{"id":10021352,"date":"2026-04-29T16:29:48","date_gmt":"2026-04-29T08:29:48","guid":{"rendered":"https:\/\/itmax.cn\/?p=10021352"},"modified":"2026-04-29T16:29:48","modified_gmt":"2026-04-29T08:29:48","slug":"ai-%e9%9d%a2%e8%af%95-copilot-%e5%a4%9a%e6%a8%a1%e6%80%81%e8%9e%8d%e5%90%88%ef%bc%88%e8%af%ad%e9%9f%b3%e6%96%87%e6%9c%ac%ef%bc%89%e6%b5%81%e5%bc%8f%e6%9d%83%e9%87%8d%e5%86%b3%e7%ad%96%ef%bc%9a6","status":"publish","type":"post","link":"https:\/\/itmax.cn\/?p=10021352","title":{"rendered":"AI \u9762\u8bd5 Copilot \u591a\u6a21\u6001\u878d\u5408\uff08\u8bed\u97f3+\u6587\u672c\uff09\u6d41\u5f0f\u6743\u91cd\u51b3\u7b56\uff1a6 \u6bb5\u6d41\u6c34\u7ebf + \u53cd\u538b\u8c03\u5ea6 + \u9519\u5cf0\u5b9e\u6d4b"},"content":{"rendered":"

AI \u9762\u8bd5 Copilot \u591a\u6a21\u6001\u878d\u5408\uff08\u8bed\u97f3 + \u6587\u672c\uff09\u6d41\u5f0f\u6743\u91cd\u51b3\u7b56\uff1a6 \u6bb5\u6d41\u6c34\u7ebf + \u53cd\u538b\u8c03\u5ea6 + \u9519\u5cf0\u5b9e\u6d4b<\/h2>\n

TL;DR\uff0830 \u79d2\u8bfb\u5b8c\uff09<\/h3>\n
    \n
  1. \u591a\u6a21\u6001\u878d\u5408\u4e0d\u662f\u7b80\u5355\u52a0\u6743<\/strong> \u2014\u2014 \u8bed\u97f3\uff08partial ASR + \u97f5\u5f8b + \u9759\u9ed8\uff09\u548c\u6587\u672c\uff08\u5019\u9009\u4eba\u7b80\u5386 + JD + \u77e5\u8bc6\u5e93\uff09\u5fc5\u987b\u6309\u6bb5\u52a8\u6001\u8c03\u6743\uff0c\u5355\u4e00\u56fa\u5b9a\u6743\u91cd\u5728\u9762\u8bd5\u573a\u666f Recall \u6389 18-23%\u3002<\/li>\n
  2. 6 \u6bb5\u6d41\u6c34\u7ebf<\/strong>\uff1a\u97f3\u9891\u91c7\u6837 \u2192 \u53cc\u8f68 ASR\uff08partial\/final\uff09\u2192 \u97f5\u5f8b\u7279\u5f81\uff08\u505c\u987f\/\u8bed\u901f\/\u80fd\u91cf\uff09\u2192 \u6587\u672c\u4e0a\u4e0b\u6587\u68c0\u7d22 \u2192 \u6743\u91cd\u878d\u5408\u5c42 \u2192 \u6d41\u5f0f LLM \u751f\u6210\u3002\u5404\u6bb5\u5fc5\u987b\u5e76\u884c + \u53cd\u538b\u3002<\/li>\n
  3. \u53cd\u538b\uff08back-pressure\uff09\u662f\u6838\u5fc3<\/strong>\uff1a\u4e0a\u6e38\u751f\u4ea7\u901f\u7387 > \u4e0b\u6e38\u6d88\u8d39\u65f6\uff0c\u5fc5\u987b\u4e22\u5f03 partial \u800c\u975e final\uff0c\u4e22\u9519\u4e00\u6b21\u4f1a\u8ba9 LLM \u7528\u534a\u53e5\u8bdd\u751f\u6210\u56de\u7b54\uff0c\u4f53\u611f\u62c9\u80ef\u3002<\/li>\n
  4. \u9519\u5cf0\u8c03\u5ea6<\/strong>\uff1aASR \u62a2 GPU 0\u3001LLM \u62a2 GPU 1\u3001TTS \u62a2 CPU\uff0c\u9519\u5f00\u5cf0\u503c\u540e P99 \u4ece 1.4s \u964d\u5230 620ms\uff08\u5b9e\u6d4b\u9762\u8bd5\u573a\u666f 30 \u5206\u949f\u4f1a\u8bdd\uff09\u3002<\/li>\n
  5. \u6743\u91cd\u4e0d\u662f\u9759\u6001\u8d85\u53c2<\/strong> \u2014\u2014 \u7528\u4e00\u4e2a 200KB \u7684\u5c0f\u578b logistic \u56de\u5f52\u5728\u7ebf\u5b66\uff0c\u7279\u5f81\u662f\u300c\u4e0a\u4e00\u8f6e final \u4e0e partial \u7684 BLEU + \u97f5\u5f8b\u7f6e\u4fe1\u5ea6 + \u9759\u9ed8\u65f6\u957f\u300d\uff0c\u6bcf 200ms \u51fa\u65b0\u6743\u91cd\u3002<\/li>\n
  6. \u9759\u9ed8 > 600ms<\/strong> \u65f6\u6743\u91cd\u7acb\u5373\u504f\u5411\u300c\u6587\u672c\u68c0\u7d22\u300d\uff08\u5019\u9009\u4eba\u5728\u601d\u8003\uff09\uff0c<200ms \u65f6\u504f\u5411\u300c\u8bed\u97f3 partial\u300d\uff08\u5019\u9009\u4eba\u5728\u62a2\u8bdd\uff09\uff0c\u4e2d\u95f4\u533a\u6bb5\uff08200-600ms\uff09\u53cc\u8f68\u540c\u7b49\u6743\u91cd\u3002<\/li>\n
  7. \u5b9e\u6d4b\u541e\u5410<\/strong>\uff1a4 \u8def\u5e76\u53d1\u9762\u8bd5\u5728\u5355\u53f0 H100 \u4e0a P99 \u7aef\u5230\u7aef 720ms\uff0c\u878d\u5408\u5c42\u672c\u8eab\u8017\u65f6 4.8ms\uff08\u4e0d\u542b LLM\uff09\u3002<\/li>\n<\/ol>\n

    \u4e0b\u9762\u662f\u5de5\u7a0b\u62c6\u89e3\u3002<\/p>\n

    \n

    \u4e00\u3001\u4e3a\u4ec0\u4e48”\u591a\u6a21\u6001\u52a0\u6743”\u5728\u9762\u8bd5\u573a\u666f\u53cd\u76f4\u89c9\u5730\u96be<\/h3>\n

    \u505a AI \u9762\u8bd5 Copilot \u8fd9\u4e00\u5e74\uff0c\u6700\u65e9\u671f\u6211\u4eec\u4ee5\u4e3a\u6d41\u7a0b\u5f88\u7b80\u5355\uff1a\u628a\u97f3\u9891\u4e22\u7ed9 ASR \u62ff\u6587\u672c\uff0c\u628a\u6587\u672c\u4e22\u7ed9 LLM \u62ff\u56de\u7b54\uff0c\u518d\u4e22\u7ed9 TTS \u51fa\u8bed\u97f3\u3002\u4e09\u6bb5\u7ba1\u7ebf\uff0c\u542c\u8d77\u6765 1 \u4e2a\u5de5\u7a0b\u5e08 2 \u5468\u5c31\u80fd\u51fa Demo\u3002<\/p>\n

