Nvidia Releases Free 40-Language Speech-to-Text Model for Recorded Audio

Nvidia released Nemotron 3.5 Transcribe on Wednesday, October 28, 2026 — a free, open-weight speech-to-text model capable of transcribing audio in up to 40 languages. The company also released version 2 of Granary, its open source speech dataset, alongside the model.

Nemotron 3.5 Transcribe can translate 31 languages to and from English, handle recordings up to an hour long, and distinguish between up to eight speakers. Nvidia says the model outperforms rivals on benchmarks, including Hugging Face’s Open ASR Leaderboard. At 30 billion parameters, the model is designed to run on a single chip, spanning Nvidia’s Jetson AGX Thor robotics controller and DGX Spark laptop chips through to Blackwell datacenter chips.

The release targets developers building applications for business meeting transcription and medical transcription — two areas where AI tools have already gained significant workplace adoption. Nvidia says developers can combine the model with the updated Granary dataset and its NeMo Speech tools to customize apps that recognize industry-specific terminology across languages. The model joins Nvidia’s existing lineup of open-weight voice tools, including Parakeet, a smaller model at 1.1 billion parameters.

One notable limitation: Nemotron 3.5 Transcribe works only with pre-recorded audio and does not support live audio streams, which could limit its usefulness for real-time applications.

The launch arrives during a period of active voice AI releases across the industry. Mistral shipped Voxtral 3 earlier in October 2026, ElevenLabs launched Realtime for instant speech in voice agents, and Salesforce released its own voice model, Launa-Omni. Nvidia’s entry into this space with a free, open-weight offering may give developers a cost-accessible alternative for building transcription-focused applications.