Nvidia Releases 30 Billion-Parameter Nemotron 3.5 for 40-Language Transcription
Updated
Updated · TechCrunch · Oct 28
Nvidia Releases 30 Billion-Parameter Nemotron 3.5 for 40-Language Transcription
3 articles · Updated · TechCrunch · Oct 28
Summary
Nemotron 3.5 Transcribe is a free, open-weight audio model that supports transcription in up to 40 languages, translation for 31 languages to and from English, and speaker separation for as many as eight voices.
Nvidia paired the launch with Granary v2, an updated open-source speech dataset, and said developers can combine both with NeMo Speech tools to tune apps for industry-specific jargon in meetings and medical transcription.
A key limitation is that Transcribe handles recorded audio rather than live streams, though Nvidia says the 30 billion-parameter model can process recordings up to an hour long.
Nvidia also says the model can run on a single chip—from Jetson AGX Thor and DGX Spark to Blackwell datacenter chips—as AI voice competition intensifies with recent releases from Mistral, ElevenLabs and Salesforce.
If Nemotron 3.5 runs on one chip from edge to cloud, could customizable on-device transcription finally challenge cloud-only speech APIs?
Why is Nvidia giving developers a 30B-parameter transcription model and dataset now—and is the real prize enterprise AI software or chip demand?
Can Nvidia’s free open-weight speech model really beat closed rivals in noisy, speaker-overlapping meetings, or do leaderboard wins hide real-world weaknesses?