Nvidia Releases Nemotron 3.5 Transcribe Open-Weight Speech Model

What Is New

On Wednesday, Nvidia released Nemotron 3.5 Transcribe, a free, open-weight speech-to-text model [1]. The model can transcribe audio in up to 40 languages and includes the ability to translate 31 languages to and from English [1]. Nvidia also released version 2 of Granary, which is an open-source speech dataset [1].

Key technical details include:

• The model has 30 billion parameters [1].
• It can process recordings up to one hour in length [1].
• It can distinguish between up to eight different speakers [1].
• It is designed to run on a single chip, including Jetson AGX Thor robotics controllers, DGX Spark laptop chips, and Blackwell datacenter chips [1].

Nvidia claims the model performs well on benchmarks such as Hugging Face’s Open ASR Leaderboard [1].

How to Use It

Developers can use the Nemotron 3.5 Transcribe model, the updated Granary dataset, and NeMo Speech tools to build and customize applications [1]. These tools allow for the recognition of specific industry jargon within different languages [1]. The model is intended for developers building software for business meetings and medical transcription [1].

Limits, Privacy and Academic Integrity

A primary limitation of Nemotron 3.5 Transcribe is that it only works with recorded audio [1]. It does not support live audio or livestreams [1].