MetaPatchAI home
AI

NVIDIA Nemotron 3.5 Transcribe: An Open Speech Model That Transcribes 40 Languages and Labels Who Said What

By MetaPatchAI Editorial Team · Published · Updated

Oct 30, 2026: NVIDIA released Nemotron 3.5 Transcribe on Wednesday, Oct. 28, an open-weight speech model that combines transcription, speaker labels, translation and audio question answering in one checkpoint. It is available on Hugging Face and ModelScope, alongside an open data pipeline called Granary v2.

What it does

According to NVIDIA's announcement on the Hugging Face blog, Nemotron 3.5 Transcribe:

  • Transcribes 40 languages, translates 31 of them to or from English, and attributes speakers in 13.
  • Takes natural-language prompts, so one recording can produce different outputs: "transcribe and label each speaker," a translation, or a transcript that spells domain terms correctly when you list them in the prompt.
  • Supports word-level timestamps.
  • Is built for recorded audio, not live streaming.

NVIDIA says it can be served with vLLM today, with quantized versions and local deployment through NeMo-Speech.cpp.

The numbers NVIDIA reports

These are NVIDIA's own figures from its post:

  • Hugging Face Open ASR Leaderboard (English): first among open-weight models at 5.33% average word error rate (WER). Multilingual: 4.62% mean WER across seven languages. Long-form: 10.98%.
  • Artificial Analysis: tied for first among open-weight models at 2.8% AA-WER, with about 4B active parameters versus 24B for the model it ties.
  • Speaker attribution: 18.26% diarization error rate across 21 languages on David AI's leaderboard, the lowest of 14 systems tested, and 32.4% cpWER on VoiceArena's 14-system English benchmark.
  • Speed: 100 hours of audio in under three minutes on one RTX PRO 6000 Blackwell Server Edition GPU, using a forthcoming NVFP4 quantized version, measured on 30-second clips at high concurrency.

What it means for creators and gamers

For streamers, YouTubers and community managers, an open model that captions recorded audio and tags speakers in one pass is directly useful: subtitles for VODs and clips, searchable transcripts of podcast or Discord recordings, and multi-language captions. Because the weights are open, it can run on your own hardware instead of a paid API. NVIDIA's own advice applies: spot-check names, speaker changes and answers against the recording.

It is not a live captioning tool. NVIDIA says it targets recorded audio, so real-time stream captions are out of scope for now.

Granary v2

NVIDIA also released Granary v2, the multilingual data and curation pipeline used in training, with NeMo Speech fine-tuning recipes. In a medical case study, NVIDIA says fine-tuning cut word error rate on held-out speakers by 45.8%. For the VoxPopuli dataset, its high-fidelity filter reduced 354,000 hours to 111,000 hours across 22 European languages.

What's confirmed / not confirmed

Confirmed (NVIDIA, Oct. 28, 2026): - Open-weight release on Hugging Face and ModelScope; Granary v2 data and pipeline released. - Language coverage (40 / 31 / 13) and the leaderboard placements listed above, as reported by NVIDIA.

Not confirmed: - Independent replication of the speed figure, which used a quantized version NVIDIA calls "forthcoming." - Exact license terms and the consumer-GPU memory needed (not covered in the post; check the model card).

Sources

Up: AI