{"id":23,"date":"2026-06-08T09:49:08","date_gmt":"2026-06-08T01:49:08","guid":{"rendered":"http:\/\/tantan.ink\/?p=23"},"modified":"2026-06-08T09:49:08","modified_gmt":"2026-06-08T01:49:08","slug":"nemotron-35-asr-streaming","status":"publish","type":"post","link":"https:\/\/www.tantan.ink\/index.php\/2026\/06\/08\/nemotron-35-asr-streaming\/","title":{"rendered":"NVIDIA Nemotron 3.5 ASR\uff1a600M \u53c2\u6570\u7684\u6d41\u5f0f\u591a\u8bed\u8a00\u8bed\u97f3\u8bc6\u522b\uff0c\u672c\u5730\u4e5f\u80fd\u8dd1"},"content":{"rendered":"<h2>\u4ece\u5f00\u6e90 ASR \u8bf4\u8d77<\/h2>\n<p>\u6700\u8fd1 NVIDIA \u5728 Hugging Face \u4e0a\u653e\u51fa\u4e86 <strong>Nemotron 3.5 ASR Streaming<\/strong>\uff0c\u4e00\u4e2a 600M \u53c2\u6570\u7684\u6d41\u5f0f\u8bed\u97f3\u8bc6\u522b\u6a21\u578b\u3002<\/p>\n<p>\u5b83\u548c\u4e4b\u524d\u6587\u7ae0\u91cc\u5199\u8fc7\u7684 Nemotron 3 Ultra \u5927\u8bed\u8a00\u6a21\u578b\u4e0d\u662f\u4e00\u56de\u4e8b\u2014\u2014\u8fd9\u4e2a ASR \u4e13\u6ce8\u505a\u4e00\u4ef6\u4e8b\uff1a\u8bed\u97f3\u8f6c\u6587\u5b57\u3002\u800c\u4e14\u5b83\u652f\u6301\u5b9e\u65f6\u6d41\u5f0f\u5904\u7406\uff0c<strong>\u4e00\u4e2a\u6a21\u578b\u8986\u76d6 40 \u79cd\u8bed\u8a00<\/strong>\uff0c\u4e0d\u9700\u8981\u9010\u4e2a\u8bed\u8a00\u6362\u6a21\u578b\u3002<\/p>\n<p>\u6211\u5173\u6ce8\u5b83\uff0c\u4e0d\u662f\u56e0\u4e3a NVIDIA \u724c\u5b50\u7684\u5149\u73af\uff0c\u800c\u662f\u56e0\u4e3a\u5b83\u6070\u597d\u5361\u5728\u51e0\u4e2a\u6211\u6bd4\u8f83\u5728\u610f\u7684\u4ea4\u53c9\u70b9\u4e0a\uff1a<\/p>\n<ul>\n<li><strong>\u672c\u5730\u53ef\u8dd1<\/strong>\u2014\u2014600M \u53c2\u6570\uff080.6B\uff09\uff0c\u5f53\u4ee3 GPU \u6216\u8005\u54ea\u6015 Apple Silicon Mac \u5e94\u8be5\u90fd\u80fd\u63a8\u7406<\/li>\n<li><strong>\u6d41\u5f0f\u67b6\u6784<\/strong>\u2014\u2014\u4e0d\u50cf\u4f20\u7edf ASR \u8981\u7b49\u8bf4\u5b8c\u4e00\u6574\u53e5\u8bdd\u624d\u51fa\u7ed3\u679c\uff0c\u5b83\u662f\u8fb9\u542c\u8fb9\u51fa\u7684<\/li>\n<li><strong>\u5f00\u6e90\u6743\u91cd<\/strong>\u2014\u2014Apache 2.0 \u7cfb\u7684\u5f00\u6e90\u534f\u8bae\uff0c\u6ca1\u6709 API \u4f9d\u8d56\uff0c\u662f\u771f\u7684\u672c\u5730\u90e8\u7f72\u8def\u7ebf<\/li>\n<li><strong>\u4e2d\u6587\u652f\u6301<\/strong>\u2014\u2014zh-CN \u5728&#8221;\u5e7f\u8986\u76d6&#8221;\u5c42\u7ea7\uff0c\u5f00\u7bb1\u53ef\u7528\uff0c\u4f46\u4e0d\u50cf\u82f1\u6587\u90a3\u4e48\u7cbe\u51c6<\/li>\n<\/ul>\n<p style=\"text-align:center;\"><img decoding=\"async\" class=\"alignnone size-full\" src=\"http:\/\/tantan.ink\/wp-content\/uploads\/2026\/06\/nemotron-overview.png\" alt=\"Nemotron 3.5 ASR Overview\" width=\"800\" \/><br \/><em>Nemotron 3.5 ASR \u6982\u89c8\uff1a\u591a\u8bed\u8a00\u8f93\u5165 \u2192 FastConformer-RNNT \u2192 \u5e26\u6807\u70b9\u548c\u8bed\u8a00\u6807\u7b7e\u7684\u6587\u5b57\u8f93\u51fa\uff08\u6765\u6e90\uff1aNVIDIA \/ Hugging Face\uff09<\/em><\/p>\n<h2>\u8fd9\u6b21\u53d1\u5e03\u7684\u5230\u5e95\u662f\u4ec0\u4e48<\/h2>\n<p>Nemotron 3.5 ASR Streaming\uff08\u6a21\u578b\u540d\uff1a<code>nvidia\/nemotron-3.5-asr-streaming-0.6b<\/code>\uff09\u662f NVIDIA \u7b2c\u4e09\u4ee3\u7684\u8bed\u97f3\u8bc6\u522b\u6a21\u578b\uff0c\u67b6\u6784\u662f <strong>FastConformer-CacheAware-RNNT<\/strong>\u3002<\/p>\n<p>\u8bf4\u4eba\u8bdd\u5c31\u662f\uff1a\u5b83\u7528\u4e86\u7f13\u5b58\u611f\u77e5\u7684\u7f16\u7801\u5668+RNNT \u89e3\u7801\u5668\u7ec4\u5408\uff0c\u4e0d\u9700\u8981\u7b49\u97f3\u9891\u5b8c\u6574\u4f20\u8f93\u5c31\u80fd\u8f93\u51fa\u6587\u5b57\u3002\u5ef6\u8fdf\u53ef\u4ee5\u8c03\u8282\uff0c\u6700\u4f4e\u5230 <strong>80ms<\/strong>\uff08\u63a5\u8fd1\u5b9e\u65f6\uff09\uff0c\u6700\u9ad8 1.12s\uff08\u8ffd\u6c42\u6700\u9ad8\u51c6\u786e\u7387\uff09\uff0c\u540c\u4e00\u4e2a\u6a21\u578b\u4e0d\u7528\u91cd\u65b0\u8bad\u7ec3\u3002<\/p>\n<p style=\"text-align:center;\"><img decoding=\"async\" class=\"alignnone size-full\" src=\"http:\/\/tantan.ink\/wp-content\/uploads\/2026\/06\/nemotron-architecture.png\" alt=\"Nemotron 3.5 ASR Architecture\" width=\"800\" \/><br \/><em>Nemotron 3.5 ASR \u67b6\u6784\uff1aFastConformer \u7f16\u7801\u5668\u4e0e\u8bed\u8a00 ID \u7f16\u7801\u62fc\u63a5\u540e\u6295\u5f71\u5230 RNNT \u89e3\u7801\u5668\uff08\u6765\u6e90\uff1aNVIDIA \/ Hugging