675f\u201d\u3002\u8fd9\u5bf9\u4f1a\u8bae\u8bb0\u5f55\u3001\u5ba2\u670d\u8d28\u68c0\u3001\u8de8\u5883\u8bed\u97f3\u8f93\u5165\u548c\u4f01\u4e1a\u77e5\u8bc6\u573a\u666f\uff0c\u6bd4\u53ea\u770b\u5355\u4e00\u6d4b\u8bd5\u96c6\u7684 SOTA \u6392\u540d\u66f4\u6709\u610f\u4e49\u3002<\/p>\n\n\n\n

                \u5341\u3001\u5c40\u9650\u6027<\/h2>\n\n\n\n