{"id":31259,"date":"2026-07-05T07:11:34","date_gmt":"2026-07-04T23:11:34","guid":{"rendered":"http:\/\/139.9.1.231\/?p=31259"},"modified":"2026-07-05T07:11:37","modified_gmt":"2026-07-04T23:11:37","slug":"open-source-asr-optimization-guide","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2026\/07\/05\/open-source-asr-optimization-guide\/","title":{"rendered":"\u5f00\u6e90ASR\u4f18\u5316\u6307\u5357"},"content":{"rendered":"\n

\u6587\u7ae0\u6765\u6e90\uff1a\u6bcf\u5468\u4e00\u4e2a\u5927\u6a21\u578b\u5e94\u7528<\/a><\/strong><\/p>\n\n\n\n

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

Whisper\u3001SenseVoice\u3001sherpa-onnx\u2026\u2026\u5f00\u6e90 ASR \u9009\u62e9\u8d8a\u6765\u8d8a\u591a\uff0c\u4f46\u771f\u6b63\u4e0a\u7ebf\u540e\uff0c\u4f60\u5f88\u53ef\u80fd\u9047\u5230\u4e24\u4e2a\u7ecf\u5178\u62b1\u6028\uff1a\u300c\u4e13\u6709\u540d\u8bcd\u8001\u8bfb\u9519\u300d\u548c\u300c\u5b9e\u65f6\u8ddf\u4e0d\u4e0a\u8bf4\u8bdd\u300d\u3002\u672c\u6587\u628a\u4f18\u5316\u624b\u6bb5\u6574\u7406\u6210\u53ef\u6267\u884c\u7684\u51b3\u7b56\u6811\u4e0e\u5206\u5c42\u673a\u5236\uff0c\u5e2e\u4f60\u5c11\u8e29\u5751\u3001\u5c11\u8bd5\u9519\u3002<\/p>\n\n\n\n

\u7ed3\u8bba\u5148\u884c\uff1a\u300c\u4e0d\u591f\u51c6\u300d\u548c\u300c\u4e0d\u591f\u5feb\u300d\u662f\u4e24\u6761\u4f18\u5316\u8def\u7ebf\uff0c\u4f46\u5e95\u5c42\u673a\u5236\u6709\u91cd\u53e0\u3002\u51c6\u786e\u7387\u4f18\u5148\u4ece\u97f3\u9891\u8d28\u91cf\u3001\u70ed\u8bcd\u3001\u8bed\u8a00\u6a21\u578b\u3001\u6362\u4e2d\u6587\u57fa\u5ea7\u5230\u9886\u57df\u5fae\u8c03\uff1b\u901f\u5ea6\u4f18\u5148\u4ece\u5c0f\u6a21\u578b\u3001INT8 \u91cf\u5316\u3001\u5408\u9002\u8fd0\u884c\u65f6\uff08\u6d41\u5f0f\u7528 sherpa-onnx\uff0c\u79bb\u7ebf\u7528 faster-whisper\uff09\u5165\u624b\u3002\u4e8c\u8005\u901a\u5e38\u4e92\u76f8\u62c9\u626f\u2014\u2014\u6ca1\u6709\u300c\u53c8\u51c6\u53c8\u5feb\u53c8\u514d\u8d39\u300d\u7684\u4e07\u80fd\u65b9\u6848\u3002<\/strong><\/p>\n\n\n\n

\u4e00\u3001\u5148\u5206\u6e05\uff1a\u4f60\u8981\u4f18\u5316\u7684\u662f\u54ea\u4e00\u7c7b\u95ee\u9898<\/strong><\/p>\n\n\n\n

\u5f88\u591a\u56e2\u961f\u4e00\u4e0a\u6765\u5c31\u6362\u66f4\u5927\u7684\u6a21\u578b\uff0c\u7ed3\u679c\u5ef6\u8fdf\u7ffb\u500d\u3001GPU \u6210\u672c\u98d9\u5347\uff0c\u51c6\u786e\u7387\u5374\u53ea\u6da8\u4e00\u70b9\u70b9\u3002<\/p>\n\n\n\n

