NVIDIA open-sources Nemotron 3.5 Transcribe: 40-language speech model topping Open ASR
realmrfakename · x · 2026-10-28
NVIDIA released Nemotron 3.5 Transcribe, its best speech model to date, fully open-sourcing weights, code, recipes, agentic skills, and the Granary-v2 dataset.
- SpeechLLM architecture backed by Nemotron 3.5 Lightning, running offline.
- Supports 40 languages and up to 8 speakers, with steered, promptable, contextual transcription and word- and speaker-level timestamps.
- Over 2000 RTFx with NVFP4 + DFlash2 on Blackwell GPUs.
- #1 open model on Open ASR, tied #1 overall on AA-WER, and #1 open on David AI multispeaker ASR.
Related event: NVIDIA Open-Sources Nemotron 3.5 Transcribe 30B Speech Model(10 posts)→