Nvidia Launches Nemotron 3.5 Transcribe, a 30‑Billion‑Parameter Open‑Source Voice Model

Open‑Source Speech Tools Target Enterprise Workflows

On Wednesday, Nvidia introduced Nemotron 3.5 Transcribe, a free, open‑weight model capable of converting spoken audio into text across up to 40 languages. The company also rolled out version 2 of Granary, its open‑source speech dataset, expanding the resources available to developers.

Nemotron 3.5 can translate 31 languages to and from English, process recordings as long as an hour, and differentiate up to eight distinct speakers. Nvidia claims the model outperforms competitors on benchmark tests, including Hugging Face’s Open ASR Leaderboard.

The tools are designed for developers building applications in two of AI’s most entrenched workplace domains: business meetings and medical transcription. By combining Nemotron 3.5 with the updated Granary dataset and the NeMo Speech model, developers can tailor solutions to recognize industry‑specific jargon in each language. A limitation remains: Transcribe works only with pre‑recorded audio and cannot handle live streams.

Despite its size, the 30‑billion‑parameter model is optimized to run on a single chip, supporting hardware ranging from the Jetson AGX Thor robotics controller and the DGX Spark laptop chip to Blackwell datacenter processors. It joins Nvidia’s portfolio of open‑weight voice models, which includes the compact yet powerful Parakeet model of 1.1 billion parameters.

The release arrives amid a wave of AI voice announcements. Earlier this month, Mistral shipped Voxtral 3; ElevenLabs launched Realtime for instant speech in voice agents; and Salesforce released Launa‑Omni, its own voice model. Together, these moves illustrate a rapidly evolving landscape where open‑source tools are becoming central to enterprise AI adoption.