    \u4f46\u771f\u4e0a\u7ebf\u5c31\u5d29\u4e86\u3002\u5d29\u5728\u54ea\uff1f\u5d29\u5728\u300c\u5019\u9009\u4eba\u8bf4\u8bdd\u7684\u6837\u5b50\u300d\u548c\u300c\u5019\u9009\u4eba\u60f3\u8868\u8fbe\u7684\u5185\u5bb9\u300d\u662f\u4e24\u4e2a\u72ec\u7acb\u4fe1\u53f7\uff0c\u5355\u7eaf\u4f9d\u8d56 ASR final \u6587\u672c\u4f1a\u8ba9 Copilot \u53d8\u6210\u4e00\u4e2a\u53cd\u5c04\u5f27\u6781\u6162\u7684\u590d\u8bfb\u673a\u3002<\/p>\n

    \u4e3e\u4e2a\u771f\u5b9e\u7684\u4f8b\u5b50\uff1a\u5019\u9009\u4eba\u88ab\u95ee\u5230”\u8bb2\u4e00\u4e0b\u4f60\u6700\u96be\u7684\u9879\u76ee”\uff0c\u4ed6\u5f00\u59cb\u8bf4\uff1a”\u55ef\u2026\u90a3\u4e2a\u2026\u6211\u505a\u8fc7\u4e00\u4e2a\u2026\u5443\u2026\u5206\u5e03\u5f0f\u2026\u7cfb\u7edf\u2026”\u3002ASR final \u7ed9\u6211\u4eec\u7684\u662f\u300c\u6211\u505a\u8fc7\u4e00\u4e2a\u5206\u5e03\u5f0f\u7cfb\u7edf\u300d\uff0c\u5e72\u51c0\u5229\u843d\u3002\u4f46\u771f\u5b9e\u7684\u5019\u9009\u4eba\u72b6\u6001\u662f\u7126\u8651\u3001\u5361\u58f3\u3001\u9700\u8981\u5f15\u5bfc<\/strong>\u3002\u8fd9\u4e2a\u72b6\u6001\u85cf\u5728\u9759\u9ed8\uff08>800ms \u4e09\u6bb5\u505c\u987f\uff09\u3001\u8bed\u901f\uff08\u6bcf\u79d2 1.2 \u5b57\uff0c\u8fdc\u4f4e\u4e8e\u9762\u8bd5\u5747\u503c 4.5 \u5b57\uff09\u3001\u80fd\u91cf\uff08\u58f0\u538b\u4f4e\u4e8e\u5747\u503c 12dB\uff09\u91cc\u2014\u2014\u97f5\u5f8b\u7279\u5f81\u3002<\/p>\n

    \u5982\u679c Copilot \u53ea\u770b final \u6587\u672c\uff0c\u5c31\u4f1a\u51b7\u51b0\u51b0\u5730\u56de\uff1a”\u597d\u7684\uff0c\u8bf7\u7ee7\u7eed\u8be6\u7ec6\u4ecb\u7ecd\u3002” \u4f46\u878d\u5408\u4e86\u97f5\u5f8b\u4e4b\u540e\uff0cCopilot \u5e94\u8be5\u8f93\u51fa\u7684\u662f STAR \u7ed3\u6784\u63d0\u793a\u5361\uff0c\u5e76\u4e14\u662f\u8f6f\u63d0\u793a\uff08\u5019\u9009\u4eba\u80fd\u770b\u5230\u3001\u9762\u8bd5\u5b98\u770b\u4e0d\u5230\uff09\u3002\u8fd9\u662f\u6211\u4eec\u5728 \u5373\u7b54\u4fa0<\/a>\uff08\u4e00\u6b3e AI \u9762\u8bd5\u5b9e\u65f6\u8f85\u52a9\u5de5\u5177\uff0c\u76ee\u524d\u670d\u52a1\u7684\u5de5\u7a0b\u5c97\u4f4d\u7528\u6237\u91cc 78% \u90fd\u9047\u5230\u8fc7\u8fd9\u79cd”\u5361\u58f3\u89e6\u53d1”\u573a\u666f\uff09\u91cc\u82b1\u4e86\u4e09\u4e2a\u6708\u624d\u6253\u78e8\u51fa\u7684\u7ec6\u8282\u3002<\/p>\n

    \u6240\u4ee5\u591a\u6a21\u6001\u878d\u5408\u7684\u672c\u8d28\u95ee\u9898\u662f\uff1aASR\u3001\u97f5\u5f8b\u3001\u6587\u672c\u68c0\u7d22\u8fd9\u4e09\u8def\u4fe1\u53f7\u5728\u4e0d\u540c\u65f6\u523b\u7684\u53ef\u4fe1\u5ea6\u5b8c\u5168\u4e0d\u540c\uff0c\u5fc5\u987b\u6d41\u5f0f\u52a8\u6001\u8c03\u6743<\/strong>\u3002\u56fa\u5b9a\u6743\u91cd\u5728\u9762\u8bd5\u573a\u666f\u5b9e\u6d4b Recall\uff08\u9ad8\u8d28\u91cf\u63d0\u793a\u89e6\u53d1\u7387\uff09\u6389 18-23%\u3002<\/p>\n

    \n

    \u4e8c\u3001\u6574\u4f53\u67b6\u6784\uff1a6 \u6bb5\u5e76\u884c\u6d41\u6c34\u7ebf<\/h3>\n

    \u6211\u4eec\u7684\u6d41\u6c34\u7ebf\u957f\u8fd9\u6837\uff08\u6bcf\u6bb5\u90fd\u662f\u72ec\u7acb\u7ebf\u7a0b\/\u8fdb\u7a0b\uff0c\u901a\u8fc7\u5185\u5b58 channel \u901a\u4fe1\uff09\uff1a<\/p>\n

    [1. \u97f3\u9891\u91c7\u6837 16kHz\/20ms \u5e27] \n        \u2193 (channel: audio_frame)\n[2. \u53cc\u8f68 ASR: partial (50ms) + final (block-end)]\n        \u2193 (channel: asr_partial \/ asr_final)\n[3. \u97f5\u5f8b\u7279\u5f81\u62bd\u53d6: pause\/rate\/energy\/pitch]\n        \u2193 (channel: prosody)\n[4. \u6587\u672c\u4e0a\u4e0b\u6587\u68c0\u7d22: \u7b80\u5386\u5411\u91cf + JD \u5411\u91cf + \u77e5\u8bc6\u5e93]\n        \u2193 (channel: context)\n[5. \u6743\u91cd\u878d\u5408\u5c42 (logistic + \u89c4\u5219)]\n        \u2193 (channel: weighted_signal)\n[6. \u6d41\u5f0f LLM \u751f\u6210 (token streaming)]\n        \u2193\n[7. \u5ba2\u6237\u7aef\u6e32\u67d3 \/ TTS (\u65c1\u8def)]<\/code><\/pre>\n
    \u6bcf\u6bb5\u7684\u5ef6\u8fdf\u9884\u7b97\uff08P99\uff09\uff1a<\/p>\n\n\n\n\n\n\n\n\n\n\n\n\n
    \u6bb5<\/th>\n\u4efb\u52a1<\/th>\nP99 \u9884\u7b97<\/th>\n\u5b9e\u6d4b P99<\/th>\n<\/tr>\n<\/thead>\n
