NVIDIA releases an open speech model that transcribes 40 languages

NVIDIA has released Nemotron 3.5 Transcribe, an open-weight speech model that combines transcription, speaker attribution, translation and audio question answering in one checkpoint. It can transcribe 40 languages, translate 31 of them to or from English, identify speakers in 13 languages and process recordings up to an hour long with as many as eight speakers.

The model pairs a speech encoder with a 30-billion-parameter mixture-of-experts language model, but activates only about 4 billion parameters per token. On Artificial Analysis’s independent non-streaming speech benchmark, it ties Mistral’s Voxtral Small for the best open-weight result at 2.8% word error rate while using roughly one-sixth as many active parameters. NVIDIA also reports leading open-model results on Hugging Face’s English ASR leaderboard and two speaker-attribution evaluations.

For builders, one prompt can request a verbatim transcript, speaker labels, a translation or answers about a recording. The model can run on a single NVIDIA GPU and is available through Hugging Face, ModelScope, vLLM and NeMo-Speech.cpp. It is designed for recorded audio, not live streaming, and NVIDIA advises checking names, speaker changes and generated answers against the source recording.