Nvidia on October 28 released Nemotron 3.5 Transcribe, an open-weight model that transcribes recorded audio in up to 40 languages. The company says the model can translate 31 languages to and from English, process recordings up to an hour long, and distinguish up to eight speakers.
The 30-billion-parameter model is designed to run on a single chip, according to Nvidia, from its Jetson AGX Thor robotics controller and DGX Spark laptop chip to Blackwell datacenter chips. The company claims Nemotron 3.5 outperforms competitors on benchmarks including Hugging Face’s Open ASR Leaderboard.
Nvidia also released version 2 of Granary, its open-source speech dataset. The company says developers can use the model, dataset and tools including its NeMo Speech framework to customize applications for industry-specific terminology in each language.
The company positions the tools for business meeting and medical transcription applications. One limitation: Nemotron 3.5 Transcribe works only with recorded audio, not livestreams. The model joins Nvidia’s existing open-weight voice offerings, including Parakeet, a 1.1-billion-parameter transcription model.