The development

What happened

NVIDIA announced Nemotron 3.5 Transcribe, an open-weight, promptable model for recorded speech, alongside the Granary v2 multilingual data-curation pipeline and public manifests. The checkpoint combines transcription, speaker attribution, translation and audio question answering; it is not intended for live streaming.

Key facts

  • NVIDIA says the model transcribes 40 languages, translates 31 to or from English and attributes speakers in 13 languages.

  • The model handles recordings up to one hour and labels up to eight speakers. Its 30 billion parameters include about 4 billion active per token.

  • Granary v2 provides manifests for 463,062 hours across 48 languages and 16 source corpora; these figures do not describe the model's full training mixture.

  • The article reports 5.33% average English word error rate on Hugging Face Open ASR and 2.80% AA-WER on Artificial Analysis.

  • The checkpoint can be served with vLLM. As of 30 October, the source describes quantized local deployment and NeMo-Speech.cpp support as available, while still calling the NVFP4 version used in its throughput example forthcoming.

The context

Why it matters

The release combines several recorded-speech tasks in one adaptable model and makes a multilingual curation pipeline available for reuse. Reported benchmark scores do not establish accuracy across all languages, accents or domains; users should check names, speaker changes and answers against the audio.