Nvidia released Nemotron 3.5 Transcribe on Wednesday, an open-weight speech model that transcribes 40 languages, translates 31 of them to or from English and labels who is speaking in 13. The model posts a 5.33% average word error rate on Hugging Face's English Open ASR Leaderboard, which Nvidia says ranks first among open-weight models.
Nemotron 3.5 Transcribe has 30 billion total parameters but activates only about 4 billion per token, according to Nvidia's Hugging Face post. It pairs a speech encoder with a hybrid Mamba-Transformer mixture-of-experts language model, and its language component draws on Nemotron 3.5 Lightning, which Nvidia released in August. Commercial use is permitted.
The model is promptable. Developers can ask in plain language for a clean or verbatim transcript, or feed it domain vocabulary such as medical or product terms. It also produces word-level timestamps and can answer questions about audio, though that question-answering mode was trained in English only.
Key facts
- 01Nvidia released Nemotron 3.5 Transcribe on October 28 as an open-weight speech model that permits commercial use.
- 02It transcribes 40 languages, translates 31 to or from English and labels speakers in 13 languages.
- 03The model has 30 billion total parameters, with about 4 billion active per token.
- 04It posts a 5.33% average word error rate on Hugging Face's English Open ASR Leaderboard, first among open-weight models.
- 05Nvidia also released Granary v2, a speech dataset of 463,062 hours across 48 languages.
Nvidia built it for recorded audio rather than live streaming. Nemotron 3.5 Transcribe handles recordings up to one hour long and can label up to eight speakers.
The efficiency numbers are the core of the pitch. On Artificial Analysis, the model scored a 2.81% word error rate, against 2.77% for Mistral's Voxtral Small, which uses about six times as many active parameters. On one RTX PRO 6000 Blackwell Server Edition GPU running at NVFP4 precision, Nvidia says the model processes 100 hours of audio in under three minutes.
Speaker attribution held up on harder tests. On VoiceArena's overlapping-speech benchmark, Nemotron 3.5 Transcribe placed first of 14 systems and missed 8.32% of reference words, compared with a median of about 31%. On David AI's 21-language diarization benchmark, it posted an 18.26% error rate, narrowly ahead of Nvidia's own Nemotron 3 Diarization at 18.38%.
Nvidia also showed how far fine-tuning can push the model. Working with Eka Care on medical speech from Indian doctors, the company cut the word error rate on held-out speakers from 22.44% to 12.17%, a 45.8% relative reduction, using about 14 GPU-hours on a single GB300. The tuned model's error rate rose slightly on general English, and Nvidia recommends testing custom versions on at least 100 verified recordings.
Alongside the model, Nvidia released Granary v2, a free speech dataset covering 463,062 hours across 48 languages, up from 25 languages in the first version. Users download the audio from each original source under that source's license.
Nemotron 3.5 Transcribe runs on a single Nvidia GPU, from Jetson AGX Thor and DGX Spark up to Blackwell data center hardware. It is available on Hugging Face and ModelScope, through hosted endpoints from Baseten, DeepInfra and Simplismart, and can be served with vLLM or NeMo-Speech.cpp. Nvidia also added it to RTX Spark Windows laptops, which start at $2,599 from ASUS, Dell, HP, Lenovo, Microsoft and MSI.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
