Oriel — AI & technology worth noticing, every day.

Nvidia Nemotron 3.5 Transcribe: open ASR for 40 languages

Web page · https://huggingface.co/blog/nvidia/nemotron-transcribe-granary

First seen by Oriel on .

AI summary

NVIDIA Nemotron 3.5 Transcribe is an open-weight, promptable multilingual speech model, while Granary v2 provides data and fine-tuning workflows for transcription, diarization, translation, audio Q&A, and domain adaptation.

  • It covers transcription in 40 languages, translates 31 languages to or from English, attributes up to eight speakers across 13 languages, and processes recordings up to one hour.
  • It achieves 5.33% average English WER on the Hugging Face Open ASR Leaderboard and 18.26% diarization error rate across 21 languages.
  • The hybrid Mamba–Transformer mixture-of-experts model activates approximately 4B of 30B parameters per token; vLLM supports deployment, while local NeMo-Speech.cpp and quantized versions are forthcoming.
  • Granary v2 publishes 463,062 hours across 48 languages and 16 corpora, with transcript variants, task labels, quality metadata, translations, captions, and contextual ASR terms.
  • In an Indian medical speech case study, fine-tuning reduced held-out-speaker WER from 22.44% to 12.17%, while general English WER increased slightly from 1.54% to 1.69%.

#speech recognition #speaker diarization #granary v2 #fine-tuning

Generated by gpt-5.6-sol. Check the original for details.

Metric trends

Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained

Points

Score reported by the source.

  • Hacker News: 126, latest observation ; +119 over observed period since

Comments

Comment count reported by the source.

  • Hacker News: 23, latest observation ; +23 over observed period since

Aliases

SourceExternal IDURL
Hugging Face Bloghttps://huggingface.co/blog/nvidia/nemotron-transcribe-granaryhttps://huggingface.co/blog/nvidia/nemotron-transcribe-granary Read discussion
Hacker News50275128https://huggingface.co/blog/nvidia/nemotron-transcribe-granary Read discussion