Nvidia Releases Nemotron 3.5 Transcribe Open Voice Model

Nvidia announced the release of Nemotron 3.5 Transcribe, a free, open-weight model designed to transcribe audio across up to 40 languages. Alongside the model release, the company introduced version 2 of Granary, its open-source speech dataset.

According to reported statements from Nvidia, Nemotron 3.5 Transcribe can translate 31 languages to and from English, process audio recordings up to an hour in length, and separate up to eight distinct speakers. The company stated that the model performs favorably against competitors on evaluation benchmarks, including Hugging Face’s Open ASR Leaderboard. Independent verification of these benchmark claims remains pending.

Target Use Cases and Technical Specifications

The newly released tools are aimed at developers building applications for business meetings and medical transcription. Nvidia reported that developers can combine the model with the updated Granary dataset and tools like the NeMo Speech model to customize applications for industry-specific terminology across various languages.

Key reported technical details include:

• Model Size: 30 billion parameters.
• Hardware Compatibility: Designed to run on a single chip, ranging from the Jetson AGX Thor robotics controller and DGX Spark laptop chip to Blackwell datacenter chips.
• Limitation: The model currently processes recorded audio files rather than live audio streams.

Context and Industry Environment

Nvidia’s release arrives amid a period of active development in AI voice technology. Other recent releases in the sector include Mistral’s Voxtral 3, ElevenLabs’ Realtime tool for speech in voice agents, and Salesforce’s Launa-Omni voice model.

Practical Implications and Verification

Developers and organizations interested in evaluating Nemotron 3.5 Transcribe can review the official announcements and technical documentation provided directly by Nvidia. Practical testing on local or cloud hardware is recommended to verify performance metrics, accuracy, and hardware compatibility against specific workload requirements before deployment.