Face\uff09<\/em><\/p>\n<p>\u6700\u5b9e\u7528\u7684\u8bbe\u8ba1\u662f\uff1a\u652f\u6301\u8bed\u8a00\u81ea\u52a8\u68c0\u6d4b\u3002\u4f60\u7ed9\u5b83\u4e00\u6bb5\u6df7\u5408\u591a\u79cd\u8bed\u8a00\u7684\u5f55\u97f3\uff0c\u5b83\u80fd\u81ea\u52a8\u8bc6\u522b\u5e76\u5207\u6362\u8bc6\u522b\u8bed\u8a00\u3002<\/p>\n<p>\u4ece\u5177\u4f53\u6570\u636e\u6765\u770b\uff0c\u5b83\u5728 FLEURS \u6d4b\u8bd5\u96c6\u4e0a\u7684\u8868\u73b0\uff1a<\/p>\n<p><strong>\u6700\u4f73\u8bed\u8a00\uff08\u5df2\u77e5\u8bed\u8a00 ID \u8f93\u5165\uff0c80ms \u4f4e\u5ef6\u8fdf\u6a21\u5f0f\uff09<\/strong><\/p>\n<ul>\n<li>\u897f\u73ed\u7259\u8bed\uff1aWER 4.87%<\/li>\n<li>\u610f\u5927\u5229\u8bed\uff1a5.23%<\/li>\n<li>\u8461\u8404\u7259\u8bed\uff1a6.29%<\/li>\n<li>\u5370\u5730\u8bed\uff1a8.13%<\/li>\n<li>\u97e9\u8bed\uff08CER\uff09\uff1a7.59%<\/li>\n<\/ul>\n<p><strong>\u4e2d\u6587\uff08zh-CN\uff09\u8868\u73b0<\/strong>\uff1aCER \u7ea6 20.56%\uff0880ms \u6a21\u5f0f\uff0c\u5df2\u77e5\u8bed\u8a00\uff09\uff0c\u5c5e\u4e8e&#8221;\u5e7f\u8986\u76d6&#8221;\u5c42\u7ea7\u2014\u2014\u80fd\u8bc6\u522b\uff0c\u4f46\u8fbe\u4e0d\u5230\u7cbe\u51c6\u7ea7\u522b\u3002<\/p>\n<p>\u5bf9\u6bd4\u4e3b\u6d41\u7684\u5546\u4e1a\u8bed\u97f3\u8bc6\u522b API\uff0c\u8fd9\u4e2a\u6570\u5b57\u4e0d\u7b97\u60ca\u8273\uff0c\u4f46\u8003\u8651\u5230\u5b83 600M \u53c2\u6570\u5c31\u80fd\u5728\u672c\u5730\u6d41\u5f0f\u8dd1\uff0c\u800c\u4e14\u662f\u4e00\u4e2a\u6a21\u578b\u625b 40 \u79cd\u8bed\u8a00\uff0c\u6211\u89c9\u5f97\u8fd9\u4e2a\u53d6\u820d\u53ef\u4ee5\u63a5\u53d7\u3002<\/p>\n<p style=\"text-align:center;\"><img decoding=\"async\" class=\"alignnone size-full\" src=\"http:\/\/tantan.ink\/wp-content\/uploads\/2026\/06\/nemotron-fleurs-wer-chart.png\" alt=\"FLEURS WER vs Chunk Size\" width=\"800\" \/><br \/><em>FLEURS \u6d4b\u8bd5\u96c6\u5e73\u5747 WER vs \u6d41\u5f0f chunk \u5927\u5c0f\uff1aLangID \u4e0e Auto-detect \u6a21\u5f0f\u5bf9\u6bd4\uff08\u6765\u6e90\uff1aNVIDIA \/ Hugging Face\uff09<\/em><\/p>\n<h2>\u5bf9\u6211\u6765\u8bf4\u503c\u5f97\u6ce8\u610f\u7684\u70b9<\/h2>\n<h3>1. \u672c\u5730\u90e8\u7f72\u7684\u53ef\u80fd\u6027<\/h3>\n<p>\u8fd9\u4e2a\u6a21\u578b\u9700\u8981 NeMo \u6846\u67b6\uff08NVIDIA \u7684\u5f00\u6e90\u8bed\u97f3 AI \u6846\u67b6\uff09\u6765\u8dd1\u63a8\u7406\u3002NeMo \u672c\u8eab\u652f\u6301 GPU\uff0c\u4e5f\u652f\u6301 TensorRT \u4f18\u5316\u3002<\/p>\n<p>\u5bf9\u4e8e ARM Mac \u7528\u6237\u6765\u8bf4\uff0c\u76ee\u524d\u6ca1\u6cd5\u76f4\u63a5\u8dd1 NeMo \u7684\u5b8c\u6574\u6d41\u6c34\u7ebf\u2014\u2014NeMo \u7684 CUDA \u4f9d\u8d56\u51b3\u5b9a\u4e86\u5b83\u4e3b\u8981\u8fd8\u662f\u9762\u5411 NVIDIA GPU \u7684\u3002\u4f46\u8fd9\u4e0d\u4ee3\u8868\u5b8c\u5168\u4e0d\u80fd\u7528\uff1a\u4e00\u4e2a\u53ef\u80fd\u8def\u5f84\u662f\u8d70 ONNX \u5bfc\u51fa\u540e\u7528 CPU \u63a8\u7406\uff0c\u53e6\u4e00\u4e2a\u7b49\u5f85\u793e\u533a\u5c01\u88c5\u3002<\/p>\n<h3>2. \u5fae\u8c03\u6f5c\u529b<\/h3>\n<p>NVIDIA \u8fd9\u6b21\u540c\u6b65\u53d1\u4e86\u4e00\u7bc7\u8be6\u7ec6\u7684\u5fae\u8c03\u6307\u5357\uff08Hugging Face Blog\uff09\uff0c\u8bb2\u4e86\u5982\u4f55\u5bf9\u65b0\u7684\u8bed\u8a00\u3001\u9886\u57df\uff08\u533b\u7597\u3001\u6cd5\u5f8b\uff09\u3001\u53e3\u97f3\u505a\u5fae\u8c03\u3002<\/p>\n<p>\u5173\u952e\u6570\u636e\u70b9\uff1a<\/p>\n<ul>\n<li>\u7528 ~2000 \u5c0f\u65f6\u6570\u636e\u5fae\u8c03\u540e\uff0c\u5e0c\u814a\u8bed WER \u4ece 35% \u964d\u5230 24%\uff08\u6539\u8fdb 32%\uff09<\/li>\n<li>\u4fdd\u52a0\u5229\u4e9a\u8bed WER \u4ece 22% \u964d\u5230 15%\uff08\u6539\u8fdb 31%\uff09<\/li>\n<li>\u6539\u8fdb\u5e45\u5ea6\u6700\u5927\u7684\u6070\u6070\u662f\u57fa\u7840\u51c6\u786e\u7387\u6700\u4f4e\u7684\u8bed\u8a00<\/li>\n<\/ul>\n<p>\u8fd9\u4e2a\u6a21\u5f0f\u5bf9\u4e2d\u6587\u6765\u8bf4\u4e5f\u662f\u7c7b\u4f3c\u7684\uff1a\u5b98\u65b9\u57fa\u7ebf\u7684 CER 20% \u4e0d\u7b97\u597d\uff0c\u4f46\u5982\u679c\u4f60\u6709\u9886\u57df\u5185\u7684\u4e2d\u6587\u8bed\u97f3\u6570\u636e\u505a\u5fae\u8c03\uff0c\u5e94\u8be5\u80fd\u663e\u8457\u63d0\u5347\u3002<\/p>\n<h3>3. \u6d41\u5f0f\u5ef6\u8fdf\u7684\u52a8\u6001\u8c03\u4f18<\/h3>\n<p>\u8fd9\u4e2a\u6a21\u578b\u8ba9\u6211\u89c9\u5f97\u6bd4\u8f83\u806a\u660e\u7684\u5730\u65b9\u662f\uff1a<strong>\u5ef6\u8fdf\u548c\u51c6\u786e\u7387\u7684\u5e73\u8861\u53ef\u4ee5\u5728\u63a8\u7406\u65f6\u52a8\u6001\u8c03\u6574<\/strong>\uff0c\u4e0d\u9700\u8981\u91cd\u65b0\u8bad\u7ec3\u3002<\/p>\n<ul>\n<li>80ms\uff1a\u6781\u4f4e\u5ef6\u8fdf\uff0c\u9002\u5408\u5