Voice AI in vogue: Nvidia releases new open transcription model

THE NEWS

What happened

Nvidia has unveiled Nemotron 3.5 Transcribe, a free open-weight AI model capable of transcribing audio in up to 40 languages and distinguishing up to eight speakers simultaneously. The release also includes Granary 2.0, an updated open-source speech dataset. Designed for enterprise applications like business meetings and medical transcription, the model supports customization for industry-specific jargon. Notably, Nemotron 3.5 boasts 30 billion parameters yet can run on a single chip, ranging from robotics controllers to new laptop GPUs and datacenter Blackwell chips.

CONTEXT

Why it matters

Nvidia has officially released Nemotron 3.5 Transcribe, a powerful open-weight model designed for enterprise voice AI. Capable of transcribing up to 40 languages and distinguishing between eight speakers, the model targets critical sectors like business meetings and medical transcription. Despite its massive 30 billion parameters, it is optimized to run on a single chip, compatible with everything from Jetson robotics controllers to new Blackwell datacenter GPUs. The release also includes Granary 2.0, an updated speech dataset. While it beats rivals on current benchmarks, developers should note the current limitation: the model processes recorded audio only, not livestreams.

AT A GLANCE

Key facts

• Nvidia released Nemotron 3.5 Transcribe, a free open-weight model supporting up to 40 languages.
• The model can distinguish between up to eight different speakers in a single recording.
• It handles audio recordings up to one hour in length and translates 31 languages to/from English.
• Nvidia also released Granary 2.0, an updated open-source speech dataset for developers.
• The model targets business meetings and medical transcription sectors with customizable industry jargon support.
• Despite having 30 billion parameters, the model can run on a single chip across various Nvidia hardware.
• Supported hardware includes Jetson AGX Thor, DGX Spark chips in new laptops, and Blackwell datacenter GPUs.
• Current limitation: The model works with recorded audio only, not livestreams.
• This launch occurs amidst competition from Mistral's Voxtral 3, ElevenLabs' Realtime, and Salesforce's Launa-Omni.
• Nemotron 3.5 claims to outperform rivals on benchmarks like Hugging Face's Open ASR Leaderboard.