{"id":31812,"date":"2026-10-30T11:55:24","date_gmt":"2026-10-30T03:55:24","guid":{"rendered":"http:\/\/139.9.1.231\/?p=31812"},"modified":"2026-10-30T11:55:26","modified_gmt":"2026-10-30T03:55:26","slug":"nemotron-3-5-transcribe","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2026\/10\/30\/nemotron-3-5-transcribe\/","title":{"rendered":"Nemotron 3.5 Transcribe \u6df1\u5ea6\u62c6\u89e3\uff1aCache-Aware FastConformer-RNNT \u4e0e Granary v2 \u8bed\u97f3\u6570\u636e\u5de5\u5382"},"content":{"rendered":"\n<ul class=\"has-light-pink-background-color has-background\"><li><strong><em><a href=\"https:\/\/github.com\/NVIDIA-NeMo\/Speech\">https:\/\/github.com\/NVIDIA-NeMo\/Speech<\/a><\/em><\/strong><\/li><li><strong><em><a href=\"https:\/\/huggingface.co\/blog\/nvidia\/nemotron-transcribe-granary\">https:\/\/huggingface.co\/blog\/nvidia\/nemotron-transcribe-granary<\/a><\/em><\/strong><\/li><li><strong><em><a href=\"https:\/\/github.com\/NVIDIA-NeMo\/Curator\/tree\/main\/tutorials\/audio\/granary_v2\">https:\/\/github.com\/NVIDIA-NeMo\/Curator\/tree\/main\/tutorials\/audio\/granary_v2<\/a><\/em><\/strong><\/li><\/ul>\n\n\n\n\\(P(y_{1:T}\\mid x)=\\prod_{t=1}^{T}P(y_t\\mid y_{&lt;t},x)\\)\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"487\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/10\/image-14-1024x487.png\" alt=\"\" class=\"wp-image-31814\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/10\/image-14-1024x487.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/10\/image-14-300x143.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/10\/image-14-768x365.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/10\/image-14.png 1413w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<h2>\u4e00\u3001\u6838\u5fc3\u7ed3\u8bba<\/h2>\n\n\n\n<ul><li>Nemotron 3.5 ASR Streaming \u662f\u7ea6 0.6B \u53c2\u6570\u7684\u591a\u8bed\u8a00\u6d41\u5f0f ASR\u3002<\/li><li>\u6838\u5fc3\u67b6\u6784\u662f 24 \u5c42 Cache-Aware FastConformer \u7f16\u7801\u5668\u3001RNN-T \u89e3\u7801\u5668\u548c\u8bed\u8a00 ID Prompt Conditioning\u3002<\/li><li>\u58f0\u5b66\u8868\u793a\u7ef4\u5ea6\u4e3a 1024\uff0c\u8bed\u8a00\u5411\u91cf\u4e3a 128 \u7ef4\uff1b\u4e00\u4e2a checkpoint \u652f\u6301 40 \u4e2a language locale\u3002<\/li><li>\u8fd0\u884c\u65f6\u652f\u6301 80 ms\u3001160 ms\u3001320 ms\u3001560 ms\u30011.12 s \u4e94\u4e2a chunk \u914d\u7f6e\u3002<\/li><li>FLEURS \u4e0a\uff0cTranscription-ready 19 \u4e2a locale \u7684\u5e73\u5747 LangID WER \u4ece 80 ms \u7684 10.38% \u964d\u5230 1.12 s \u7684 8.84%\uff1bAuto-detect \u4ece 11.14% \u964d\u5230 9.21%\u3002<\/li><li>\u5355\u5f20 H100 \u4e0a\uff0c80 ms \u65f6\u7ea6\u652f\u6301 240 \u8def\u5e76\u53d1\u6d41\uff0cParakeet RNNT 1.1B buffered streaming \u7ea6 14 \u8def\uff1b1.12 s \u65f6\u7ea6\u4e3a 2,400 \u8def\u5bf9 400 \u8def\u3002<\/li><\/ul>\n\n\n\n<p>\u8bbe\u8ba1\u76ee\u6807\u4e0d\u662f\u53ea\u8ffd\u6c42\u79bb\u7ebf WER\uff0c\u800c\u662f\u5728\u51c6\u786e\u7387\u3001\u7aef\u5230\u7aef\u5ef6\u8fdf\u3001\u663e\u5b58\u548c\u5e76\u53d1\u4e4b\u95f4\u63d0\u4f9b\u53ef\u8c03\u7684 Pareto \u70b9\u3002<\/p>\n\n\n\n<h2>\u4e8c\u3001\u6a21\u578b\u67b6\u6784<\/h2>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"503\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/10\/image-15-1024x503.png\" alt=\"\" class=\"wp-image-31815\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/10\/image-15-1024x503.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/10\/image-15-300x147.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/10\/image-15-768x377.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/10\/image-15.png 1459w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<h3>2.1 Cache-Aware FastConformer<\/h3>\n\n\n\n<p>FastConformer \u7ed3\u5408\u5377\u79ef\u5bf9\u5c40\u90e8\u58f0\u5b66\u6a21\u5f0f\u7684\u5efa\u6a21\u80fd\u529b\u548c\u6ce8\u610f\u529b\u5bf9\u957f\u7a0b\u4f9d\u8d56\u7684\u5efa\u6a21\u80fd\u529b\u3002Nemotron 3.5 \u4f7f\u7528 24 \u5c42 Cache-Aware FastConformer\u3002\u7f13\u5b58\u540c\u65f6\u8986\u76d6 self-attention \u548c convolution \u72b6\u6001\uff0c\u56e0\u6b64\u5f53\u524d chunk \u53ef\u4ee5\u590d\u7528\u4e0a\u4e00 chunk \u7684\u9690\u85cf\u72b6\u6001\u3002<\/p>\n\n\n\n<p>\u8bbe\u7b2c k \u4e2a\u97f3\u9891\u5757\u4e3a x_k\uff0c\u4e0a\u4e00\u6b65\u7f13\u5b58\u4e3a C_(k-1)\uff1a<\/p>\n\n\n\n\\((H_k, C_k) = Encoder(x_k, C_(k-1))\\)\n\n\n\n<p>\u4f20\u7edf buffered streaming \u4f1a\u628a\u4e0a\u4e00\u6bb5\u4e0a\u4e0b\u6587\u4e0e\u5f53\u524d\u97f3\u9891\u91cd\u65b0\u62fc\u63a5\uff0c\u5bfc\u81f4\u91cd\u53e0\u5e27\u91cd\u590d\u8ba1\u7b97\uff1bCache-Aware \u65b9\u6848\u53ea\u8ba1\u7b97\u65b0\u5230\u8fbe\u7684\u5e27\uff0c\u7f13\u5b58\u8d1f\u8d23\u4fdd\u7559\u4e0a\u4e0b\u6587\u3002<\/p>\n\n\n\n<h3>2.2 \u8bed\u8a00 ID Prompt Conditioning<\/h3>\n\n\n\n<p>\u7f16\u7801\u5668\u8f93\u51fa