b9e\u65f6\u8bed\u97f3\u52a9\u624b<\/li>\n<li>160ms\uff1a\u4f4e\u5ef6\u8fdf\uff0c\u9002\u5408\u4ea4\u4e92\u5f0f\u5bf9\u8bdd<\/li>\n<li>320ms\uff1a\u4e2d\u7b49\u5ef6\u8fdf\uff0c\u9002\u5408\u5bf9\u8bdd\u8bb0\u5f55\/\u4f1a\u8bae<\/li>\n<li>560ms\uff1a\u8f83\u9ad8\u5ef6\u8fdf\uff0c\u9002\u5408\u9ad8\u51c6\u786e\u7387\u573a\u666f<\/li>\n<li>1.12s\uff1a\u6700\u9ad8\u5ef6\u8fdf\uff0c\u9002\u5408\u6700\u9ad8\u51c6\u786e\u7387<\/li>\n<\/ul>\n<p>\u540c\u4e00\u4efd\u8bad\u7ec3\u597d\u7684 checkpoint\uff0c\u6362\u4e00\u4e2a\u53c2\u6570\u5c31\u80fd\u5207\u6362\u573a\u666f\u3002<\/p>\n<h2>\u51e0\u4e2a\u5b9e\u9645\u5236\u7ea6<\/h2>\n<ul>\n<li><strong>NeMo \u751f\u6001\u8f83\u91cd<\/strong>\u2014\u2014\u4e0d\u662f pip install \u76f4\u63a5\u8dd1\u90a3\u4e48\u7b80\u5355\uff0c\u9700\u8981 clone \u6574\u4e2a NeMo \u4ed3\u5e93\uff0c\u8d70\u5b83\u7684 pipeline\u3002\u5bf9\u4e8e\u60f3&#8221;\u4e0b\u4e2a\u6a21\u578b\u7acb\u523b\u7528&#8221;\u7684\u6298\u817e\u515a\u6765\u8bf4\uff0c\u95e8\u69db\u6709\u70b9\u9ad8<\/li>\n<li><strong>\u4e2d\u6587\u7cbe\u51c6\u5ea6\u4ecd\u9700\u5fae\u8c03<\/strong>\u2014\u201420% \u7684 CER \u5728\u751f\u4ea7\u73af\u5883\u4e0d\u592a\u591f\u7528\uff0c\u5c24\u5176\u5bf9\u4e8e\u4e2d\u6587\u8fd9\u79cd\u540c\u97f3\u5b57\u5bc6\u96c6\u7684\u8bed\u8a00<\/li>\n<li><strong>\u534f\u8bae\u662f OpenMDW-1.1<\/strong>\u2014\u2014\u4e0d\u662f\u7eaf Apache 2.0\uff0c\u9700\u8981\u786e\u8ba4\u5546\u7528\u6761\u6b3e\u662f\u5426\u7b26\u5408\u4f60\u7684\u573a\u666f<\/li>\n<li><strong>NIM \u90e8\u7f72\u7248\u672c\u8fd8\u672a\u51fa<\/strong>\u2014\u2014NVIDIA \u540e\u7eed\u4f1a\u63d0\u4f9b gRPC \u6d41\u5f0f\u90e8\u7f72\u65b9\u6848\uff08NIM\uff09\uff0c\u4f46\u76ee\u524d\u8fd8\u5728\u8def\u4e0a<\/li>\n<\/ul>\n<h2>\u6211\u7684\u5224\u65ad<\/h2>\n<p>Nemotron 3.5 ASR \u4e0d\u662f\u4e00\u4e2a&#8221;\u88c5\u4e0a\u5c31\u80fd\u66ff\u4ee3 Whisper&#8221;\u7684\u4e1c\u897f\u2014\u2014Whisper \u7684\u4e2d\u6587\u51c6\u786e\u7387\u3001\u751f\u6001\u6210\u719f\u5ea6\u548c\u4f7f\u7528\u4fbf\u5229\u6027\u90fd\u660e\u663e\u66f4\u9ad8\u3002<\/p>\n<p>\u4f46\u5b83\u4ee3\u8868\u4e86\u4e00\u4e2a\u5f88\u6709\u8da3\u7684\u65b9\u5411\uff1a<strong>\u591a\u8bed\u8a00\u6d41\u5f0f