\u5176\u5b9e ASR \u7684\u300c\u4e0d\u51c6\u300d\u548c\u300c\u4e0d\u5feb\u300d\uff0c\u80cc\u540e\u662f\u5b8c\u5168\u4e0d\u540c\u7684\u74f6\u9888\u3002<\/p>\n\n\n\n

\u258e <\/strong>\u5e38\u89c1\u73b0\u8c61\u5bf9\u7167\u8868<\/strong><\/p>\n\n\n\n

\u25cf  <\/strong>\u4e13\u6709\u540d\u8bcd\u3001\u884c\u4e1a\u8bcd\u9519 \u2192 \u8bad\u7ec3\u6570\u636e\u91cc\u6ca1\u6709\uff0c\u4f18\u5148\u70ed\u8bcd \/ \u5fae\u8c03 \/ \u8bed\u8a00\u6a21\u578b<\/p>\n\n\n\n

\u25cf  <\/strong>\u566a\u58f0\u3001\u8fdc\u573a\u3001\u6df7\u54cd \u2192 \u524d\u7aef\u97f3\u9891\u5dee\uff0c\u4f18\u5148\u964d\u566a + VAD + \u6362\u6a21\u578b<\/p>\n\n\n\n

\u25cf  <\/strong>\u65b9\u8a00\u3001\u53e3\u97f3 \u2192 \u57fa\u5ea7\u8986\u76d6\u4e0d\u8db3\uff0c\u6362\u4e2d\u6587\/\u65b9\u8a00\u6a21\u578b\u6216\u5fae\u8c03<\/p>\n\n\n\n

\u25cf  <\/strong>\u5ef6\u8fdf\u9ad8\u3001\u8ddf\u4e0d\u4e0a\u8bf4\u8bdd \u2192 \u6d41\u5f0f\u914d\u7f6e\u6216\u6a21\u578b\u592a\u5927\uff0c\u5c0f\u6a21\u578b + INT8 + \u771f\u6d41\u5f0f<\/p>\n\n\n\n

\u25cf\u00a0\u00a0<\/strong>\u957f\u97f3\u9891\u8f6c\u5199\u6162 \u2192 \u975e\u6279\u5904\u7406\u6216 CPU \u63a8\u7406\uff0cfaster-whisper \/ GPU \/ \u5206\u6bb5\u5e76\u884c<\/p>\n\n\n\n

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

\u300c<\/strong>\u51c6\u786e\u7387\u548c\u901f\u5ea6\u901a\u5e38\u4e92\u76f8\u62c9\u626f\uff1a\u6a21\u578b\u8d8a\u5927\u8d8a\u51c6\u4f46\u8d8a\u6162\uff1bbeam \u8d8a\u5927\u8d8a\u51c6\u4f46\u8d8a\u6162\uff1bINT8 \u66f4\u5feb\u4f46\u53ef\u80fd\u7565\u635f\u7cbe\u5ea6\u3002\u4f18\u5316\u524d\u5148\u660e\u786e\u4f60\u7684\u4e3b KPI \u662f CER \u8fd8\u662f RTF\u3002<\/em>\u300d<\/strong><\/p>\n\n\n\n

\u4e8c\u3001\u4e94\u5c42\u4f18\u5316\u6846\u67b6\uff1a\u4ece\u8f93\u5165\u5230\u540e\u5904\u7406<\/strong><\/p>\n\n\n\n

\u4e0d\u7ba1\u7528\u54ea\u5957\u5f00\u6e90 ASR\uff0c\u90fd\u53ef\u4ee5\u6309\u4e94\u5c42\u6765\u601d\u8003\u4f18\u5316\u8def\u5f84\u3002<\/p>\n\n\n\n

\u81ea\u4e0b\u800c\u4e0a\u5206\u522b\u662f\uff1a\u6539\u8f93\u5165\u3001\u6539\u6a21\u578b\u3001\u6539\u89e3\u7801\u3001\u6539\u90e8\u7f72\u3001\u6539\u540e\u5904\u7406\u3002<\/p>\n\n\n\n