    1<\/td>\n\u97f3\u9891\u91c7\u6837<\/td>\n20ms<\/td>\n21ms<\/td>\n<\/tr>\n
    2a<\/td>\npartial ASR<\/td>\n80ms<\/td>\n92ms<\/td>\n<\/tr>\n
    2b<\/td>\nfinal ASR<\/td>\n250ms<\/td>\n310ms<\/td>\n<\/tr>\n
    3<\/td>\n\u97f5\u5f8b\u62bd\u53d6<\/td>\n30ms<\/td>\n28ms<\/td>\n<\/tr>\n
    4<\/td>\n\u68c0\u7d22\uff08HNSW\uff09<\/td>\n60ms<\/td>\n54ms<\/td>\n<\/tr>\n
    5<\/td>\n\u878d\u5408\uff08\u5728\u7ebf logistic\uff09<\/td>\n8ms<\/td>\n4.8ms<\/td>\n<\/tr>\n
    6<\/td>\nLLM \u9996 token<\/td>\n280ms<\/td>\n240ms<\/td>\n<\/tr>\n
    \u7aef\u5230\u7aef<\/td>\npartial \u89e6\u53d1\u5230\u9996 token<\/td>\n800ms<\/td>\n720ms<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
    \u6ce8\u610f\u7b2c 2\u30013\u30014 \u6bb5\u662f\u5e76\u884c<\/strong>\u7684\u2014\u2014final ASR \u6ca1\u51fa\u6765\u65f6\uff0cpartial ASR + \u97f5\u5f8b + \u68c0\u7d22\u65e9\u5c31\u8dd1\u5b8c\u4e86\u3002\u878d\u5408\u5c42\u6bcf 200ms \u51fa\u4e00\u6b21\u51b3\u7b56\uff08\u4e0d\u662f\u7b49 final\uff09\u3002<\/p>\n
    \n
    \u4e09\u3001\u53cc\u8f68 ASR\uff1apartial \u548c final \u7684\u53d6\u820d<\/h3>\n
    \u5f88\u591a\u4eba\u7b2c\u4e00\u6b21\u505a\u5b9e\u65f6 ASR \u90fd\u6389\u8fc7\u8fd9\u4e2a\u5751\uff1a\u7528 Azure \/ Deepgram \/ Whisper Streaming \u62ff partial \u6587\u672c \u2192 \u76f4\u63a5\u9001\u7ed9 LLM \u2192 LLM \u7528\u534a\u53e5\u8bdd\u751f\u6210 \u2192 \u5168\u662f\u5e7b\u89c9\u3002<\/p>\n
    \u6b63\u786e\u505a\u6cd5\u662f partial \u548c final \u53cc\u8f68\u5e76\u884c\uff0c\u878d\u5408\u5c42\u51b3\u5b9a\u54ea\u4e2a\u6743\u91cd\u9ad8<\/strong>\uff1a<\/p>\n
      \n
    • partial\uff0850-80ms \u51fa\u4e00\u6b21\uff09<\/strong>\uff1a\u4f4e\u5ef6\u8fdf\uff0c\u4f46\u8bcd\u9519\u7387\uff08WER\uff09\u5e73\u5747 22%\u3002\u5728\u5019\u9009\u4eba\u62a2\u8bdd<\/strong>\uff08\u9759\u9ed8 < 200ms\uff09\u65f6\u4f18\u5148\u7528\uff0c\u56e0\u4e3a final \u8fd8\u6ca1\u51fa\u6765\u7528\u6237\u5df2\u7ecf\u8bb2\u4e0b\u4e00\u53e5\u4e86\u3002<\/li>\n
    • final\uff08block-end \u89e6\u53d1\uff0c250-350ms\uff09<\/strong>\uff1a\u9ad8\u51c6\u786e\uff0cWER 5-7%\u3002\u5728\u5019\u9009\u4eba\u6b63\u5e38\u9648\u8ff0<\/strong>\uff08\u9759\u9ed8 200-600ms\uff09\u65f6\u4f18\u5148\u7528\u3002<\/li>\n
    • \u9759\u9ed8 > 600ms<\/strong>\uff1a\u5019\u9009\u4eba\u5728\u601d\u8003\uff0c\u4e24\u4e2a ASR \u90fd\u6ca1\u65b0\u5185\u5bb9\uff0c\u6743\u91cd\u5207\u7ed9\u6587\u672c\u68c0\u7d22\uff08\u91cd\u65b0\u8bfb\u4e00\u904d JD + \u7b80\u5386\uff0c\u627e\u4e0b\u4e00\u4e2a\u5f15\u5bfc\u95ee\u9898\uff09\u3002<\/li>\n<\/ul>\n
      \u4ee3\u7801\u7684\u6838\u5fc3\u7ed3\u6784\uff08Python\uff0c\u7701\u7565\u9519\u8bef\u5904\u7406\uff09\uff1a<\/p>\n
      async def asr_dual_track(audio_stream):\n    partial_q = asyncio.Queue(maxsize=8)\n    final_q = asyncio.Queue(maxsize=4)\n    \n    async def partial_loop():\n        async for chunk in audio_stream:\n            text, conf = await asr_partial.feed(chunk)  # 50ms\n            try:\n                partial_q.put_nowait({\n                    'text': text, 'conf': conf, 'ts': time.time(), 'kind': 'partial'\n                })\n            except asyncio.QueueFull:\n                # \u53cd\u538b\uff1a\u4e22\u6700\u8001\u7684 partial\uff08\u4e0d\u662f\u6700\u65b0\uff09\n                _ = partial_q.get_nowait()\n                partial_q.put_nowait({...})\n    \n    async def final_loop():\n        async for chunk in audio_stream:\n            text, conf = await asr_final.feed(chunk)  # block-end\n            await final_q.put({'text': text, 'conf': conf, 'kind': 'final'})\n            # final \u6c38\u8fdc\u4e0d\u4e22\uff0c\u4e22\u4e86 LLM \u62ff\u4e0d\u5230\u5b8c\u6574\u53e5\u5b50\n    \n    await asyncio.gather(partial_loop(), final_loop())<\/code><\/pre>\n
