Nvidia released Nemotron 3.5 Transcribe on Wednesday, October 28, as an open-weight speech model covering 40 languages. One checkpoint handles transcription, speaker labels, translation and spoken-question answering on recordings up to an hour long.
At a Glance
- Nemotron 3.5 Transcribe translates 31 languages to or from English and attributes speakers in 13, telling up to eight apart.
- The model averages a 5.33% word error rate on the English track of Hugging Face’s Open ASR Leaderboard.
- Transcribe posted the lowest speaker-attribution error rate among 14 systems on David AI’s 21-language diarization leaderboard.
- Granary v2, Nvidia’s updated open speech dataset, lists 463,062 hours across 48 languages, up from 25 in version 1.
Nvidia folds speaker labels into the transcript
Nvidia is pitching the model at developers building meeting tools and medical dictation, one of the older footholds for AI in healthcare. One prompt can turn a meeting recording into a speaker-labeled transcript, and another can pull out the decisions made. Nvidia trained audio question answering only in English, and the model card warns that language support differs by task.
Transcribe pairs a speech encoder with a hybrid Mamba-Transformer mixture-of-experts language model. Only about 4 billion of its 30 billion parameters are active per token. That keeps each token far cheaper than the full size suggests. Speaker cues from Nvidia’s Sortformer reach the language model alongside the acoustic features, so it tags each turn while it writes.
Meet NVIDIA Nemotron 3.5 Transcribe 🎙️ An open-weight, promptable 30B MoE model with ~4B active parameters that brings transcription in 40 languages, speaker attribution, translation, contextual prompting and audio understanding into one checkpoint. Watch it keep track of who said what when everyone’s talking at once. 1/ 🧵
— NVIDIA AI (@NVIDIAAI) October 28, 2026
That English score on the Open ASR Leaderboard ranks first among open-weight models. Nvidia’s chart for that ranking leaves API models out, a detail worth reading twice. Artificial Analysis, which ranks speech-to-text models independently, puts Transcribe at 2.8% on its AA-WER index. That ties for first among open-weight models, with about 4 billion active parameters against 24 billion for the model it ties.
Speaker attribution is where the release looks strongest. Transcribe scored an 18.26% diarization error rate on David AI’s multilingual track. Nvidia’s dedicated Nemotron 3 Diarization model placed second at 18.38%, which made Nvidia its own closest rival. VoiceArena’s English benchmark also ranked Transcribe first, at 32.4% cpWER, a word error score counted after matching speaker labels.
The speed numbers depend on an unreleased build
Nvidia says Transcribe can run on a single GPU, whether a Jetson AGX Thor, a DGX Spark or a Blackwell data-center card. That claim depends on quantized versions the company lists as coming soon. The same caveat covers its showiest figure, 100 hours of audio processed in under three minutes. Nvidia measured that on one RTX PRO 6000 Blackwell Server Edition GPU, using NVFP4 precision and 30-second clips at high concurrency.
Developers can serve the current checkpoint with vLLM today. Local deployment through NeMo-Speech.cpp waits on the same quantized release, and the model handles recorded audio only, with no live streaming. Teams testing it should spot-check names, speaker changes and answers against the original recording, which is Nvidia’s own guidance.
Granary v2 and NeMo Speech fine-tuning recipes ship alongside the model for teams that need domain jargon. Nvidia’s medical case study shows what that buys. A prompt naming English and the Latin alphabet cut full-benchmark WER from 30.8% to 23.8% before any training. Fine-tuning then cut WER on held-out speakers from 22.44% to 12.17%, a 45.8% relative drop.
The test used Eka Care’s open benchmark of recordings from Indian doctors and medical students. The full run took about 14 GPU-hours on one GB300. The adapted model also edged an open medical-specialist model, which scored 13.79% on the same speakers. It still gave ground elsewhere, starting with general English, where LibriSpeech test-clean WER rose from 1.54% to 1.69%.
The specialist also kept its lead on exact matches for brand-name drugs and diagnostics. Nvidia’s study names real recordings of clinicians saying brand names as the next step.
Why It Matters?
Transcribe joins Nvidia’s smaller open voice models, including the 1.1-billion-parameter Parakeet. It also lands in a crowded month. Mistral announced Voxtral 3 on X earlier in October, and Salesforce released Launa-Omni, its voice model. ElevenLabs, meanwhile, launched Realtime for instant speech inside voice agents.
That last release points at the gap. Rivals keep pushing toward real-time voice models, while Transcribe waits for whole files and works on the meeting archive and the dictation backlog. Its next markers are the quantized builds and the NeMo-Speech.cpp release for local use. Nvidia’s announcement gives a date for neither.
