Nvidia has launched Nemotron 3.5 Transcribe, an open-weight speech-to-text model that supports transcription in up to 40 languages. This new AI tool excels in handling recorded audio and is capable of identifying up to eight speakers in conversations lasting as long as an hour. According to Nvidia, Nemotron 3.5 outperforms competitors on key benchmarks such as the Hugging Face Open ASR Leaderboard, positioning it as a leading solution for transcription needs, particularly in business meetings and medical settings.

Alongside Nemotron 3.5, Nvidia released version 2 of Granary, its open-source speech dataset, enabling developers to enhance transcription accuracy by tailoring models to industry-specific jargon across multiple languages. Although Nemotron 3.5 excels in transcription of recorded files, it does not support live audio streams, which remain the domain of alternatives like ElevenLabs' real-time voice tools.

Nvidia emphasizes that its 30 billion-parameter model can operate efficiently on a variety of hardware platforms, ranging from the Jetson AGX Thor robotics controller and Dell’s DGX Spark laptop chips to the Blackwell datacenter processors. This efficiency opens opportunities for deployment across many devices while maintaining high performance. Nemotron joins a suite of Nvidia’s open-weight voice AI models, including the smaller yet powerful Parakeet model with 1.1 billion parameters.

The release comes amid a surge in voice AI developments from major players, with recent launches including Mistral’s Voxtral 3 and Salesforce’s Launa-Omni voice model. Nvidia's announcement highlights the growing importance and competitiveness of voice AI technologies in the enterprise sector, as companies seek more accurate and scalable transcription solutions for professional use cases.