NVIDIA published Nemotron 3.5 Transcribe on October 28, 2026, as an open-weight, promptable speech model on Hugging Face and ModelScope, alongside Granary v2. A single checkpoint transcribes 40 languages, translates 31 of them to or from English, and attributes speakers in 13. It handles recordings up to about one hour with up to eight speakers, supports natural-language prompts, word-level timestamps, and English-trained audio question answering. It is built for recorded audio, not live streaming.
Key Facts
| Details | |
|---|---|
| Released | October 28, 2026 (NVIDIA Hugging Face blog) |
| Distribution | Open weights on Hugging Face and ModelScope; free per TechCrunch |
| Tasks | Transcription (40 languages); EN translation (31); speaker attribution (13); audio Q&A (English-trained) |
| Limits | Recordings up to ~1 hour; up to 8 speakers; not live streaming |
| Size | ~30B total parameters; ~4B active (hybrid Mamba–Transformer MoE) |
| HF Open ASR (English, open weights) | First place; 5.33% average WER |
| Artificial Analysis (non-streaming) | About 2.8% AA-WER; HF blog says tied for first among open weights; THE DECODER cites 2.81% vs Voxtral Small 2.77% |
| Granary v2 (HF primary) | Public manifest 463,062 hours / 48 languages / 16 source corpora |
| Granary v2 (THE DECODER) | About 844,000 hours in 48 languages (differs from HF manifest figure) |
| Serving | vLLM today; quantized / NeMo-Speech.cpp coming; hosted endpoints named on the blog |
Background
Developers steer output with prompts such as speaker-labeled transcripts, English translation, or questions about decisions in a meeting. Domain terms can be listed as hints that the model should use only when the audio supports them. Language support differs by task; the blog says to spot-check names, speaker changes, and answers.
On speaker attribution, NVIDIA cites the lowest diarization error on David AI's 21-language track among 14 systems tested (18.26% DER) and first place on VoiceArena's English full-session comparison (32.4% cpWER), with deletions at 8.32% of reference words versus a median near 31%. On AA-AgentTalk (half of AA-WER), the blog reports 1.8% WER versus 2.3% for the next-best open-weight model cited there.
Granary v2 is the open curation pipeline and public manifests behind much of the training story. The Hugging Face post is explicit that the public manifest covers 463,062 hours across 48 languages and 16 corpora, up from 25 languages in Granary v1, and that these figures describe the dataset release, not Transcribe's full training mixture. THE DECODER instead reports about 844,000 hours. Until NVIDIA reconciles the two numbers in one place, treat the HF manifest figure as the primary public count and the Decoder figure as a separate secondary claim.
A medical fine-tuning case study on Eka Care's Indian clinical speech set reports held-out WER falling from 22.44% with the best prompt to 12.17% after adaptation (45.8% relative), using mostly synthetic speech plus some real medical audio, in about 14 GPU-hours on one GB300. TechCrunch frames the release for business meetings and medical transcription and notes that Transcribe does not do livestreams.
Reactions and Impact
TechCrunch places the launch in a busy week for voice AI, citing Mistral Voxtral 3, ElevenLabs Realtime, and Salesforce Launa-Omni. THE DECODER emphasizes the open weights, the single-GPU active-parameter footprint, and the medical adaptation numbers, while repeating that audio Q&A was trained in English and that the model is recording-only.
What We Don't Know Yet
- Why Granary v2 hour counts differ between the HF public manifest (463,062) and THE DECODER (~844,000)
- When NVFP4 quantized builds and local NeMo-Speech.cpp packaging ship
- How AA-WER rankings shift once more labs re-run the same non-streaming suite
- License and commercial-use terms beyond "open weight" as summarized by secondary outlets (check the model card before production use)
Sources
- NVIDIA Nemotron 3.5 Transcribe and Granary v2 — NVIDIA on Hugging Face, October 28, 2026 (primary)
- Voice AI in vogue: Nvidia releases new open transcription model — TechCrunch, October 28, 2026
- Nvidia's open Nemotron 3.5 Transcribe combines transcription, translation, and speaker recognition — THE DECODER, October 28, 2026