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

\u2460\u00a0\u00a0<\/strong>\u6539\u8f93\u5165\uff1a16 kHz \u5355\u58f0\u9053\u3001\u964d\u566a\u3001VAD \u5207\u9759\u97f3\u2014\u2014\u96f6\u6210\u672c\uff0coften \u88ab\u5ffd\u7565<\/p>\n\n\n\n

\u2461  <\/strong>\u6539\u6a21\u578b\uff1a\u6362\u66f4\u5f3a\u57fa\u5ea7\uff08SenseVoice \/ Qwen3-ASR\uff09\u6216\u9886\u57df\u5fae\u8c03<\/p>\n\n\n\n

\u2462  <\/strong>\u6539\u89e3\u7801\uff1abeam size\u3001\u70ed\u8bcd\u504f\u7f6e\u3001LM \u91cd\u6253\u5206<\/p>\n\n\n\n

\u2463  <\/strong>\u6539\u90e8\u7f72\uff1aONNX + INT8\u3001sherpa-onnx \u6d41\u5f0f\u3001faster-whisper \u6279\u5904\u7406<\/p>\n\n\n\n

\u2464  <\/strong>\u6539\u540e\u5904\u7406\uff1a\u6807\u70b9\u6a21\u578b\u3001ITN \u6570\u5b57\u89c4\u6574\u3001\u81ea\u5b9a\u4e49\u8bcd\u5178\u66ff\u6362<\/p>\n\n\n\n

\u5f88\u591a\u300c\u6a21\u578b\u4e0d\u51c6\u300d\u5176\u5b9e\u662f\u8f93\u5165\u97f3\u9891\u548c\u8bad\u7ec3\u5206\u5e03\u4e0d\u4e00\u81f4\u3002<\/p>\n\n\n\n

\u5148\u628a\u97f3\u9891\u6807\u51c6\u5316\u505a\u597d\uff0c\u518d\u8c08\u6362\u6a21\u578b\uff0c\u5f80\u5f80\u4e8b\u534a\u529f\u500d\u3002<\/p>\n\n\n\n

\u4e09\u3001\u63d0\u5347\u8bc6\u522b\u7387\u7684\u516d\u5927\u673a\u5236<\/strong><\/p>\n\n\n\n

\u4e0b\u9762\u6309\u300c\u4e0d\u6539\u6a21\u578b \u2192 \u6362\u6a21\u578b \u2192 \u5fae\u8c03 \u2192 \u89e3\u7801 \u2192 \u540e\u5904\u7406\u300d\u7684\u987a\u5e8f\uff0c\u4ece\u6613\u5230\u96be\u68b3\u7406\u63d0\u5347\u51c6\u786e\u7387\u7684\u624b\u6bb5\u3002<\/p>\n\n\n\n

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

 <\/strong>1. \u70ed\u8bcd \/ \u4e0a\u4e0b\u6587\u504f\u7f6e\uff08\u6700\u5feb\u89c1\u6548\uff09<\/strong><\/p>\n\n\n\n

\u673a\u5236\uff1a\u5728\u89e3\u7801\u65f6\u63d0\u9ad8\u7279\u5b9a\u8bcd\uff08\u516c\u53f8\u540d\u3001\u4ea7\u54c1\u540d\u3001\u4eba\u540d\uff09\u7684\u5f97\u5206\u6743\u91cd\u3002<\/p>\n\n\n\n

FunASR\u3001WeNet\u3001sherpa-onnx \u7b49\u5747\u652f\u6301\uff0c\u51e0\u4e4e\u4e0d\u7528\u91cd\u8bad\uff0c\u6539\u914d\u7f6e\u5373\u53ef\u3002<\/p>\n\n\n\n

\u9002\u5408\u8bcd\u8868\u56fa\u5b9a\u3001\u9886\u57df\u672f\u8bed\u591a\u7684\u573a\u666f\u3002<\/p>\n\n\n\n

\u258e <\/strong>2. \u8bed\u8a00\u6a21\u578b\u91cd\u6253\u5206\uff08LM Rescoring\uff09<\/strong><\/p>\n\n\n\n

