Nvidia strengthens audio transcription with a unified open-weight model
Nvidia has introduced Nemotron 3.5 Transcribe, an open-weight model that combines transcription, translation, diarization, and audio question-answering. At the same time, the company releases Granary v2, an open dataset of 463,062 hours.
On October 28, 2026, Nvidia unveiled Nemotron 3.5 Transcribe, an open-weight audio model that integrates multilingual transcription, speaker attribution, translation, and audio question-answering into a single checkpoint. The goal is clear: to simplify uses that until now often required multiple software components. At the same time, the company releases Granary v2, an open dataset and curation pipeline covering 463,062 hours in 48 languages.
A single model for multiple audio tasks
Nemotron 3.5 Transcribe transcribes 40 languages, translates 31 to or from English, and identifies speakers in 13 languages, according to the model card published by Nvidia on Hugging Face. The system accepts one-hour recordings and can attribute up to eight voices. In other words, it targets broader uses than simple dictation, with a model designed for minutes, meetings, or customer support.
Technically, the model combines a voice encoder, a language backbone, and a temporal classification head for word-level timestamps. Nvidia also indicates a context of 65,536 tokens and quantized variants in BF16, FP8, and NVFP4. The whole remains an offline model: internal chunk processing does not turn it into a real-time streaming service. A small but important detail, especially for integrators eager to tick all the boxes.
Granary v2, the discreet but central piece
The second announcement of the day is Granary v2. It is not a consumer product but the database and preparation that feed the training. Its public manifesto covers 463,062 hours spread over 48 languages and 16 source corpora, compared to 25 languages in Granary v1. Nvidia specifies that these figures describe the published dataset, not the entire training mix of the model.
The curation pipeline, built with NVIDIA NeMo Curator, generates multiple versions of each transcription, subtitles, question-answer pairs, contextual bias terms, and translations with quality scores. The pipeline works in four steps, with hallucination filters and agreement checks between models. On VoxPopuli, it reduced 354,000 hours to 111,000 hours in 22 European languages, proving that sorting can also clean up audio data.
Solid scores, but to be read with scope in mind
Nvidia claims that Transcribe ranks first among open-weight models on the English Open ASR leaderboard on Hugging Face, with a 5.33% average word error rate, excluding API models. The company also claims a 4.62% average WER on seven languages in multilingual mode and 10.98% on long formats, always in the open model comparison included in the card.
In its own tables, captured on October 12, 2026, Nvidia reports an English average of 4.71% WER on nine datasets, compared to 4.94% for Qwen3-ASR-1.7B. For diarization, the model shows a 15.80 average cpWER multi-speaker rate versus 16.83 for DiCoW v3.3 paired with DiariZen. In the multilingual speech field, the company also places itself at 4.62% average WER on seven languages [to be verified], with comparisons strongly dependent on the chosen protocol.
What Nvidia aims to prove with the medical version
To illustrate adaptation, Nvidia cites a medical use case based on 3,619 recordings of Indian doctors and medical students. In this scenario, specifying “English” and the Latin alphabet in the prompt lowers the WER from 30.8% to 23.8%, then fine-tuning brings it down to 11.8%. On 2,764 clips kept for evaluation, the company says it reduced the WER from 22.44% to 12.17%, a relative decrease of 45.8%.
These results remain highly domain-dependent. The LibriSpeech test-clean WER slightly increases from 1.54% to 1.69%, reminding us that a model optimized for one context does not necessarily improve everything else. In medical, accuracy matters more than hype; this is also where these systems are judged, not just on a well-lit slide.
Deployment, speed, and industrial reality
Nvidia says the model can run on a single GPU, from Jetson AGX Thor and DGX Spark to Blackwell systems in datacenters. With the optimized NIM implementation, an RTX PRO 6000 Blackwell Server Edition would process up to 2,238 seconds of audio per second in NVFP4 precision with DFlash2, according to the announcement. The company also claims to process 100 hours of audio in under three minutes on 30-second clips under heavy concurrency.
These performances are not yet part of the open-source version, the model card specifies. Execution runs through vLLM 0.32.0 with the NeMo Speech plugin or locally via NeMo-Speech.cpp, while a fine-tuning agent can automate the adaptation recipe. The offering thus resembles a fairly complete set for teams wanting to test, deploy, and adjust without rebuilding the entire pipeline. This is often where the market decides.
Key points
- October 28, 2026: release of Nemotron 3.5 Transcribe.
- 463,062 hours: public manifesto of Granary v2.
- 40 languages transcribed, 31 translated, 13 diarization.
- 5.33% WER: English Open ASR leaderboard on Hugging Face.
- 18.26% DER: best score cited on DAI-ASR-I18N.
- Nvidia and Baseten, DeepInfra, Simplismart offer endpoints.
By the numbers
- 34 billion parameters — size of the Transcribe checkpoint, according to Nvidia.
- 1.72 million audio hours — unique indexed volume used for training.
- 5.6 billion text tokens — text data associated with training.
- 3,619 recordings — Eka Care medical benchmark used for adaptation.
- 14 GPU-hours — announced time for medical fine-tuning on a NVIDIA GB300.