Voice AI in vogue: Nvidia releases new open transcription model
AI-assisted digest, checked against the original source before publishing. How we work
Nvidia ships an open-weight, 40-language transcription model for meeting and medical tools — but only for recordings, not live audio.
Nvidia released Nemotron 3.5 Transcribe on Wednesday, a free, open-weight speech-to-text model covering up to 40 languages, alongside version 2 of Granary, its open-source speech dataset. The release targets developers building transcription apps for business meetings and medical documentation — two of AI's most established workplace use cases.
The model translates between English and 31 languages, processes recordings up to an hour long, and distinguishes among as many as eight speakers. Nvidia claims Nemotron 3.5 Transcribe outperforms rivals on benchmarks including Hugging Face's Open ASR Leaderboard. The company also positions the model alongside its NeMo Speech tooling so developers can teach apps an industry's jargon in each language.
At 30 billion parameters, Transcribe is designed to run on a single chip, ranging from the Jetson AGX Thor robotics controller and the DGX Spark chip found in new laptops to Blackwell datacenter GPUs. It joins Nvidia's existing open-weight voice roster, which includes Parakeet, a 1.1-billion-parameter model.
The clearest limitation: Transcribe handles recorded audio only, not live streams. That leaves real-time applications such as voice agents open to rivals — ElevenLabs launched Realtime earlier for instant speech in voice agents. The broader voice AI field is crowded right now: Mistral shipped Voxtral 3 earlier this month, and Salesforce released its Launa-Omni voice model.
What to watch: whether Nvidia closes the live-streaming gap, and how much the open-weight, single-chip economics shift adoption for meeting and medical transcription builders weighing self-hosted models against hosted APIs.
Nvidia's claims of benchmark superiority will face scrutiny as developers put the 40-language coverage and eight-speaker diarization to work on real-world recordings, where domain jargon in medicine and enterprise meetings tends to be the hard part. The Granary dataset update could matter as much as the model itself if it materially improves fine-tuning for specialized vocabularies. The bigger open question is whether single-chip inference at 30 billion parameters makes on-premise transcription economical for regulated industries like healthcare, where audio often can't leave the premises.
Key facts
- • Nvidia released Nemotron 3.5 Transcribe, a free open-weight model, Wednesday.
- • Transcribe supports up to 40 languages and 31 languages to and from English.
- • The 30-billion-parameter model handles hour-long recordings and up to eight speakers.
- • Nvidia also released version 2 of its open-source Granary speech dataset.
- • Transcribe runs on a single chip, from Jetson AGX Thor to Blackwell GPUs.
Why it matters
Builders shipping meeting and medical transcription tools gain a free, open-weight option that runs on one chip, lowering hardware and licensing barriers. Because it lacks live-stream support, teams building voice agents will still depend on realtime-focused vendors like ElevenLabs — until Nvidia ships streaming.
Spotted an error? Send a correction.
Get the day's AI news in one email
The Daily Epoch · 60-second read · unsubscribe anytime