      \u53cd\u538b\u7b56\u7565\u7684\u5173\u952e\u70b9<\/strong>\uff1apartial \u961f\u5217\u6ee1\u65f6\u4e22\u6700\u8001\u7684\uff08FIFO \u53cd\u5411\uff09\uff0cfinal \u961f\u5217\u6ee1\u65f6\u963b\u585e\u4e0a\u6e38<\/strong>\u2014\u2014\u4e0d\u80fd\u4e22 final\uff0c\u4e22\u4e86 LLM \u62ff\u4e0d\u5230\u5b8c\u6574\u8bed\u4e49\uff0c\u8f93\u51fa\u4f1a\u53d1\u6563\u3002<\/p>\n
      \n
      \u56db\u3001\u97f5\u5f8b\u7279\u5f81\uff1a\u90a3\u4e9b\u85cf\u5728\u505c\u987f\u91cc\u7684\u4fe1\u53f7<\/h3>\n
      ASR \u7ed9\u7684\u662f\u300c\u5b57\u300d\uff0c\u97f5\u5f8b\u7ed9\u7684\u662f\u300c\u4eba\u300d\u3002\u6211\u4eec\u62bd 4 \u7c7b\u7279\u5f81\uff1a<\/p>\n
        \n
      1. \u9759\u9ed8\u65f6\u957f\uff08pause_ms\uff09<\/strong>\uff1a\u8fde\u7eed RMS \u80fd\u91cf\u4f4e\u4e8e\u9608\u503c\u7684\u65f6\u95f4\u3002> 600ms \u8868\u793a\u601d\u8003\u4e2d\u3002<\/li>\n
      2. \u8bed\u901f\uff08rate_cps\uff09<\/strong>\uff1a\u6bcf\u79d2\u8f93\u51fa\u5b57\u6570\u3002\u9762\u8bd5\u573a\u666f\u5747\u503c 4.5\uff0c\u5019\u9009\u4eba\u7d27\u5f20\u65f6\u964d\u5230 1.2-2.0\u3002<\/li>\n
      3. \u80fd\u91cf\u65b9\u5dee\uff08energy_var\uff09<\/strong>\uff1a\u80fd\u91cf\u7684\u6ed1\u52a8\u65b9\u5dee\u3002\u4f4e\u65b9\u5dee = \u5e73\u94fa\u76f4\u53d9\uff0c\u9ad8\u65b9\u5dee = \u60c5\u611f\u6295\u5165\u3002<\/li>\n
      4. \u57fa\u9891\u65b9\u5dee\uff08pitch_var\uff09<\/strong>\uff1aF0 \u7684\u6ed1\u52a8\u65b9\u5dee\u3002\u4f4e = \u5355\u8c03\uff0c\u53ef\u80fd\u662f\u80cc\u8bf5\uff1b\u9ad8 = \u6709\u601d\u8003\u3002<\/li>\n<\/ol>\n
        \u5b9e\u73b0\u4e0a\u7528 librosa + \u81ea\u5df1\u5199\u7684 RMS\/F0 \u6ed1\u7a97\uff0830ms hop\uff09\u3002\u6574\u6bb5 28ms \u8dd1\u5b8c\uff0c\u8ddf partial ASR \u540c wall-clock\u3002<\/p>\n
        def prosody_features(audio_chunk_pcm16):\n    rms = librosa.feature.rms(y=audio_chunk_pcm16, hop_length=480)[0]\n    pitch, _ = librosa.piptrack(y=audio_chunk_pcm16, sr=16000, hop_length=480)\n    pause_ms = compute_pause(rms, threshold=0.005)  # \u81ea\u5b9a\u4e49\u9608\u503c\n    rate = words_in_chunk \/ chunk_duration\n    return {\n        'pause_ms': pause_ms,\n        'rate_cps': rate,\n        'energy_var': float(np.var(rms)),\n        'pitch_var': float(np.var(pitch[pitch > 0])) if (pitch > 0).any() else 0.0\n    }<\/code><\/pre>\n
        \u53cd\u76f4\u89c9\u70b9<\/strong>\uff1a\u80fd\u91cf\u65b9\u5dee\u548c\u57fa\u9891\u65b9\u5dee\u662f\u914d\u5bf9<\/strong>\u7684\u3002\u65b9\u5dee\u90fd\u4f4e = \u5019\u9009\u4eba\u5728\u80cc\u8bf5\u7b54\u6848\uff08\u5e94\u8be5\u63d0\u793a Copilot \u6539\u7528\u8ffd\u95ee\uff09\uff1b\u65b9\u5dee\u90fd\u9ad8 = \u5019\u9009\u4eba\u6fc0\u52a8 \/ \u7d27\u5f20\uff08\u5e94\u8be5\u63d0\u793a\u7a33\u5b9a\u8bdd\u672f\uff09\uff1b\u65b9\u5dee\u4e00\u9ad8\u4e00\u4f4e = \u4fe1\u53f7\u566a\u58f0\u5927\uff0c\u6743\u91cd\u964d\u5230 0.3\u3002<\/p>\n
        \n
        \u4e94\u3001\u6743\u91cd\u878d\u5408\u5c42\uff1a\u5728\u7ebf logistic + \u4e09\u6761\u786c\u89c4\u5219<\/h3>\n
        \u878d\u5408\u5c42\u662f\u6574\u4e2a\u7cfb\u7edf\u7684\u5927\u8111\u3002\u5b83\u6bcf 200ms \u6536\u5230\u4e00\u6b21 {partial, final, prosody, context}<\/code> \u56db\u8def\u4fe1\u53f7\uff0c\u8f93\u51fa\u4e00\u4e2a {w_partial, w_final, w_prosody, w_context}<\/code> \u6743\u91cd\u5411\u91cf\uff0c\u5582\u7ed9 LLM\u3002<\/p>\n
        \u6211\u4eec\u8bd5\u8fc7\u4e09\u79cd\u5b9e\u73b0\uff1a<\/p>\n\n\n\n\n\n\n\n
        \u65b9\u6848<\/th>\n\u8bad\u7ec3\u6210\u672c<\/th>\n\u63a8\u7406\u5ef6\u8fdf<\/th>\n\u4e0a\u7ebf\u6548\u679c<\/th>\n<\/tr>\n<\/thead>\n
        \u56fa\u5b9a\u6743\u91cd<\/td>\n0<\/td>\n0.1ms<\/td>\nRecall -23%<\/td>\n<\/tr>\n
        \u89c4\u5219\u5f15\u64ce\uff0830 \u6761 if-else\uff09<\/td>\n0<\/td>\n1.2ms<\/td>\nRecall -8%\uff0c\u4f46\u8c03\u8bd5\u96be<\/td>\n<\/tr>\n
        \u5728\u7ebf logistic \u56de\u5f52\uff08200KB\uff09<\/td>\n4h \u6807\u6ce8 + 5min \u8bad\u7ec3<\/td>\n4.8ms<\/td>\n\u57fa\u7ebf<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
        \u6700\u540e\u9009\u7684\u662f logistic + 3 \u6761\u786c\u89c4\u5219\u8986\u76d6<\/strong>\uff1a<\/p>\n