H_a \u7684\u5f62\u72b6\u662f (1024,T)\u3002\u8bed\u8a00 ID \u88ab\u7f16\u7801\u4e3a 128 \u7ef4\u5411\u91cf e_lang\uff0c\u6cbf\u65f6\u95f4\u8f74\u5e7f\u64ad\u4e3a H_l\uff0c\u4e8c\u8005\u62fc\u63a5\u540e\u7ecf\u8fc7\u6295\u5f71\u5c42\u9001\u5165 RNNT\uff1a<\/p>\n\n\n\n\\(H_a = FastConformer(x_1:T), H_a in R^(1024 x T)\\)\n\n\n\n\\(H_l = Broadcast(e_lang,T), e_lang in R^128\\)\n\n\n\n\\(H_f = Proj([H_a;H_l]), [H_a;H_l] in R^(1152 x T)\\)\n\n\n\n<p>\u5df2\u77e5\u8bed\u8a00\u65f6\u53ef\u4ee5\u8f93\u5165 en-US\u3001de-DE\u3001zh-CN \u7b49 locale\uff1b\u672a\u77e5\u8bed\u8a00\u65f6\u4f7f\u7528 auto \u6a21\u5f0f\u3002\u8bed\u8a00\u5411\u91cf\u662f\u6bcf\u4e2a\u65f6\u95f4\u5e27\u90fd\u53ef\u89c1\u7684\u6761\u4ef6\u4fe1\u53f7\uff0c\u800c\u4e0d\u662f\u9644\u52a0\u5728\u6700\u7ec8\u6587\u672c\u540e\u7684\u6807\u7b7e\u3002<\/p>\n\n\n\n<h3>2.3 RNN-T<\/h3>\n\n\n\n<p>RNN-T \u7528 prediction network \u5efa\u6a21\u5df2\u8f93\u51fa\u7684 token \u5386\u53f2\uff0c\u7528 joint network \u878d\u5408\u58f0\u5b66\u72b6\u6001\u4e0e\u6807\u7b7e\u72b6\u6001\uff1a<\/p>\n\n\n\n\\(z_(t,u) = f_joint(h_t,g_u), P(y_(t,u)|x,y_&lt;u) = softmax(z_(t,u))\\)\n\n\n\n<p>\u5b83\u80fd\u5728\u97f3\u9891\u6d41\u6301\u7eed\u5230\u8fbe\u65f6\u9010\u6b65\u53d1\u5c04 token\uff0c\u6bd4\u5fc5\u987b\u7b49\u5f85\u6574\u6bb5\u97f3\u9891\u7684\u79bb\u7ebf encoder-decoder \u66f4\u9002\u5408\u5b9e\u65f6\u8bed\u97f3\u4ee3\u7406\u3002<\/p>\n\n\n\n<h2>\u4e09\u3001\u521b\u65b0\u70b9<\/h2>\n\n\n\n<ul><li>\u7f13\u5b58\u7ea7\u6d41\u5f0f\u8ba1\u7b97\uff0c\u51cf\u5c11 buffered streaming \u7684\u91cd\u53e0\u91cd\u7b97\u3002<\/li><li>\u8bed\u8a00\u63d0\u793a\u8ba9\u4e00\u4e2a\u5171\u4eab checkpoint \u8986\u76d6\u591a\u8bed\u8a00 locale\u3002<\/li><li>chunk size \u53ef\u5728\u63a8\u7406\u65f6\u5207\u6362\uff0c\u4e0d\u9700\u8981\u91cd\u65b0\u8bad\u7ec3\u3002<\/li><li>\u539f\u751f\u652f\u6301 punctuation \u548c capitalization\u3002<\/li><li>0.6B \u53c2\u6570\u914d\u5408\u7f13\u5b58\u5f0f\u63a8\u7406\uff0c\u5728 H100 \u4e0a\u83b7\u5f97\u9ad8\u5e76\u53d1\u3002<\/li><\/ul>\n\n\n\n<h2>\u56db\u3001\u8bad\u7ec3\u6570\u636e\uff1a\u89c4\u6a21\u3001\u6765\u6e90\u4e0e\u6807\u7b7e<\/h2>\n\n\n\n<p>\u5b98\u65b9\u6a21\u578b\u5361\u7ed9\u51fa\u7684\u8bad\u7ec3\u6570\u636e\u8303\u56f4\u662f 10,000 \u5230 1,000,000 \u5c0f\u65f6\uff0c\u8986\u76d6 40 \u4e2a language locale\u3002\u6765\u6e90\u5305\u62ec NVIDIA Riva multilingual ASR training