\u673a\u5236\uff1aASR \u5148\u51fa N \u4e2a\u5019\u9009\uff08beam search\uff09\uff0c\u518d\u7528 N-gram \u6216\u795e\u7ecf\u7f51\u7edc LM \u9009\u66f4\u300c\u50cf\u4eba\u8bdd\u300d\u7684\u7ed3\u679c\u3002<\/p>\n\n\n\n

\u9002\u5408\u540c\u97f3\u5b57\u3001\u8bed\u6cd5\u7ea6\u675f\uff08\u5982\u300c\u5728\/\u518d\u300d\u300c\u7684\/\u5730\/\u5f97\u300d\uff09\u3002<\/p>\n\n\n\n

icefall \/ WeNet \u8bad\u7ec3\u94fe\u91cc\u5e38\u89c1\uff0c\u53ef\u5bfc\u51fa\u5230 sherpa-onnx \u90e8\u7f72\u3002<\/p>\n\n\n\n

\u258e <\/strong>3. Prompt \/ \u4e0a\u4e0b\u6587\uff08\u5927\u6a21\u578b ASR\uff09<\/strong><\/p>\n\n\n\n

Qwen3-ASR\u3001Fun-ASR-Nano \u7b49\u652f\u6301\u5728\u8f93\u5165\u91cc\u52a0\u63d0\u793a\uff0c\u4f8b\u5982\u300c\u8bed\u97f3\u8f6c\u5199\u6210\u82f1\u6587\uff1a\u300d\u6216\u300cSpeech transcription without text normalization:\u300d\u3002<\/p>\n\n\n\n

\u5229\u7528 LLM \u7684\u8bed\u4e49\u5148\u9a8c\uff0c\u7ea6\u675f\u8f93\u51fa\u683c\u5f0f\u548c\u9886\u57df\u3002<\/p>\n\n\n\n

\u258e <\/strong>4. VAD + \u5206\u6bb5<\/strong><\/p>\n\n\n\n

\u5148\u5207\u6389\u9759\u97f3\u548c\u65e0\u6548\u6bb5\uff0c\u53ea\u5bf9\u6709\u6548\u8bed\u97f3\u505a\u8bc6\u522b\uff0c\u51cf\u5c11\u300c\u628a\u566a\u58f0\u5f53\u5b57\u300d\u7684\u5e7b\u89c9\u3002<\/p>\n\n\n\n

FunASR \u81ea\u5e26 FSMN-VAD\uff1b\u672c\u5730\u5b9e\u65f6\u5de5\u5177 asr_tool \u7528 endpoint \u68c0\u6d4b\u53e5\u672b\u3002<\/p>\n\n\n\n

\u258e <\/strong>5. \u97f3\u9891\u524d\u7aef\u5904\u7406<\/strong><\/p>\n\n\n\n

\u25cf  <\/strong>\u91cd\u91c7\u6837\u5230 16 kHz \u5355\u58f0\u9053\uff1a\u4e0e\u6a21\u578b\u8bad\u7ec3\u5206\u5e03\u4e00\u81f4<\/p>\n\n\n\n

\u25cf  <\/strong>\u964d\u566a\uff08RNNoise\u3001DeepFilterNet\uff09\uff1a\u8fdc\u573a\u3001\u4f1a\u8bae\u573a\u666f<\/p>\n\n\n\n

\u25cf  <\/strong>AGC \/ \u97f3\u91cf\u5f52\u4e00\u5316\uff1a\u907f\u514d\u8fc7\u5c0f\u6216\u524a\u6ce2<\/p>\n\n\n\n

\u258e <\/strong>6. \u9886\u57df\u5fae\u8c03\uff08ROI \u901a\u5e38\u6700\u9ad8\uff09<\/strong><\/p>\n\n\n\n

\u7528\u4f60\u81ea\u5df1\u7684\u300c\u97f3\u9891 + \u6807\u6ce8\u300d\u628a\u6a21\u578b\u62c9\u5411\u76ee\u6807\u5206\u5e03\u3002<\/p>\n\n\n\n

\u6570\u636e < 10h \u5efa\u8bae\u70ed\u8bcd + LM\uff1b10\uff5e100h \u53ef LoRA \/ adaptor \u5fae\u8c03\uff1b100\uff5e1000h \u5fae\u8c03 encoder + adaptor\uff1b> 1000h \u53ef\u8003\u8651\u5168\u53c2\u3002<\/p>\n\n\n\n