        def fuse_weights(partial, final, prosody, context):\n    # \u7279\u5f81\u5411\u91cf\n    feats = np.array([\n        partial['conf'],\n        final['conf'] if final else 0,\n        prosody['pause_ms'] \/ 1000,\n        prosody['rate_cps'] \/ 5,\n        prosody['energy_var'],\n        bleu(partial['text'], final['text']) if final else 0,\n        context['retrieval_score'],\n        time_since_last_final()\n    ])\n    # logistic 4 \u8def\u6743\u91cd\n    w = sigmoid(W @ feats + b)  # W: 4x8, b: 4\n    \n    # \u786c\u89c4\u5219\u8986\u76d6\n    if prosody['pause_ms'] > 600:\n        w = [0.1, 0.2, 0.2, 0.5]  # \u9759\u9ed8\uff1a\u504f\u5411\u68c0\u7d22\n    if prosody['pause_ms'] < 200 and partial['conf'] > 0.7:\n        w = [0.6, 0.1, 0.2, 0.1]  # \u62a2\u8bdd\uff1a\u504f\u5411 partial\n    if final and final['ts'] - partial['ts'] < 50:\n        w = [0.0, 0.7, 0.2, 0.1]  # final \u521a\u5230\uff1a\u5b8c\u5168\u7528 final\n    \n    return w \/ w.sum()<\/code><\/pre>\n
        \u4e3a\u4ec0\u4e48\u7559\u786c\u89c4\u5219<\/strong>\uff1f\u56e0\u4e3a logistic \u662f\u7edf\u8ba1\u6a21\u578b\uff0c\u957f\u5c3e\u573a\u666f\u8bad\u7ec3\u6570\u636e\u4e0d\u591f\uff08\u5019\u9009\u4eba\u54fd\u54bd\u3001\u7b11\u58f0\u3001\u54b3\u55fd\uff09\u3002\u786c\u89c4\u5219\u662f\u515c\u5e95\uff0c\u5f53\u6a21\u578b\u7f6e\u4fe1\u5ea6\u4f4e\u4e8e 0.5 \u65f6\u76f4\u63a5\u8d70\u89c4\u5219\u3002<\/p>\n
        \n
        \u516d\u3001\u53cd\u538b\u8c03\u5ea6\uff1a\u9519\u5cf0\u8ba9 P99 \u4ece 1.4s \u964d\u5230 720ms<\/h3>\n
        \u6700\u65e9\u6240\u6709\u6bb5\u90fd\u8dd1\u5728\u540c\u4e00\u4e2a CPU\/GPU \u4e0a\uff0cP99 \u7ecf\u5e38\u98d9\u5230 1.4-1.8s\u3002\u5256\u6790\u540e\u53d1\u73b0\uff1aASR \u548c LLM \u5728 GPU \u4e0a\u62a2\u8d44\u6e90\uff0cTTS \u53c8\u8ddf\u68c0\u7d22\u5728 CPU \u4e0a\u62a2\u3002<\/p>\n
        \u9519\u5cf0\u8c03\u5ea6\uff1a<\/p>\n\n\n\n\n\n\n\n\n
        \u8d44\u6e90<\/th>\n\u4e3b\u4efb\u52a1<\/th>\n\u6b21\u4efb\u52a1<\/th>\n\u89e6\u53d1\u6761\u4ef6<\/th>\n<\/tr>\n<\/thead>\n
        GPU 0 (H100 SXM)<\/td>\nASR (Whisper-medium)<\/td>\n\u97f5\u5f8b F0 \u62bd\u53d6<\/td>\nASR idle<\/td>\n<\/tr>\n
        GPU 1 (H100 SXM)<\/td>\nLLM (Qwen-72B \u91cf\u5316)<\/td>\n–<\/td>\n\u72ec\u5360<\/td>\n<\/tr>\n
        CPU 0-15<\/td>\n\u68c0\u7d22 + \u878d\u5408<\/td>\nTTS \u7f16\u7801<\/td>\nTTS \u4ec5 generate \u65f6\u5207\u5165<\/td>\n<\/tr>\n
        CPU 16-31<\/td>\n\u7f51\u7edc IO + WS push<\/td>\n\u5ba2\u6237\u7aef\u7f16\u7801<\/td>\n–<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
        \u5b9e\u6d4b 4 \u8def\u5e76\u53d1\uff084 \u4e2a\u5019\u9009\u4eba\u540c\u65f6\u5728\u9762\u8bd5\uff09\u4e0b\uff1a<\/p>\n\n\n\n\n\n\n\n
        \u914d\u7f6e<\/th>\nP50<\/th>\nP99<\/th>\n\u5907\u6ce8<\/th>\n<\/tr>\n<\/thead>\n
        \u5355 GPU \u5168\u538b<\/td>\n480ms<\/td>\n1420ms<\/td>\nLLM \u62a2 ASR \u8d44\u6e90<\/td>\n<\/tr>\n
        \u53cc GPU \u4e0d\u9519\u5cf0<\/td>\n380ms<\/td>\n1080ms<\/td>\nTTS \u5361 CPU<\/td>\n<\/tr>\n
        \u53cc GPU + \u9519\u5cf0<\/td>\n320ms<\/td>\n720ms<\/td>\n\u4e0a\u7ebf\u7248<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
        \u53cd\u538b\u5728\u9519\u5cf0\u91cc\u4e5f\u6709\u4f5c\u7528<\/strong>\uff1a\u5f53 GPU 1 LLM \u961f\u5217\u957f\u5ea6 > 3 \u65f6\uff0c\u878d\u5408\u5c42\u4e3b\u52a8\u964d\u4f4e partial \u9891\u7387\uff08\u4ece 50ms \u4e00\u6b21\u964d\u5230 100ms \u4e00\u6b21\uff09\uff0c\u8ba9 LLM \u6709\u65f6\u95f4\u6d88\u5316\u3002\u8fd9\u662f\u8f6f\u53cd\u538b\uff08\u4e0d\u4e22\u6570\u636e\uff0c\u964d\u901f\u7387\uff09\u3002<\/p>\n
        \n
        \u4e03\u3001\u4e09\u4e2a\u751f\u4ea7\u4e8b\u6545\u590d\u76d8<\/h3>\n
        \u4e8b\u6545 1\uff1apartial \u548c final \u65f6\u5e8f\u5012\u6302<\/h4>\n
        \u67d0\u6b21 final ASR \u56e0\u4e3a GPU \u62a2\u5360\u5ef6\u8fdf\u5230 700ms \u624d\u51fa\u6765\uff0cpartial \u65e9\u5c31\u63a8\u8fdb\u5230\u4e0b\u4e00\u53e5\u4e86\u3002\u878d\u5408\u5c42\u62ff\u5230\u65e7 final + \u65b0 partial \u7684\u7ec4\u5408\uff0cBLEU \u7b97\u51fa\u6765\u662f 0.03\uff08\u5b8c\u5168\u4e0d\u5339\u914d\uff09\uff0clogistic \u7ed9\u51fa\u6781\u7aef\u6743\u91cd [0.95, 0.0, 0.05, 0.0]<\/code>\uff0cLLM \u7528\u9519\u4f4d\u7684 partial \u751f\u6210\u4e86\u7b54\u975e\u6240\u95ee\u7684\u63d0\u793a\u3002<\/p>\n
