Nvidia open-sourced a 34B speech model that transcribes 30 seconds of audio in 0.13 seconds and posts 4.71% average English WER, beating Alibaba's Qwen3-ASR-1.7B at 4.94%. It supports 40 languages and

What happened: Nvidia released and open-sourced the speech model Nemotron 3.5 Transcribe, which combines transcription, multi-speaker recognition and multilingual processing in one model. Key numbers: About 34 billion total parameters, about 4.3 billion activated per inference. Why it matters: Nvidia positions the model as reducing the complexity of combining multiple speech tools, as user demand shifts from plain transcription to more complex tasks. Performance in real business and multilingual scenarios still awaits developer feedback.

Nvidia released and open-sourced the speech model Nemotron 3.5 Transcribe, which combines transcription, multi-speaker recognition and multilingual processing in one model.

Nvidia positions the model as reducing the complexity of combining multiple speech tools, as user demand shifts from plain transcription to more complex tasks. Performance in real business and multilingual scenarios still awaits developer feedback.