Nvidia ships an open speech model that transcribes 40 languages
Nemotron 3.5 Transcribe combines transcription, speaker attribution, translation and audio Q&A in one 30-billion-parameter model, under Nvidia's OpenMDW licence.
Nvidia released Nemotron 3.5 Transcribe on 28 October, an open-weight speech model that brings transcription, speaker attribution, translation and audio question answering into one checkpoint. It transcribes 40 languages, translates 31 of them to or from English, and attributes speakers in 13, with word-level timestamps included.
The model pairs a speech encoder with a hybrid Mamba-Transformer mixture-of-experts language model: 30 billion parameters total, with only about 4 billion active per token. Nvidia says it’s first among open-weight models on English transcription at 5.33% average word error rate, and ties for first on Artificial Analysis’s benchmark at 2.8% WER using roughly a sixth of the active parameters of the model it ties. It’s built for recorded audio rather than live streaming, handles recordings up to an hour long, and labels up to eight speakers. Weights are on Hugging Face and ModelScope under Nvidia’s OpenMDW-1.1 licence, with hosted endpoints also available through Baseten, DeepInfra and Simplismart.
Why it matters: this is a serious open alternative to proprietary transcription APIs, strong enough on Nvidia’s own numbers to self-host for multilingual meeting notes, call transcripts or subtitle pipelines, with diarization and translation already built in rather than bolted on as separate services.
The caveat: every benchmark figure here comes from Nvidia’s own evaluation; there’s no independent leaderboard entry yet to check them against, and it’s not a fit for anything that needs live transcription.