Nvidia released Nemotron 3.5 Transcribe and Granary v2 on 28 October 2026, giving developers an open-weight speech model they can instruct to produce transcripts, identify speakers, translate audio or answer questions about a recording. The accompanying open multilingual dataset and curation pipeline are intended to support further adaptation. Nvidia says the model transcribes 40 languages, translates 31 of them to or from English, and attributes speech to speakers in 13.
Key points
- The model accepts natural-language instructions and handles recordings of up to an hour, with labels for up to eight speakers.
- The English Hugging Face Open ASR Leaderboard places the model first among open-weight models at 5.33% average word error rate.
- Granary v2 covers 843,579 hours of speech across 48 languages and 19 source corpora.
One Nemotron checkpoint, several speech tasks
Nemotron 3.5 Transcribe uses one checkpoint for tasks that can require separate stages: writing down speech, attaching speaker labels to turns, translating, and answering a question about what was said. A developer can ask for different outputs from the same recording by changing the instruction. Nvidia says the model also supports word-level timestamps, while its audio question-answering capability was trained in English.
Taking meeting notes could therefore involve asking for a transcript with speaker labels, or asking what decisions the recording contains. A recording in another language could instead be rendered in English where translation is supported. Nvidia advises checking names, changes of speaker and answers against the audio.
The checkpoint pairs a speech encoder with a mixture-of-experts language model containing 30 billion parameters. About 4 billion are active for each token it generates: like consulting a few volumes from a large shelf for each answer, the system draws on part of its stored weights at a time. That distinction concerns computation during generation, rather than the total size of the checkpoint. Nvidia says the model runs on a single Nvidia GPU, including systems ranging from Jetson AGX Thor to Blackwell data-centre hardware.
On one RTX PRO 6000 Blackwell Server Edition GPU using NVFP4 precision, Nvidia says the model processed 100 hours of audio in under three minutes. That measurement used 30-second clips at high concurrency. It describes throughput when many short recordings are processed together, while Nvidia says the model can accept an individual recording of up to an hour. The model is built for recorded audio rather than live streaming, and Nvidia lists vLLM for serving the checkpoint and NeMo-Speech.cpp for local use.
A 5.33% result on English transcription
Nvidia reports that Nemotron 3.5 Transcribe ranks first among open-weight models on the English track of the Hugging Face Open ASR Leaderboard, at 5.33% average word error rate. That measure counts transcription mistakes against reference text on public test sets, with lower scores indicating fewer errors. API models are excluded from the ranking Nvidia cites.
The company also reports a 4.62% mean word error rate across seven languages in the model card’s multilingual Open ASR comparison, and 10.98% in its long-form comparison. On Artificial Analysis’s AA-WER index, Nvidia says the model ties for first among open-weight models at 2.8%. The company describes AA-AgentTalk, a held-out voice-agent speech set, as accounting for 50% of that index.
For speaker attribution, Nvidia reports an 18.26% diarization error rate on David AI’s evaluation across 21 languages, the lowest among 14 systems tested there. Diarization measures who spoke when, so it tests a different part of the output from a transcription score. Nvidia also reports a 32.4% score on VoiceArena’s English full-session measure, which counts word errors after matching words to speakers in sessions that include overlapping speech.
Nvidia’s figures for prompting differ across the evaluations it describes. The company reported a 30.8% error rate without prompting and 20.8% with prompting, SiliconANGLE reported. Nvidia reports a 5.33% average word error rate on the English Hugging Face Open ASR Leaderboard.
Granary v2 spans 48 languages
Granary v2, the open multilingual speech dataset released alongside the model, contains 843,579 hours across 48 languages and 19 source corpora, SiliconANGLE reported. Nvidia built its reusable curation pipeline with NeMo Curator. The reported dataset totals describe the released data rather than the model’s full training mixture or the proportions used to sample it.
The pipeline produces three transcript styles for each recording, according to SiliconANGLE: one retains filler sounds, one changes spoken numbers into written figures, and one removes filler words. That gives developers examples of the written forms they can request with a prompt. Nvidia makes Granary v2 available on Hugging Face and ModelScope alongside NeMo Speech fine-tuning tools for adapting recognition to specialised vocabulary and accents.
In a healthcare prompting example, Nvidia suggests supplying possible medical terms and instructing the model to use them only when the audio supports them. In a medical case study, the company says fine-tuning reduced word error rate on held-out speakers by 45.8% relative to the best prompt.
