NVIDIA 发布 Nemotron 3.5 Transcribe 开源权重语音模型，支持 40 种语言
NVIDIA Launches Open-Weight Transcription Model for 40 Languages
NVIDIA 于 10 月 28 日发布 Nemotron 3.5 Transcribe，这是一款开源权重的语音模型，可转写 40 种语言的音频、翻译语音并识别说话人，开发者可用自然语言指令控制转写格式而无需重新训练。该模型由 Transformer 语音编码器、说话人分离模型和 Nemotron 3.5 Lightning 语言模型组成，约 43 亿激活参数，发布检查点标注为 30B-A4B。NVIDIA 称其在英语 Open ASR Leaderboard 九个公开测试集上平均词错误率为 4.71%，在多语言榜单七个语言上为 4.62%，并称其为两个榜单上最好的开源权重模型；模型以 BF16、FP8、NVFP4 格式及 GGUF 文件发布，采用 OpenMDW-1.1 许可。
正文
NVIDIA released Nemotron 3.5 Transcribe on October 28, an open-weight speech model that can transcribe audio in 40 languages, translate speech and identify speakers. Developers can use natural-language instructions to control transcript formatting without retraining the model, according to NVIDIA.
The model supports verbatim transcripts, cleaned-up written text, domain-specific terminology and code-switched speech. It can process recordings of up to one hour and generate speaker-attributed word-level timestamps on request. NVIDIA notes that individual capabilities vary by language and task.
Nemotron 3.5 Transcribe combines a Transformer speech encoder and a speaker diarisation model with NVIDIA’s Nemotron 3.5 Lightning language model. The architecture has approximately 4.3 billion active parameters. Its released checkpoints use the designation 30B-A4B.
NVIDIA reported a mean word error rate (WER) of 4.71% across nine public test sets on the English Open ASR Leaderboard. Its mean WER on the multilingual Open ASR Leaderboard was 4.62%, covering seven languages. A lower WER indicates fewer transcription errors. NVIDIA describes the model as the best open-weight model on both leaderboards.
The company also reported a six-fold reduction in named-entity errors on ContextASR-Bench when the model received a list of expected entities. This feature is intended to improve the transcription of names and specialised terms.
For inference speed, NVIDIA reported throughput of up to 2,238 seconds of audio per second on a single RTX PRO 6000 Blackwell GPU. This result uses the NVFP4 checkpoint, DFlash2 speculative decoding and NVIDIA’s Nemotron 3.5 Transcribe inference software.
However, these serving optimisations are not yet part of the open-weight release. NVIDIA says the release-day vLLM implementation is expected to run at roughly half the speed of the optimised setup. Actual performance will depend on the hardware, configuration and workload, according to the company.
NVIDIA has released model checkpoints in BF16, FP8 and NVFP4 formats, along with GGUF files for NeMo-Speech.cpp. The model is available under the OpenMDW-1.1 licence, and a fine-tuning toolkit is available through NeMo Speech.
The company says developers can adapt the model to new languages, domains and tasks. Its technical report describes experiments across 17 tasks, with adaptation typically taking between 0.5 and 2.3 GPU-hours per task.