Nvidia has released Nemotron 3.5 Transcribe, a free, open-weight model that transcribes audio in up to 40 languages and translates 31 of them to and from English.
The semiconductor and AI computing company paired the launch with version 2 of Granary, its open source speech dataset, giving developers both the model and the tools to adapt it to specialist vocabulary.
The model can process recordings of up to an hour, distinguish as many as eight speakers, and handle transcription, translation, speaker diarization, timestamps and audio question-answering within a single system, according to AI Insider's report.
It runs at 30 billion parameters but uses a mixture-of-experts architecture with roughly 4 billion parameters active per inference, a design meant to deliver large-model performance at lower running cost.
Granary v2's public manifest spans 463,062 hours across 48 languages and 16 source corpora, up from Granary v1's 25 languages.
Nvidia said it outperforms rivals on benchmarks including Hugging Face's Open ASR Leaderboard, a platform that Nvidia has reported hosting more than 18 million users and 3 million models.
The model targets business meetings and medical transcription, two established enterprise AI use cases, and can run on hardware ranging from Jetson AGX Thor robotics controllers to Blackwell data centre processors.
It handles only recorded audio, not live streams, and joins Nvidia's existing open-weight voice lineup including the 1.1-billion-parameter Parakeet.
The release lands amid a crowded stretch for voice AI, following Mistral's Voxtral 3, ElevenLabs' Realtime and Salesforce's Launa-Omni voice model.
