Nvidia Opens 40-Language Transcribe Model
Nvidia on Oct. 28 released Nemotron 3.5 Transcribe, an open-weight speech model that transcribes 40 languages, translates 31 to or from English and labels speakers. It leads open models on Hugging Face's English ASR leaderboard at 5.33% WER but does not handle live audio.
Nvidia on Oct. 28 released Nemotron 3.5 Transcribe, an open-weight speech model that transcribes 40 languages, translates 31 to or from English and labels who said what, all from one checkpoint that takes plain-language prompts. The company also published Granary v2, a 48-language speech dataset, according to a Hugging Face blog post by Nvidia's speech team.
Key takeaways
- The model card lists 34 billion total parameters, with about 4.3 billion active per token, under the OpenMDW-1.1 license, which allows commercial use.
- Nvidia says it ranks first among open-weight models on Hugging Face's Open ASR English leaderboard, with a 5.33% average word error rate.
- It handles recorded audio up to one hour long but does not support live streaming.
What Nemotron 3.5 Transcribe does
Speech pipelines usually chain separate models for transcription, speaker separation and translation. Nvidia folds those jobs into one model: a speech encoder feeding a hybrid Mamba-Transformer mixture-of-experts language model. Users steer the output with instructions, such as naming the expected language or supplying technical terms that should appear only if they are actually spoken.
Per Nvidia's blog, the model attributes speakers in 13 languages and labels up to eight speakers. It also produces word-level timestamps and answers questions about a recording, though audio question answering was trained in English only.
Nemotron 3.5 Transcribe benchmark results
| Benchmark | Metric | Result |
|---|---|---|
| Open ASR leaderboard, English | Average WER | 5.33% |
| Open ASR, multilingual (7 languages) | Mean WER | 4.62% |
| Open ASR, long-form | Mean WER | 10.98% |
| Artificial Analysis | AA-WER | 2.8% |
| VoiceArena Diarization-Bench (English) | cpWER | 32.4% |
On Artificial Analysis, Nvidia says the model tied for first among open-weight systems while using about 4 billion active parameters, against 24 billion for the model it tied. The Decoder identified that rival as Mistral's Voxtral Small. On the English leaderboard, The Decoder reported Qwen3-ASR-1.7B at 5.4% and IBM Granite Speech at 5.68%. These figures come from Nvidia's post and third-party leaderboards, not from independent testing by AI Newsway.
On speed, Nvidia says one RTX PRO 6000 Blackwell Server Edition GPU running NVFP4 processed 100 hours of audio in under three minutes using 30-second clips at high concurrency.
Context: an open data push alongside the model
The Granary v2 dataset card lists about 463,062 hours of speech across 48 languages and 16 source corpora, up from 25 European languages in the first release. Labels are under OpenMDW-1.1, but source audio must be downloaded separately under each corpus's terms. Nvidia cautions that the dataset does not describe the model's full training mix.
Nvidia also showed how far fine-tuning can move results. In a medical case study using an Eka Care benchmark, word error rate on held-out speakers fell from 22.44% to 12.17% after about 14 GPU-hours on one GB300. General English accuracy slipped slightly, with LibriSpeech test-clean WER rising from 1.54% to 1.69%.
The release lands in a busy stretch for voice AI. Mistral launched its Voxtral 3 speech models on Oct. 20, and OpenAI opened a full-duplex voice model to developers in September (OpenAI opens its full-duplex voice model to developers).
What's next
The weights are on Hugging Face with a demo, plus hosted endpoints from Baseten and DeepInfra. Nvidia says the model runs on a single GPU, from DGX Spark to Blackwell data-center parts. The model card lists a single H100 80GB as the reference deployment and notes that the open vLLM integration is expected to run about twice as slow as Nvidia's optimized NIM serving stack. Nvidia advises users to "spot-check names, speaker changes, and answers" against the recording.
FAQ
Can Nemotron 3.5 Transcribe caption live audio?
No. Nvidia says the model is built for recorded audio, up to one hour per file, and does not support real-time streaming.
Is it free for commercial use?
The model card lists the OpenMDW-1.1 license and says the model is ready for commercial or non-commercial use.
References
- NVIDIA Nemotron 3.5 Transcribe and Granary v2: Intelligent Multilingual Transcription You Can Prompt and Adapt — Nvidia, Hugging Face blog (Oct. 28, 2026)
- nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B — Hugging Face model card
- nvidia/Granary-v2 — Hugging Face dataset card
- Nvidia's open Nemotron 3.5 Transcribe combines transcription, translation, and speaker recognition — The Decoder (Oct. 28, 2026)
Editor responsible for AI Newsway. Writes about AI, developer tools, cloud and productivity technology.
Comments 0
Comments are currently disabled.