On the 28th of October, Nvidia released Nemotron 3.5 Transcribe, a free speech model for recorded audio. Specifically, the model transcribes 40 languages, translates 31 of them, and labels up to eight speakers in 13 languages. Nvidia reports first place among open models on Hugging Face’s English Open ASR leaderboard.
Nvidia Releases Open Transcription Model
The company designed Nemotron 3.5 Transcribe to follow written instructions. As a result, users type a request before the model listens, and it returns a word-for-word or cleaned transcript. Users can also tell the model how to format numbers.
Likewise, developers can supply names and technical terms, which improves recognition of specialist vocabulary. Another prompt requests speaker labels, so listeners can know the person talking. Also, the model can be prompted to record, although it only records for an hour.
Alongside the model, Nvidia released Granary v2, a public dataset covering 463,062 hours of speech in 48 languages. Granary v2 gives each recording three transcript styles, so users can request one with a prompt.
In addition, developers can run the dataset’s curation pipeline on their own recordings to prepare tuning data. The model has 30 billion parameters, the values it learns during training. Only about 4 billion are active at a time, so the model stays fast and cheap to run. It runs on one Nvidia chip, from Jetson AGX Thor edge computers to Blackwell data-center GPUs.
Nvidia says a compressed setup on one Blackwell GPU processes 100 hours of audio in under three minutes. Its language model comes from Nemotron 3.5 Lightning, which the company released in August.
Benchmark Results and Trade-Offs
First, the model scores a 5.33% average word error rate on Hugging Face’s Open ASR leaderboard. The rate counts wrong, missing, or extra words, so 5.33% means roughly five errors per 100 words. The first-place ranking excludes paid API models.
Second, the company reports first place among 14 systems on VoiceArena’s English speaker-labeling test, with a 32.4% error rate. The runner-up scored a 34.2% error rate. Third, tuning the model on medical speech cut the error rate from 22.44% to 12.17% on unseen speakers.
The change cuts errors by 45.8%. The tuning took about 14 hours on one GPU, so specialist teams can adapt the model quickly. Because the model is open-weight, developers can download it and run it themselves.
The OpenMDW-1.1 license allows commercial use, so companies can build products on the model. However, its size is a trade-off, since traditional transcription systems often have under one billion parameters.
What Comes Next for Voice AI
Nvidia built the model for recorded audio, so live voice assistants still need a streaming model. Because Nvidia supplied the benchmark results, independent tests should confirm them before companies replace transcription systems.
For now, the company rates each language as Leading, Production-ready, or Supported, so users can check accuracy first. Developers can download the model from Hugging Face and ModelScope, and a NeMo Speech toolkit supports tuning. The company also plans compressed versions, which will use less memory on local devices.
