Nvidia has introduced an AI model for transcribing and translating audio.

Nvidia has released Nemotron 3.5 Transcribe, an open-source AI model that transcribes audio recordings, translates speech, and identifies speakers. It can also generate brief meeting summaries and answer questions based on the content of the recordings.

The model supports 40 languages and allows multiple tasks to be performed with a single instruction. For example, it can prepare a meeting transcript with speaker names, translate a conversation into English, or extract key decisions.

Nemotron 3.5 Transcribe runs on a single Nvidia GPU. The model is suitable for processing recordings up to an hour long, involving multiple participants, including those who interrupt each other.

Alongside the model, the company introduced Granary v2, an open dataset of audio recordings in 48 languages. Developers can use it to train their own models and adapt speech recognition for specific domains, such as medicine.

Nemotron 3.5 Transcribe and Granary v2 are available on Hugging Face. The model works with pre-recorded audio files but does not yet support real-time speech recognition.

Source: huggingface.co