Nvidia’s new AI model can transcribe, translate, and summarize long recordings

Nvidia has introduced Nemotron 3.5 Transcribe, an open weight AI model that combines multilingual transcription, translation, speaker labeling, and audio question answering in one system. Kyt Dotson reports for SiliconANGLE that the company is also releasing Granary v2, an open source speech dataset and data pipeline intended to help developers adapt the model.

The model can process recordings in 40 languages, distinguish up to eight speakers, and answer questions about audio files lasting as long as 100 hours, according to Nvidia. Rather than using separate tools for transcription, speaker identification, and analysis, users can request these tasks through prompts in a single model pass.

For example, a user can ask the system to create a transcript with speaker names, translate a recording into English, or identify decisions made in a meeting. Users can also supply industry terminology or sample dialogue before transcription to improve results in specialized fields.

Nemotron 3.5 Transcribe uses a mixture of experts design with 30 billion total parameters, though only 4 billion are active for a given task. Nvidia says the model can run on one graphics processor, from edge devices such as the Jetson AGX Thor to data center hardware. On a Blackwell Server Edition GPU, the company says it can process 100 hours of audio in less than three minutes with compression and lower precision settings.

Nvidia reports a 30.8 percent word error rate without additional guidance, falling to 20.8 percent with prompting. The company says a version fine tuned for Eka Care’s Indian medical speech benchmark reached an 11.8 percent error rate.

Granary v2 includes a public manifest covering 463,062 hours of speech across 48 languages and 16 source collections. It provides three transcript styles for each recording, including verbatim text, normalized numbers, and a version without filler words. The model checkpoint and a Hugging Face demo are available for developers.