set\uff08\u4e13\u6709\uff09\u3001NVIDIA Granary\u3001Multilingual LibriSpeech\u3001Mozilla Common Voice\u3001FLEURS\u3001VoxPopuli \u548c Europarl-ASR\u3002<\/p>\n\n\n\n<p>\u8bad\u7ec3\u6807\u7b7e\u5305\u62ec\u4eba\u5de5\u6807\u7b7e\u548c\u5408\u6210\u6807\u7b7e\u3002\u5408\u6210 ASR \u6807\u7b7e\u6765\u81ea Canary\u3001Parakeet Multilingual 1.1B RNNT\u3001Parakeet CTC 1.1B\u3001Whisper large-v3 \u4e0e FunASR \u7684\u6a21\u578b\u96c6\u6210\uff1b\u6807\u70b9\u548c\u5927\u5c0f\u5199\u7531 Qwen3-32B \u751f\u6210\uff1a<\/p>\n\n\n\n\\(y_hat_i = Consensus(Canary(x_i), Parakeet(x_i), Whisper(x_i), FunASR(x_i))\\)\n\n\n\n\\(y_PnC_i = Qwen3-32B(y_hat_i, language_i)\\)\n\n\n\n<p>Consensus \u4e0d\u80fd\u88ab\u64c5\u81ea\u89e3\u91ca\u4e3a\u7b80\u5355\u591a\u6570\u6295\u7968\uff0c\u56e0\u4e3a\u5b98\u65b9\u6ca1\u6709\u62ab\u9732\u96c6\u6210\u89c4\u5219\u3002\u5de5\u7a0b\u4e0a\u8fd8\u5e94\u7ed3\u5408\u8bed\u8a00\u4e00\u81f4\u6027\u3001\u5b57\u7b26\u901f\u7387\u3001\u91cd\u590d\u68c0\u6d4b\u3001\u7a7a\u8f93\u51fa\u548c\u957f\u5ea6\u6bd4\u8fc7\u6ee4\u3002<\/p>\n\n\n\n<h3>4.1 Granary \u6570\u91cf\u7ea7<\/h3>\n\n\n\n<p>Granary \u8986\u76d6 25 \u79cd\u6b27\u6d32\u8bed\u8a00\uff0c\u805a\u5408\u89c4\u6a21\u7ea6 1M \u5c0f\u65f6\uff1bASR \u4fa7\u7ea6 643,238 \u5c0f\u65f6\uff0cAST \u4fa7\u7ea6 351,000 \u5c0f\u65f6\u3002<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table><thead><tr><th>\u6765\u6e90<\/th><th>\u8bed\u8a00\u6570<\/th><th>\u8fc7\u6ee4\u540e\u5c0f\u65f6\u6570<\/th><th>\u683c\u5f0f<\/th><\/tr><\/thead><tbody><tr><td>YODAS-Granary<\/td><td>23<\/td><td>192,172<\/td><td>16 kHz WAV<\/td><\/tr><tr><td>VoxPopuli<\/td><td>24<\/td><td>206,116<\/td><td>FLAC<\/td><\/tr><tr><td>YouTube-Commons<\/td><td>24<\/td><td>122,475<\/td><td>WAV<\/td><\/tr><tr><td>LibriLight<\/td><td>\u82f1\u8bed<\/td><td>\u7ea6 23,500<\/td><td>FLAC<\/td><\/tr><tr><td>ASR \u5408\u8ba1<\/td><td>25<\/td><td>643,238<\/td><td>\u6df7\u5408\u683c\u5f0f<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p>Granary \u4e3b\u8981\u63d0\u4f9b NeMo JSONL manifest \u548c\u5143\u6570\u636e\uff0c\u90e8\u5206\u6765\u6e90\u9700\u8981\u53e6\u884c\u4e0b\u8f7d\u97f3\u9891\u3002\u5178\u578b\u5b57\u6bb5\u5305\u62ec audio_filepath\u3001text\u3001source_lang\u3001target_lang\u3001original_source_id \u548c dataset_source\u3002<\/p>\n\n\n\n<h2>\u4e94\u3001\u8bad\u7ec3\u9636\u6bb5<\/h2>\n\n\n\n<ul><li><strong>\u9636\u6bb5 