\u540c\u4e00\u5957\u533b\u7597\/\u6cd5\u5f8b\u8bed\u6599\u5fae\u8c03\u540e\uff0c\u4e13\u6709\u540d\u8bcd\u51c6\u786e\u7387 often \u660e\u663e\u63d0\u5347\u3002<\/p>\n\n\n\n

\u56db\u3001\u63d0\u5347\u8bc6\u522b\u901f\u5ea6\u7684\u4e94\u6761\u8def\u5f84<\/strong><\/p>\n\n\n\n

\u901f\u5ea6\u4f18\u5316\u548c\u51c6\u786e\u7387\u5f80\u5f80\u662f\u53cd\u5411\u7684\u3002<\/p>\n\n\n\n

\u5b9e\u65f6\u573a\u666f\u8981\u7684\u662f\u9996\u5b57\u5ef6\u8fdf\u548c RTF\uff08\u5b9e\u65f6\u7387\uff09\uff0c\u79bb\u7ebf\u573a\u666f\u8981\u7684\u662f\u541e\u5410\u3002<\/p>\n\n\n\n

\u522b\u7528\u79bb\u7ebf\u5927\u6a21\u578b\u786c\u505a\u6d41\u5f0f\u2014\u2014\u5ef6\u8fdf\u4f1a\u975e\u5e38\u9ad8\u3002<\/p>\n\n\n\n

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

\u6a21\u578b\u4fa7\uff1a\u53d8\u5c0f\u3001\u53d8\u8f7b<\/strong><\/p>\n\n\n\n

\u25cf  <\/strong>\u84b8\u998f\uff08distil-whisper\uff09\uff1a\u5927\u6a21\u578b\u6559\u5c0f\u6a21\u578b\uff0c6\uff5e10\u00d7 \u52a0\u901f\uff0c\u7cbe\u5ea6\u7565\u964d<\/p>\n\n\n\n

\u25cf  <\/strong>\u91cf\u5316 INT8\uff1a\u6743\u91cd 8bit\uff0cCPU \u53cb\u597d\uff0c2\uff5e4\u00d7 \u52a0\u901f<\/p>\n\n\n\n

\u25cf  <\/strong>\u6362\u5c0f\u67b6\u6784\uff1aZipformer 14M vs 1.7B LLM\uff0c\u6570\u91cf\u7ea7\u5dee\u5f02<\/p>\n\n\n\n

\u258e <\/strong>\u8fd0\u884c\u65f6\uff1a\u6362\u5f15\u64ce<\/strong><\/p>\n\n\n\n

\u25cf  <\/strong>sherpa-onnx\uff1aCPU \u6d41\u5f0f\u3001\u4f4e\u5ef6\u8fdf\uff0cRust \u96c6\u6210\uff0c\u9002\u5408\u684c\u9762\u5b9e\u65f6<\/p>\n\n\n\n

\u25cf  <\/strong>faster-whisper\uff08CTranslate2\uff09\uff1aGPU \u6279\u5904\u7406\u957f\u97f3\u9891\u5feb<\/p>\n\n\n\n

\u25cf  <\/strong>whisper.cpp\uff1a\u8fb9\u7f18\u8bbe\u5907\u3001\u65e0 Python<\/p>\n\n\n\n

\u25cf  <\/strong>vLLM\uff1aQwen3-ASR \u7b49\u670d\u52a1\u5316\u90e8\u7f72<\/p>\n\n\n\n

\u258e <\/strong>\u6d41\u5f0f\u67b6\u6784\uff08\u5ef6\u8fdf\u7684\u5173\u952e\uff09<\/strong><\/p>\n\n\n\n