        \u4fee\u590d\uff1a\u6240\u6709\u4fe1\u53f7\u5fc5\u987b\u5e26 ts<\/code> \u65f6\u95f4\u6233\uff0c\u878d\u5408\u5c42\u53ea\u63a5\u53d7\u65f6\u95f4\u7a97\u53e3 \u00b1200ms \u5185<\/strong>\u7684\u4fe1\u53f7\u7ec4\u5408\uff0c\u8d85\u7a97\u7684 final \u76f4\u63a5\u4e22\u5f03\uff0c\u7b49\u4e0b\u4e00\u7ec4\u3002<\/p>\n
        \u4e8b\u6545 2\uff1a\u97f5\u5f8b\u9759\u9ed8\u628a\u5019\u9009\u4eba\u5413\u5230<\/h4>\n
        \u5019\u9009\u4eba\u8bf4\u4e00\u534a\uff0c\u505c\u987f 700ms \u5728\u60f3\u8bcd\u3002\u7cfb\u7edf\u5224\u5b9a\u300c\u601d\u8003\u4e2d\u300d\u7acb\u5373\u63a8\u9001 STAR \u63d0\u793a\u5361\u3002\u4f46\u5019\u9009\u4eba\u90a3 700ms \u4e0d\u662f\u60f3\u8bcd\uff0c\u662f\u559d\u6c34<\/strong>\u3002\u63d0\u793a\u5361\u7a81\u7136\u5f39\u51fa\u53cd\u800c\u6253\u65ad\u4e86\u4ed6\u7684\u601d\u8def\uff0c\u4ed6\u540e\u9762\u56de\u7b54\u9519\u4e71\u4e86\u3002<\/p>\n
        \u4fee\u590d\uff1a\u52a0\u5165\u300c\u80fd\u91cf+\u9759\u9ed8\u300d\u8054\u5408\u5224\u5b9a\u3002\u9759\u9ed8 > 600ms \u4f46\u524d 1s \u80fd\u91cf\u65b9\u5dee < 0.001<\/strong>\uff08\u8bf4\u8bdd\u505c\u4e86\u4f46\u6ca1\u6709\u601d\u8003\u7684\u8bed\u8c03\u8d77\u4f0f\uff09= \u7269\u7406\u505c\u987f\uff08\u559d\u6c34\/\u54b3\u55fd\uff09\uff0c\u4e0d\u89e6\u53d1\u63d0\u793a\u3002<\/p>\n
        \u4e8b\u6545 3\uff1a\u878d\u5408\u6743\u91cd\u5728\u957f\u4f1a\u8bdd\u6f02\u79fb<\/h4>\n
        30 \u5206\u949f\u4f1a\u8bdd\u540e\u671f\uff0clogistic \u6743\u91cd\u6162\u6162\u504f\u5411 w_partial<\/code>\uff0c\u56e0\u4e3a\u540e\u671f\u5019\u9009\u4eba\u8bb2\u5f97\u8d8a\u6765\u8d8a\u5feb\u3001\u8d8a\u6765\u8d8a\u81ea\u4fe1\uff0cpartial confidence \u90fd\u9ad8\u3002\u4f46\u5176\u5b9e\u5019\u9009\u4eba\u662f\u5728\u81ea\u6211\u91cd\u590d\uff08\u9648\u8ff0\u7ed3\u675f\u9636\u6bb5\uff09\uff0c\u5e94\u8be5\u51cf\u5c11 Copilot \u5e72\u9884\u3002<\/p>\n
        \u4fee\u590d\uff1a\u6bcf 5 \u5206\u949f\u505a\u4e00\u6b21\u300c\u6743\u91cd\u91cd\u7f6e\u300d\uff0c\u628a logistic \u8f93\u51fa\u505a\u6307\u6570\u79fb\u52a8\u5e73\u5747\uff08EMA, alpha=0.7\uff09\uff0c\u6291\u5236\u5355\u6b21\u6781\u7aef\u503c\u3002<\/p>\n
        \n
        \u516b\u3001\u53ef\u91cf\u5316\u7684\u76d1\u63a7\u57cb\u70b9<\/h3>\n
        \u6bcf\u8def\u4fe1\u53f7\u57cb\u4e00\u7ec4\u6307\u6807\uff0c\u6240\u6709\u6307\u6807\u8d70 OpenTelemetry \u2192 Prometheus\uff1a<\/p>\n\n\n\n\n\n\n\n\n\n\n\n
        \u6307\u6807\u540d<\/th>\n\u542b\u4e49<\/th>\n\u62a5\u8b66\u9608\u503c<\/th>\n<\/tr>\n<\/thead>\n
        fusion.weight_drift<\/code><\/td>\n\u6743\u91cd EMA \u4e0e\u77ac\u65f6\u504f\u5dee<\/td>\n> 0.4 \u6301\u7eed 1min<\/td>\n<\/tr>\n
        asr.partial.lag<\/code><\/td>\npartial \u961f\u5217\u5806\u79ef<\/td>\n> 5 \u5e27<\/td>\n<\/tr>\n
        asr.final.timeout<\/code><\/td>\nfinal \u8d85 500ms<\/td>\n\u51fa\u73b0\u5373\u62a5<\/td>\n<\/tr>\n
        prosody.feat.nan<\/code><\/td>\n\u97f5\u5f8b\u7279\u5f81 NaN<\/td>\n\u51fa\u73b0\u5373\u62a5\uff08\u9664\u96f6\uff09<\/td>\n<\/tr>\n
        llm.ttft<\/code><\/td>\nLLM \u9996 token \u65f6\u95f4<\/td>\nP99 > 350ms<\/td>\n<\/tr>\n
        e2e.partial_to_token<\/code><\/td>\n\u7aef\u5230\u7aef\u9996 token \u5ef6\u8fdf<\/td>\nP99 > 800ms<\/td>\n<\/tr>\n
        gpu.util.mismatch<\/code><\/td>\nGPU 0\/1 \u5229\u7528\u7387\u5dee<\/td>\n\u6301\u7eed > 30%<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
        \u9519\u5cf0\u8c03\u5ea6\u6709\u6548\u7684\u5224\u65ad\u6807\u51c6\uff1agpu.util.mismatch<\/code> \u5e94\u8be5\u56f4\u7ed5 5-10% \u6ce2\u52a8\uff0c\u800c\u4e0d\u662f 30-50%\uff08\u9519\u5cf0\u5931\u8d25 = \u62a2\u5360\uff09\u3002<\/p>\n
        \n
        \u4e5d\u3001\u7ed9\u505a\u76f8\u4f3c\u7cfb\u7edf\u7684\u51e0\u6761\u5efa\u8bae<\/h3>\n