A\uff1a\u6536\u96c6\u548c\u7edf\u4e00 manifest\u3002<\/strong>\u4fdd\u7559\u8def\u5f84\u3001\u65f6\u957f\u3001\u91c7\u6837\u7387\u3001\u8bed\u8a00\u3001\u6765\u6e90\u3001\u8bb8\u53ef\u8bc1\u548c\u539f\u59cb ID\u3002<\/li><li><strong>\u9636\u6bb5 B\uff1a\u5207\u5206\u3001VAD\u3001\u8bf4\u8bdd\u4eba\u548c\u8bed\u8a00\u8d28\u68c0\u3002<\/strong>\u68c0\u67e5\u65f6\u957f\u3001\u9759\u97f3\u6bd4\u4f8b\u3001\u5b57\u7b26\u901f\u7387\u3001\u8bed\u8a00\u4e00\u81f4\u6027\u3001\u7a7a\u6587\u672c\u548c\u5f02\u5e38\u91cd\u590d\u3002<\/li><li><strong>\u9636\u6bb5 C\uff1a\u591a\u6a21\u578b\u4f2a\u6807\u6ce8\u4e0e PnC\u3002<\/strong>\u751f\u6210\u5019\u9009\u3001\u505a\u4e00\u81f4\u6027\u7b5b\u9009\uff0c\u518d\u7528 Qwen3-32B \u7edf\u4e00\u6807\u70b9\u548c\u5927\u5c0f\u5199\u3002<\/li><li><strong>\u9636\u6bb5 D\uff1a\u6d41\u5f0f\u591a\u8bed\u8a00 ASR \u8bad\u7ec3\u3002<\/strong>\u4f7f\u7528 Cache-Aware FastConformer-RNNT\u3001\u8bed\u8a00 ID \u6761\u4ef6\u548c\u6df7\u5408\u591a\u8bed\u8a00\u6570\u636e\u3002<\/li><li><strong>\u9636\u6bb5 E\uff1a\u591a\u5ef6\u8fdf\u70b9\u8bc4\u6d4b\u3002<\/strong>\u4ee5 80 ms \u4e3a\u57fa\u672c\u5e27\uff0c\u6d4b\u8bd5 80 ms\u3001160 ms\u3001320 ms\u3001560 ms\u30011.12 s\u3002<\/li><\/ul>\n\n\n\n\\(D_train = Filter(Normalize(D_public union D_private))\\)\n\n\n\n\\(theta* = argmin_theta E[-log P_theta(y | x, language)]\\)\n\n\n\n<h2>\u516d\u3001Granary v2 \u6d41\u6c34\u7ebf<\/h2>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"557\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/10\/image-16-1024x557.png\" alt=\"\" class=\"wp-image-31816\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/10\/image-16-1024x557.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/10\/image-16-300x163.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/10\/image-16-768x418.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/10\/image-16.png 1260w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>Granary v2 \u628a\u539f\u59cb\u5f55\u97f3\u52a0\u5de5\u6210\u5e26\u5143\u6570\u636e\u3001\u53ef\u8ffd\u6eaf\u3001\u53ef\u7ee7\u7eed\u5904\u7406\u7684\u8bad\u7ec3\u6570\u636e\u3002\u5b98\u65b9\u6559\u7a0b\u5305\u62ec Metadata extraction\u3001Indic ASR\u3001Qwen Omni ASR \u548c Text pipeline\u3002<\/p>\n\n\n\n<h3>6.1 Metadata extraction<\/h3>\n\n\n\n<ul><li>\u8bfb\u53d6 recording manifest\uff0c\u91cd\u91c7\u6837\u5230\u5355\u58f0\u9053 16 kHz WAV\u3002<\/li><li>Sortformer \u505a\u8bf4\u8bdd\u4eba\u5206\u79bb\u5e76\u8f93\u51fa RTTM\u3002<\/li><li>Silero TensorRT VAD \u5207\u5206\u8bed\u97f3\u6bb5\u3002<\/li><li>CNN14 TensorRT \u9884\u6d4b\u58f0\u97f3\u4e8b\u4ef6\u3002<\/li><li>SpeechBrain\u3001Indic Canary\u3001Whisper \u505a\u591a\u7ea7\u8bed\u8a00\u8bc6\u522b\u3002<\/li><li>\u6839\u636e\u4e00\u81f4\u6027\u9009\u62e9 source_lang\uff0c\u4fdd\u7559\u62d2\u7edd\u539f\u56e0\u3002<\/li><li>\u7247\u6bb5\u7f16\u7801\u4e3a Opus\uff0c\u5e76\u5199\u51fa JSONL metadata\u3002<\/li><\/ul>\n\n\n\n<p>\u6ca1\u6709\u8bed\u97f3\u7684\u5f55\u97f3\u4fdd\u7559 vad_empty \u5ba1\u8ba1\u884c\uff1b\u8bed\u8a00\u610f\u89c1\u4e0d\u4e00\u81f4\u65f6\u4fdd\u7559 skip \u6807\u8bb0\u548c\u9884\u6d4b\u7ed3\u679c\uff0c\u907f\u514d\u628a\u201c\u65e0\u8bed\u97f3\u201d\u548c\u201c\u6587\u4ef6\u7f3a\u5931\u201d\u6df7\u4e3a\u4e00\u8c08\u3002<\/p>\n\n\n\n<h3>6.2 Indic ASR<\/h3>\n\n\n\n<p>Hindi \u793a\u4f8b\u6709 11 \u4e2a\u9636\u6bb5\uff1a\u8bfb\u53d6 manifest\uff1b\u51c6\u5907\u5355\u58f0\u9053 16 kHz \u97f3\u9891\uff1bCanary \u4e3b\u8f6c\u5199\uff1b\u4e3b\u7ed3\u679c hallucination \u68c0\u67e5\uff1bParakeet \u6062\u590d\u5019\u9009\uff1b\u6062\u590d\u7ed3\u679c\u68c0\u67e5\uff1b\u9009\u62e9\u6700\u4f73\u9884\u6d4b\uff1b\u6b63\u5219\u6e05\u6d17\uff1b\u7f29\u5199\u5408\u5e76\uff1b\u8ba1\u7b97 WER\uff1b\u5199\u56de JSONL\u3002\u4e24\u4e2a ASR \u6a21\u578b\u4f1a\u5bf9\u6bcf\u4e2a\u6709\u6548\u7247\u6bb5\u90fd\u8fd0\u884c\u3002<\/p>\n\n\n\n<p>\u793a\u4f8b\u9650\u5236 0.5 \u5230 40 \u79d2\uff1b\u6700\u7ec8\u5b57\u6bb5\u5305\u62ec primary_model_prediction\u3001fallback_model_prediction\u3001best_prediction\u3001cleaned_text\u3001abbreviated_text \u548c wer_pct\u3002<\/p>\n\n\n\n<h3>6.3 Qwen Omni ASR \u4e0e Text pipeline<\/h3>\n\n\n\n<p>Qwen3-Omni \u751f\u6210\u4e3b\u8f6c\u5199\uff0c\u5df2\u6709 transcript \u53ea\u4f5c\u4e3a\u56de\u9000\u6587\u672c\uff0c\u4e0d\u53d1\u9001\u7ed9 Qwen\u3002\u968f\u540e\u6267\u884c\u89c4\u5219\u578b hallucination \u68c0\u67e5\u3001\u6700\u4f73\u9884\u6d4b\u9009\u62e9\u3001\u6807\u70b9\u548c\u7a7a\u767d\u6e05\u7406\u3001\u7f29\u5199\u5408\u5e76\u3002Text pipeline \u4e0d\u8bfb\u53d6\u97f3\u9891\uff0c\u8d1f\u8d23\u5b9e\u4f53\u6062\u590d\u3001PnC\u3001\u8bed\u8a00\u8bc6\u522b\u4e0e\u9a8c\u8bc1\u3001\u9006\u6587\u672c\u89c4\u8303\u5316\u3001captioning\u3001code switching \u548c speech QA\u3002<\/p>\n\n\n\n<h3>6.4 Xenna \u4e0e Ray Data \u541e\u5410<\/h3>\n\n\n\n<figure class=\"wp-block-table\"><table><thead><tr><th>\u7ba1\u9053<\/th><th>\u57fa\u51c6\u6570\u636e<\/th><th>Xenna<\/th><th>Ray Data<\/th><\/tr><\/thead><tbody><tr><td>Metadata<\/td><td>30,631 \u5f55\u97f3 \/ 107.76357 \u5c0f\u65f6<\/td><td>472.309 audio-hours\/hour<\/td><td>166.251<\/td><\/tr><tr><td>Indic ASR<\/td><td>216,164 Hindi \u7247\u6bb5 \/ 531.75763 \u5c0f\u65f6<\/td><td>3,604.471<\/td><td>387.649<\/td><\/tr><tr><td>Qwen Omni ASR<\/td><td>22,710 \u82f1\u8bed\u7247\u6bb5 \/ 80.61259 \u5c0f\u65f6<\/td><td>406.615<\/td><td>228.424<\/td><\/tr><tr><td>Text pipeline<\/td><td>14,290 \u884c transcript<\/td><td>30.283 rows\/s<\/td><td>12.217 rows\/s<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p>\u5b98\u65b9\u5bf9\u6bd4\u4f7f\u7528 8 \u5f20 H100 80 GB GPU\u3001\u76f8\u540c cohort\u3001\u955c\u50cf\u3001\u6a21\u578b\u548c stage \u8bbe\u7f6e\u3002\u97f3\u9891\u7ba1\u9053\u541e\u5410\u5305\u542b\u6a21\u578b\u542f\u52a8\u65f6\u95f4\uff1bText pipeline \u590d\u7528\u4e86\u5df2\u542f\u52a8\u7684\u6a21\u578b\u670d\u52a1\u5668\u3002<\/p>\n\n\n\n<h2>\u4e03\u3001\u5b9e\u9a8c\u5bf9\u6bd4\uff1aFLEURS \u4e0e\u6d41\u5f0f\u5ef6\u8fdf<\/h2>\n\n\n\n<figure class=\"wp-block-table\"><table><thead><tr><th>\u5206\u7ec4<\/th><th>\u6a21\u5f0f<\/th><th>80 ms<\/th><th>160 ms<\/th><th>320 ms<\/th><th>560 ms<\/th><th>1.12 s<\/th><\/tr><\/thead><tbody><tr><td>Transcription-ready 19 locales<\/td><td>LangID \u5e73\u5747 WER<\/td><td>10.38<\/td><td>10.00<\/td><td>9.49<\/td><td>9.12<\/td><td>8.84<\/td><\/tr><tr><td>Transcription-ready 19 locales<\/td><td>Auto-detect \u5e73\u5747 WER<\/td><td>11.14<\/td><td>10.67<\/td><td>10.05<\/td><td>9.63<\/td><td>9.21<\/td><\/tr><tr><td>Broad-coverage 13 locales<\/td><td>LangID \u5e73\u5747 WER<\/td><td>25.86<\/td><td>25.14<\/td><td>24.11<\/td><td>23.09<\/td><td>22.13<\/td><\/tr><tr><td>Broad-coverage 13 locales<\/td><td>Auto-detect \u5e73\u5747 WER<\/td><td>27.62<\/td><td>26.65<\/td><td>25.41<\/td><td>24.44<\/td><td>23.09<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p>\u5728 1.12 s\u3001LangID \u6761\u4ef6\u4e0b\uff0c\u5b98\u65b9\u8868\u4e2d\u82f1\u8bed WER \u4e3a 7.91%\uff0c\u897f\u73ed\u7259\u8bed 4.11%\uff0c\u610f\u5927\u5229\u8bed 4.25%\uff0c\u8461\u8404\u7259\u8bed 5.48%\uff0cHindi 6.81%\uff0c\u5fb7\u8bed 8.31%\uff0c\u6cd5\u8bed 9.03%\uff0c\u4fc4\u8bed 9.17%\uff0c\u4e4c\u514b\u5170\u8bed 13.07%\u3002Hindi \u5728 80 ms \u4e0b LangID \u4e3a 8.13%\uff0cAuto-detect \u4e3a 11.47%\uff1b1.12 s \u4e0b\u5206\u522b\u4e3a 6.81% \u548c 8.23%\u3002\u65e5\u8bed\u3001\u97e9\u8bed\u548c\u666e\u901a\u8bdd\u4f7f\u7528 CER\u3002<\/p>\n\n\n\n<p>chunk \u589e\u5927\u901a\u5e38\u5e26\u6765\u66f4\u591a\u53f3\u4fa7\u4e0a\u4e0b\u6587\uff0c\u5e73\u5747 WER \u4e0b\u964d\u4f46\u7b49\u5f85\u65f6\u95f4\u4e0a\u5347\uff1a<\/p>\n\n\n\n\\(WER(c_1) &gt;= WER(c_2), c_1 &lt; c_2\\)\n\n\n\n<h2>\u516b\u3001\u541e\u5410\u5bf9\u6bd4\u4e0e\u7cfb\u7edf\u610f\u4e49<\/h2>\n\n\n\n<figure class=\"wp-block-table\"><table><thead><tr><th>\u914d\u7f6e<\/th><th>Nemotron 3.5 ASR<\/th><th>Parakeet RNNT 1.1B<\/th><th>\u4f18\u52bf<\/th><\/tr><\/thead><tbody><tr><td>80 ms<\/td><td>\u7ea6 240 \u8def<\/td><td>\u7ea6 14 \u8def<\/td><td>\u7ea6 17 \u500d<\/td><\/tr><tr><td>1.12 s<\/td><td>\u7ea6 2,400 \u8def<\/td><td>\u7ea6 400 \u8def<\/td><td>\u7ea6 6 \u500d<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p>\u4f18\u52bf\u6765\u81ea\u53c2\u6570\u89c4\u6a21\u3001\u7f13\u5b58\u590d\u7528\u3001chunk \u7c92\u5ea6\u548c H100 \u63a8\u7406\u5b9e\u73b0\u7684\u5171\u540c\u4f5c\u7528\uff0c\u800c\u4e0d\u662f\u5355\u7eaf\u53c2\u6570\u66f4\u5c0f\uff1a<\/p>\n\n\n\n\\(Cost_per_stream proportional to GPU_hourly_cost \/ sustainable_concurrent_streams\\)\n","protected":false},"excerpt":{"rendered":"<p>https:\/\/github.com\/NVIDIA-NeMo\/Speech https:\/\/huggingfa &hellip; <a href=\"http:\/\/139.9.1.231\/index.php\/2026\/10\/30\/nemotron-3-5-transcribe\/\" class=\"more-link\">\u7ee7\u7eed\u9605\u8bfb<span class=\"screen-reader-text\">Nemotron 3.5 Transcribe \u6df1\u5ea6\u62c6\u89e3\uff1aCache-Aware FastConformer-RNNT \u4e0e Granary v2 \u8bed\u97f3\u6570\u636e\u5de5\u5382<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[42,40,4,9,38,43,34],"tags":[],"_links":{"self":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/31812"}],"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=31812"}],"version-history":[{"count":4,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/31812\/revisions"}],"predecessor-version":[{"id":31819,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/31812\/revisions\/31819"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=31812"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=31812"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=31812"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}