\u771f\u6d41\u5f0f pipeline\uff1a\u9ea6\u514b\u98ce \u2192 \u56fa\u5b9a chunk\uff08\u5982 100ms\uff09\u2192 \u589e\u91cf decode \u2192 partial \u7ed3\u679c \u2192 endpoint \u68c0\u6d4b \u2192 final \u2192 reset\u3002<\/p>\n\n\n\n

chunk_size \u8d8a\u5c0f\u5ef6\u8fdf\u8d8a\u4f4e\uff0c\u4f46 CPU \u8c03\u5ea6\u5f00\u9500\u4e0a\u5347\uff1bnum_threads 4\uff5e8 \u5e38\u662f CPU \u751c\u70b9\u3002<\/p>\n\n\n\n

\u258e <\/strong>\u7cfb\u7edf\u7ea7\u4f18\u5316<\/strong><\/p>\n\n\n\n

\u25cf  <\/strong>GPU \/ NPU\uff1a\u79bb\u7ebf\u6279\u5904\u7406\u3001\u5927\u6a21\u578b\u5fc5\u5907<\/p>\n\n\n\n

\u25cf  <\/strong>\u8df3\u8fc7\u9759\u97f3\uff08VAD\uff09\uff1a\u4e0d\u7b97\u65e0\u6548\u6bb5\uff0c\u603b\u8017\u65f6\u4e0b\u964d<\/p>\n\n\n\n

\u25cf  <\/strong>\u5e76\u884c\u5206\u6bb5\uff1a\u957f\u97f3\u9891\u5207\u591a\u6bb5\u591a GPU \u8dd1<\/p>\n\n\n\n

\u25cf\u00a0\u00a0<\/strong>\u51cf\u5c0f beam\uff1abeam=1 often \u5feb 2\uff5e3\u00d7<\/p>\n\n\n\n

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

\u4e94\u3001\u6309\u6295\u5165\u4ea7\u51fa\u6392\u5e8f\u7684\u63a8\u8350<\/strong><\/p>\n\n\n\n

\u258e <\/strong>\u60f3\u63d0\u5347\u51c6\u786e\u7387\uff08\u4ece\u6613\u5230\u96be\uff09<\/strong><\/p>\n\n\n\n

\u2460  <\/strong>\u97f3\u9891\u6807\u51c6\u5316\uff0816 kHz\u3001\u964d\u566a\u3001VAD\uff09\u2014\u2014 \u96f6\u6210\u672c<\/p>\n\n\n\n

\u2461  <\/strong>\u70ed\u8bcd \/ Prompt \u2014\u2014 \u51e0\u5c0f\u65f6\u914d\u7f6e<\/p>\n\n\n\n

\u2462  <\/strong>\u8c03 beam + \u52a0\u6807\u70b9\/ITN \u2014\u2014 \u4e00\u5929\u5185<\/p>\n\n\n\n

\u2463  <\/strong>\u6362\u66f4\u5f3a\u7684\u4e2d\u6587\u57fa\u5ea7\uff08SenseVoice \/ Qwen3-ASR\uff09\u2014\u2014 \u6539\u90e8\u7f72<\/p>\n\n\n\n

\u2464  <\/strong>LM \u91cd\u6253\u5206 \u2014\u2014 \u9700\u989d\u5916 LM \u8d44\u6e90<\/p>\n\n\n\n

\u2465  <\/strong>\u9886\u57df\u5fae\u8c03 \u2014\u2014 \u9700\u6807\u6ce8\u6570\u636e\uff0cROI \u901a\u5e38\u6700\u9ad8<\/p>\n\n\n\n

\u258e <\/strong>\u60f3\u63d0\u5347\u901f\u5ea6\uff08\u4ece\u6613\u5230\u96be\uff09<\/strong><\/p>\n\n\n\n

\u2460  <\/strong>INT8 \u91cf\u5316 + \u51cf beam \u2014\u2014 \u7acb\u523b\u89c1\u6548<\/p>\n\n\n\n

\u2461  <\/strong>\u6362 faster-whisper \/ sherpa-onnx \u2014\u2014 \u6539\u8fd0\u884c\u65f6<\/p>\n\n\n\n

\u2462  <\/strong>\u6362\u5c0f\u6a21\u578b \/ \u84b8\u998f\u6a21\u578b \u2014\u2014 \u63a5\u53d7\u7565\u964d\u7cbe\u5ea6<\/p>\n\n\n\n

