NVIDIA releases open speech model that can tell who is talking

NVIDIA has launched Nemotron 3.5 Transcribe, an open-weight speech transcription model, along with its training data pipeline Granary v2. The model combines transcription, speaker labeling, translation, and audio Q&A. It is built for meeting notes, class recordings, interviews, and multilingual customer service.

The model uses a mixture-of-experts design with about 34 billion total parameters. It handles recordings up to one hour long, transcribes 40 languages, and can tell apart up to 8 speakers. It labels who said what in the text, even when people talk over each other.

In tests, it topped the Hugging Face English speech leaderboard for open models with a 5.33% average word error rate. It also logged the lowest error rate, 18.26%, on a 21-language speaker separation benchmark. The model is free to use.

The Granary v2 pipeline covers 463,000 hours of audio across 48 languages. It turns raw audio into transcripts, translations, and subtitles for training speech models.