          \n
        1. \u4e0d\u8981\u4ece\u300c\u6700\u4f73\u6a21\u578b\u300d\u5f00\u59cb<\/strong>\u3002\u5148\u628a\u6d41\u6c34\u7ebf\u9aa8\u67b6\u8dd1\u901a\uff0c\u6bcf\u6bb5\u90fd\u7528\u6700\u5c0f\u53ef\u7528\u6a21\u578b\uff08Whisper-tiny + 7B LLM\uff09\uff0c\u8ba9\u5168\u94fe\u8def\u6709\u5b9e\u65f6\u56de\u653e\u3002<\/li>\n
        2. \u76d1\u63a7 > \u6a21\u578b<\/strong>\u3002\u878d\u5408\u5c42\u662f\u9ed1\u7bb1\uff0c\u6ca1\u6709 OpenTelemetry trace \u4f60\u6839\u672c\u4e0d\u77e5\u9053\u54ea\u6bb5\u6162\u4e86\u3002\u63d0\u524d\u57cb\u70b9\u6bd4\u540e\u671f\u8865\u57cb\u6210\u672c\u4f4e 5 \u500d\u3002<\/li>\n
        3. \u53cd\u538b\u662f\u5fc5\u9700\uff0c\u4e0d\u662f\u4f18\u5316<\/strong>\u3002\u751f\u4ea7\u7cfb\u7edf\u5fc5\u6709\u7a81\u53d1\u6d41\u91cf\uff08\u5019\u9009\u4eba\u6fc0\u52a8\u3001\u9762\u8bd5\u5b98\u591a\u6b21\u6253\u65ad\uff09\uff0c\u961f\u5217\u963b\u585e\u4e0d\u6b7b\u4e5f\u5f97\u5361\u3002<\/li>\n
        4. \u97f5\u5f8b\u7279\u5f81\u8981\u505a chunk \u7ea7\u522b\u5f52\u4e00\u5316<\/strong>\u3002\u7edd\u5bf9\u503c\u566a\u58f0\u5927\uff0c\u76f8\u5bf9\u503c\uff08\u8fd9\u4e00\u6bb5\u76f8\u5bf9\u524d 3s \u7684\u65b9\u5dee\uff09\u7a33\u5b9a\u3002<\/li>\n
        5. logistic \u6bd4 deep model \u597d\u7ef4\u62a4<\/strong>\u3002200KB \u6a21\u578b\uff0c\u53ef\u8bfb\u6743\u91cd\uff0c\u8c03\u53c2\u5feb\u3002\u795e\u7ecf\u7f51\u7edc\u5728\u878d\u5408\u5c42\u8fd9\u79cd\u5c0f\u95ee\u9898\u4e0a\u662f\u8fc7\u5ea6\u5de5\u7a0b\u3002<\/li>\n
        6. \u9519\u5cf0\u8c03\u5ea6\u770b NUMA<\/strong>\u3002\u591a GPU \u7cfb\u7edf\u6ce8\u610f PCIe lane \u5206\u5e03\uff0c\u8de8 socket \u8c03\u5ea6\u4f1a\u8ba9\u6570\u636e\u642c\u8fd0\u5403 50ms+\u3002<\/li>\n
        7. \u6743\u91cd\u4e0d\u8981\u5728\u524d 30s \u8bad\u7ec3<\/strong>\u3002\u4f1a\u8bdd\u5f00\u5934\u566a\u58f0\u5927\uff0c\u70ed\u542f\u52a8\u7528\u5168\u5c40\u5148\u9a8c\u6743\u91cd\uff0c30s \u540e\u518d\u5207\u5230 in-session logistic\u3002<\/li>\n<\/ol>\n
          \n
          \u5e38\u89c1\u95ee\u9898\uff08FAQ\uff09<\/h3>\n
          Q1: \u4e3a\u4ec0\u4e48\u4e0d\u7528 end-to-end \u7684\u591a\u6a21\u6001\u5927\u6a21\u578b\uff08GPT-4o realtime\uff09\uff1f<\/strong>
          A: \u4e09\u4e2a\u539f\u56e0\u3002\u4e00\u662f\u5ef6\u8fdf\uff0crealtime API \u7684 P99 \u5728\u8de8\u5883\u7f51\u7edc\u4e0b\u5230 1.5s \u8d77\u6b65\uff1b\u4e8c\u662f\u6210\u672c\uff0c\u6bcf\u5206\u949f $0.06\uff0c\u957f\u4f1a\u8bdd\u4e0d\u53ef\u63a7\uff1b\u4e09\u662f\u53ef\u63a7\u6027\uff0c\u878d\u5408\u6743\u91cd\u662f\u9ed1\u76d2\uff0c\u65e0\u6cd5\u6309\u9762\u8bd5\u573a\u666f\u5355\u72ec\u8c03\u3002\u81ea\u7814\u6d41\u6c34\u7ebf\u867d\u7136\u5de5\u7a0b\u91cf\u5927\uff0c\u4f46\u6bcf\u6bb5\u53ef\u72ec\u7acb\u4f18\u5316\uff0c\u5355\u5206\u949f\u6210\u672c $0.008\u3002<\/p>\n
          Q2: partial ASR \u7528\u4ec0\u4e48\u6a21\u578b\uff1f<\/strong>
          A: Whisper-medium\uff08\u91cf\u5316\u5230 INT8\uff0c\u7ea6 800MB\uff09\u914d\u5408 streaming wrapper\uff08faster-whisper + \u81ea\u5df1\u5199\u7684 chunk \u8c03\u5ea6\uff09\u3002\u5f00\u6e90\u751f\u6001\u91cc\u76ee\u524d\u6700\u7a33\u5b9a\u7684\u65b9\u6848\u3002Whisper-large-v3 \u51c6\u786e\u5ea6\u9ad8 3-5%\uff0c\u4f46\u5ef6\u8fdf\u7ffb\u500d\uff0c\u4e0d\u9002\u5408 partial\u3002<\/p>\n
          Q3: \u878d\u5408\u5c42\u4e3a\u4ec0\u4e48\u4e0d\u7528 transformer\uff1f<\/strong>
          A: \u8bd5\u8fc7 4 \u5c42 transformer encoder\uff0c\u63a8\u7406\u5ef6\u8fdf 24ms\uff0c\u6bd4 logistic \u6162 5 \u500d\uff0c\u51c6\u786e\u5ea6\u53ea\u9ad8 1.8%\uff08A\/B \u6d4b\u8bd5\uff09\u3002\u4e0d\u5212\u7b97\u3002\u878d\u5408\u5c42\u7684\u8f93\u5165\u7ef4\u5ea6\u53ea\u6709 8-12 \u7ef4\uff0clogistic \u5df2\u7ecf\u662f\u6027\u4ef7\u6bd4\u4e0a\u9650\u3002<\/p>\n
          Q4: \u97f5\u5f8b\u7279\u5f81\u600e\u4e48\u6807\u6ce8\u8bad\u7ec3\u6570\u636e\uff1f<\/strong>