\u2463  <\/strong>GPU + \u6279\u5904\u7406 \u2014\u2014 \u79bb\u7ebf\u573a\u666f<\/p>\n\n\n\n

\u2464  <\/strong>\u771f\u6d41\u5f0f\u67b6\u6784 \u2014\u2014 \u5b9e\u65f6\u573a\u666f\u5fc5\u505a<\/p>\n\n\n\n

  \u00b7 \u00b7 \u00b7  <\/p>\n\n\n\n

\u516d\u3001\u573a\u666f\u5316\u843d\u5730\u7ec4\u5408<\/strong><\/p>\n\n\n\n

\u7ed3\u5408 voice_repo \u91cc\u7684\u5de5\u5177\u94fe\uff0c\u4ee5\u4e0b\u662f\u51e0\u79cd\u5e38\u89c1\u573a\u666f\u7684\u7ec4\u5408\u5efa\u8bae\u3002<\/p>\n\n\n\n

\u258e <\/strong>\u684c\u9762\u5b9e\u65f6\u4e2d\u6587<\/strong><\/p>\n\n\n\n

\u51c6\u786e\u7387\uff1a\u70ed\u8bcd + \u66f4\u597d\u7684ASR\u6a21\u578b\u3002<\/p>\n\n\n\n

\u901f\u5ea6\uff1a\u4fdd\u6301sherpa-onnx INT8 \u771f\u6d41\u5f0f\u3002<\/p>\n\n\n\n

\u9ed8\u8ba4\u6a21\u578b streaming-***-zh-14M INT8\uff0c\u9762\u5411 CPU \u5b9e\u65f6\u3002<\/p>\n\n\n\n

\u258e <\/strong>\u4f1a\u8bae\u5f55\u97f3\u8f6c\u5199<\/strong><\/p>\n\n\n\n

\u51c6\u786e\u7387\uff1aSenseVoice \/ Qwen3-ASR + \u6807\u70b9\u6a21\u578b\u3002<\/p>\n\n\n\n

\u901f\u5ea6\uff1afaster-whisper GPU \u6279\u5904\u7406\uff0c\u957f\u97f3\u9891\u5206\u6bb5\u5e76\u884c\u3002<\/p>\n\n\n\n

\u258e <\/strong>\u5782\u76f4\u9886\u57df\uff08\u533b\u7597 \/ \u6cd5\u5f8b \/ \u91d1\u878d\uff09<\/strong><\/p>\n\n\n\n

\u51c6\u786e\u7387\uff1aFunASR \u5fae\u8c03 + \u70ed\u8bcd + LM \u91cd\u6253\u5206\u3002<\/p>\n\n\n\n

\u901f\u5ea6\uff1a\u5fae\u8c03\u540e\u5bfc\u51fa ONNX \u5230 sherpa-onnx\uff0c\u63a5\u81ea\u7814\u670d\u52a1\u3002<\/p>\n\n\n\n

\u258e <\/strong>\u591a\u8bed\u79cd \/ \u65b9\u8a00<\/strong><\/p>\n\n\n\n

\u51c6\u786e\u7387\uff1aQwen3-ASR + Prompt \u7ea6\u675f\u3002<\/p>\n\n\n\n

\u901f\u5ea6\uff1avLLM \u670d\u52a1\u5316\uff0c\u6216 0.6B \u5c0f\u6a21\u578b\u505a\u8fb9\u7f18\u90e8\u7f72\u3002<\/p>\n\n\n\n

  \u00b7 \u00b7 \u00b7  <\/p>\n\n\n\n

\u4e03\u3001\u600e\u4e48\u9a8c\u8bc1\u300c\u771f\u7684\u53d8\u597d\u4e86\u300d<\/strong><\/p>\n\n\n\n

\u522b\u51ed\u611f\u89c9\uff0c\u7528\u540c\u4e00\u5957 test set \u5bf9\u6bd4\u3002<\/p>\n\n\n\n