ASR \u7684\u5c0f\u578b\u5316<\/strong>\u3002600M \u53c2\u6570\u5c31\u80fd\u505a\u5230 40 \u8bed\u8a00\u7684\u5b9e\u65f6\u8f6c\u5f55\uff0c\u8fd9\u5728\u4e00\u5e74\u524d\u8fd8\u662f\u9700\u8981\u66f4\u5927\u6a21\u578b\u6216\u8005\u5546\u4e1a API \u624d\u80fd\u5b9e\u73b0\u7684\u3002<\/p>\n<p>\u5982\u679c\u4f60\u6b63\u5728\u505a\u4e00\u4e2a\u591a\u8bed\u8a00\u8bed\u97f3\u52a9\u624b\u3001\u6216\u8005\u9700\u8981\u672c\u5730\u7684\u5b9e\u65f6\u5b57\u5e55\u65b9\u6848\uff0c\u800c\u4e14\u4f60\u7684\u4e3b\u8981\u8bed\u8a00\u4e0d\u662f\u4e2d\u6587\uff08\u82f1\/\u897f\/\u610f\u7b49\u4f18\u5148\u652f\u6301\u7684\u8bed\u8a00\uff09\uff0c\u90a3\u4e48\u8fd9\u4e2a\u6a21\u578b\u503c\u5f97\u5173\u6ce8\u540e\u7eed\u7684\u793e\u533a\u5c01\u88c5\u8fdb\u5c55\u3002\u5982\u679c\u4f60\u4e3b\u8981\u505a\u4e2d\u6587 ASR\uff0c\u76ee\u524d Whisper \u4ecd\u7136\u662f\u66f4\u52a1\u5b9e\u7684\u9009\u62e9\u3002<\/p>\n<p>\u7b49\u540e\u7eed NIM \u90e8\u7f72\u5305\u6216\u8005\u793e\u533a Docker \u955c\u50cf\u51fa\u6765\u540e\uff0c\u6211\u5982\u679c\u771f\u53bb\u8bd5\u4e86\uff0c\u518d\u8865\u4e00\u7bc7\u5b9e\u64cd\u8bb0\u5f55\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>NVIDIA \u5f00\u6e90 Nemotron 3.5 ASR Streaming\uff0c\u4e00\u4e2a 600M \u53c2\u6570\u7684\u6d41\u5f0f\u591a\u8bed\u8a00\u8bed\u97f3\u8bc6\u522b\u6a21\u578b\uff0c\u652f\u6301 40 \u79cd\u8bed\u8a00\u5b9e\u65f6\u8f6c\u5f55\uff0c\u53ef\u672c\u5730\u90e8\u7f72\u3002<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[],"tags":[],"class_list":["post-23","post","type-post","status-publish","format-standard","hentry"],"_links":{"self":[{"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/posts\/23","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/comments?post=23"}],"version-history":[{"count":0,"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/posts\/23\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/media?parent=23"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/categories?post=23"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/tags?post=23"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}