NVIDIA 发布 Nemotron 3.5 Transcribe 多语种转录模型与 Granary v2 数据集Chinese only
NVIDIA 的开放权重模型 Nemotron 3.5 Transcribe 已上线 Hugging Face 和 ModelScope，可转录 40 种语言，并在同一模型中提供说话人标注、翻译和音频问答。同步开放的 Granary v2 数据集覆盖 48 种语言、463,062 小时音频；一项医疗语音适配案例中，微调后对未参与调参的说话人，词错误率较最佳提示词降低 45.8%。模型针对录音设计，不支持实时流式转录。
1 developments2 reports2 sources3 verified facts
Verified facts 3
Voice AI in vogue: Nvidia releases new open transcription model
英伟达发布开放权重的音频转录模型Nemotron 3.5 Transcribe，支持转录最多40种语言的音频。
Voice AI in vogue: Nvidia releases new open transcription model
英伟达表示，Nemotron 3.5能处理多语言翻译、长达一小时的录音及说话人区分，并称其基准测试表现优于竞争对手。
Event developments 1
NVIDIA releases Nemotron 3.5 Transcribe and the Granary v2 multilingual dataset
NVIDIA’s open-weight Nemotron 3.5 Transcribe is available on Hugging Face and ModelScope. It transcribes 40 languages and combines speaker attribution, translation and audio question answering in one model. The accompanying Granary v2 dataset covers 463,062 hours across 48 languages. In a medical speech adaptation case study, fine-tuning reduced word error rate on held-out speakers by 45.8% relative to the best prompt. The model is built for recorded audio, not live streaming.
View source reports · 2 sources · 2 reports
- Hugging Face 官方博客 · NVIDIA releases Nemotron 3.5 Transcribe and the Granary v2 multilingual dataset
- TechCrunch AI 报道 · Nvidia releases open-weight transcription model and updates speech dataset
All reports 2
Nvidia releases open-weight transcription model and updates speech dataset
Nvidia released Nemotron 3.5 Transcribe, a free, open-weight model that transcribes recordings in up to 40 languages, alongside version 2 of its Granary speech dataset. Nvidia says the model translates between English and 31 languages and distinguishes up to eight speakers, but it works with recorded audio, not live streams. Mistral also shipped Voxtral 3, ElevenLabs launched Realtime, and Salesforce released Launa-Omni.
1 more source reported this
- Hugging Face 官方博客
NVIDIA releases Nemotron 3.5 Transcribe and the Granary v2 multilingual dataset
NVIDIA’s open-weight Nemotron 3.5 Transcribe is available on Hugging Face and ModelScope. It transcribes 40 languages and combines speaker attribution, translation and audio question answering in one model. The accompanying Granary v2 dataset covers 463,062 hours across 48 languages. In a medical speech adaptation case study, fine-tuning reduced word error rate on held-out speakers by 45.8% relative to the best prompt. The model is built for recorded audio, not live streaming.
1 more source reported this
- TechCrunch AI