\u56fa\u5b9a 100\uff5e500 \u6761\u4ee3\u8868\u6027\u6837\u672c\uff0c\u6bcf\u6b21\u53ea\u6539\u4e00\u4e2a\u53d8\u91cf\uff08\u4f8b\u5982\u53ea\u52a0\u70ed\u8bcd\u3001\u53ea\u6362\u6a21\u578b\uff09\uff0c\u5426\u5219\u65e0\u6cd5\u5224\u65ad\u54ea\u62db\u6709\u6548\u3002<\/p>\n\n\n\n

\u25cf  <\/strong>CER \/ WER\uff1a\u5b57\/\u8bcd\u9519\u8bef\u7387\uff0c\u4e2d\u6587\u5e38\u7528 CER<\/p>\n\n\n\n

\u25cf  <\/strong>RTF\uff1a\u5b9e\u65f6\u7387\uff0c< 1 \u624d\u7b97\u5b9e\u65f6<\/p>\n\n\n\n

\u25cf  <\/strong>\u9996\u5b57\u5ef6\u8fdf\uff1a\u6d41\u5f0f\u573a\u666f\u7684\u5173\u952e\u6307\u6807<\/p>\n\n\n\n

\u25cf  <\/strong>\u9886\u57df\u8bcd\u51c6\u786e\u7387\uff1a\u5355\u72ec\u7edf\u8ba1\u70ed\u8bcd\u8868\u547d\u4e2d\u7387<\/p>\n\n\n\n

\u300c<\/strong>\u5f00\u6e90 ASR \u7684\u4f18\u5316\u4e0d\u662f\u300c\u6362\u4e00\u4e2a\u66f4\u5927\u7684\u6a21\u578b\u5c31\u5b8c\u4e8b\u300d\uff0c\u800c\u662f\u6309\u573a\u666f\u9009\u5bf9\u673a\u5236\u3001\u6309\u6570\u636e\u91cf\u9009\u5bf9\u6df1\u5ea6\u3001\u7528\u6307\u6807\u9a8c\u8bc1\u6bcf\u4e00\u6b65\u3002<\/em>\u300d<\/strong><\/p>\n\n\n\n

\u603b\u7ed3          
<\/strong>\u4e0d\u591f\u51c6\uff1a\u97f3\u9891\u8d28\u91cf \u2192 \u70ed\u8bcd\/LM \u2192 \u6362\u4e2d\u6587\u6a21\u578b \u2192 \u9886\u57df\u5fae\u8c03          
\u2022 \u4e0d\u591f\u5feb\uff1a\u5c0f\u6a21\u578b + INT8 + \u5408\u9002\u8fd0\u884c\u65f6\uff08\u6d41\u5f0f sherpa-onnx\uff0c\u79bb\u7ebf faster-whisper\uff09          
\u2022 \u4e94\u5c42\u6846\u67b6\uff1a\u6539\u8f93\u5165 \u2192 \u6539\u6a21\u578b \u2192 \u6539\u89e3\u7801 \u2192 \u6539\u90e8\u7f72 \u2192 \u6539\u540e\u5904\u7406          
\u2022 \u51c6\u786e\u4e0e\u901f\u5ea6\u4e92\u76f8\u62c9\u626f\uff0c\u4f18\u5316\u524d\u5148\u660e\u786e\u4e3b KPI          
\u2022 \u56fa\u5b9a test set\uff0c\u6bcf\u6b21\u53ea\u6539\u4e00\u4e2a\u53d8\u91cf\uff0c\u7528 CER\/RTF \u9a8c\u8bc1<\/p>\n","protected":false},"excerpt":{"rendered":"

\u6587\u7ae0\u6765\u6e90\uff1a\u6bcf\u5468\u4e00\u4e2a\u5927\u6a21\u578b\u5e94\u7528 Whisper\u3001SenseVoice\u3001sherpa-onnx\u2026\u2026\u5f00\u6e90 ASR \u9009 … \u7ee7\u7eed\u9605\u8bfb\u5f00\u6e90ASR\u4f18\u5316\u6307\u5357<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[4,38,34],"tags":[],"_links":{"self":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/31259"}],"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=31259"}],"version-history":[{"count":6,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/31259\/revisions"}],"predecessor-version":[{"id":31271,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/31259\/revisions\/31271"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=31259"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=31259"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=31259"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}