          A: 4 \u5c0f\u65f6\u4eba\u5de5\u6807\u6ce8 + \u81ea\u84b8\u998f\u3002\u7528 GPT-4 \u5f53\u4f2a\u6807\u6ce8\u5668\uff0c\u7ed9\u5b9a\u6587\u672c + \u97f5\u5f8b\u5411\u91cf\uff0c\u8ba9\u5b83\u5224\u65ad\u300c\u662f\u5426\u5728\u601d\u8003\u300d\u300c\u662f\u5426\u62a2\u8bdd\u300d\u300c\u662f\u5426\u7269\u7406\u505c\u987f\u300d\u30023 \u4e2a\u6807\u7b7e\u7684 Cohen’s Kappa 0.78\uff0c\u591f\u7528\u3002<\/p>\n
          Q5: \u53cd\u538b\u961f\u5217\u5927\u5c0f\u600e\u4e48\u5b9a\uff1f<\/strong>
          A: partial \u961f\u5217 = 8 \u5e27\uff08160ms \u7f13\u51b2\uff09\uff0cfinal \u961f\u5217 = 4 \u4e2a block\uff08\u7ea6 1s \u7f13\u51b2\uff09\u3002LLM \u8f93\u5165\u961f\u5217 = 3\u3002\u539f\u5219\u662f\u4e0a\u6e38\u6c38\u8fdc\u6bd4\u4e0b\u6e38\u77ed\u4e00\u4e9b\uff0c\u8ba9\u53cd\u538b\u80fd\u66f4\u5feb\u611f\u77e5\u5230\u4e0b\u6e38\u5835\u585e\u3002<\/p>\n
          Q6: \u957f\u4f1a\u8bdd\uff08>30min\uff09\u6743\u91cd\u6f02\u79fb\u600e\u4e48\u5904\u7406\uff1f<\/strong>
          A: 5 \u5206\u949f EMA \u91cd\u7f6e + \u6bcf 10 \u5206\u949f\u8dd1\u4e00\u6b21 in-session \u5fae\u8c03\uff08\u7528 LoRA \u5fae\u8c03 logistic\uff0c120ms \u5b8c\u6210\uff09\u3002\u540c\u65f6\u76d1\u63a7 fusion.weight_drift<\/code> \u6307\u6807\uff0c\u8d85\u9608\u503c\u4eba\u5de5\u4ecb\u5165\u3002<\/p>\n
          Q7: \u8fd9\u5957\u67b6\u6784\u9002\u7528\u4e8e\u5ba2\u670d \/ \u6559\u80b2 \/ \u533b\u7597 Copilot \u5417\uff1f<\/strong>
          A: \u6d41\u6c34\u7ebf\u9aa8\u67b6\u9002\u7528\uff0c\u4f46\u878d\u5408\u6743\u91cd\u9700\u8981\u91cd\u8bad\u3002\u5ba2\u670d\u573a\u666f\u9759\u9ed8\u89e6\u53d1\u6743\u91cd\u5e94\u8be5\u504f\u5411 FAQ \u68c0\u7d22\uff1b\u6559\u80b2\u573a\u666f\u5bf9\u97f5\u5f8b\u7f6e\u4fe1\u5ea6\u8981\u6c42\u66f4\u4f4e\uff08\u5b66\u751f\u8868\u8fbe\u672c\u8eab\u5c31\u7d27\u5f20\uff09\uff1b\u533b\u7597\u573a\u666f\u8981\u52a0\u672f\u8bed\u6821\u9a8c\u5c42\u3002\u5efa\u8bae\u9aa8\u67b6\u590d\u7528\uff0c\u878d\u5408\u5c42 + \u68c0\u7d22\u5c42\u5b9a\u5236\u3002<\/p>\n
          \n
          \u6536\u5c3e<\/h3>\n
          \u591a\u6a21\u6001\u878d\u5408\u5728\u9762\u8bd5 Copilot \u573a\u666f\u6ca1\u6709\u94f6\u5f39\u3002\u6211\u4eec\u82b1\u4e86\u4e09\u4e2a\u6708\u624d\u628a P99 \u4ece 1.4s \u538b\u5230 720ms\uff0c\u6838\u5fc3\u5fc3\u5f97\u662f\uff1a\u522b\u60f3\u7740\u4e00\u6b21\u505a\u5bf9\uff0c\u628a\u6d41\u6c34\u7ebf\u9aa8\u67b6\u5148\u8dd1\u8d77\u6765\uff0c\u6bcf\u6bb5\u72ec\u7acb\u4f18\u5316\u3001\u72ec\u7acb\u76d1\u63a7\uff0c\u53cd\u538b\u548c\u9519\u5cf0\u662f\u5de5\u7a0b\u7eaa\u5f8b\u4e0d\u662f\u7b97\u6cd5<\/strong>\u3002<\/p>\n
          \u5982\u679c\u4f60\u4e5f\u5728\u505a\u5b9e\u65f6\u591a\u6a21\u6001\u7cfb\u7edf\uff0c\u6b22\u8fce\u8bc4\u8bba\u533a\u4ea4\u6d41\u8e29\u5751\u7ec6\u8282\u3002<\/p>\n
          \u6807\u7b7e\u5efa\u8bae\uff1a\u4eba\u5de5\u667a\u80fd\u3001\u6027\u80fd\u4f18\u5316\u3001python\u3001\u524d\u7aef\u3001\u9762\u8bd5<\/p><\/blockquote>\n
          https:\/\/segmentfault.com\/a\/1190000047737137<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"
          AI \u9762\u8bd5 Copilot \u591a\u6a21\u6001\u878d\u5408\uff08\u8bed\u97f3 + \u6587\u672c\uff09\u6d41\u5f0f\u6743\u91cd\u51b3\u7b56\uff1a6 \u6bb5\u6d41\u6c34\u7ebf + \u53cd\u538b\u8c03\u5ea6 + \u9519\u5cf0\u5b9e\u6d4b […]<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_jetpack_memberships_contains_paid_content":false,"footnotes":""},"categories":[4],"tags":[],"class_list":["post-10021352","post","type-post","status-publish","format-standard","hentry","category-share"],"jetpack_sharing_enabled":true,"jetpack_featured_media_url":"","_links":{"self":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts\/10021352","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=10021352"}],"version-history":[{"count":0,"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts\/10021352\/revisions"}],"wp:attachment":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=10021352"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=10021352"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=10021352"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}