Nvidia launches open-source voice AI model for multilingual transcription

Nvidia has launched Nemotron 3.5 Transcribe, an open-source model capable of transcribing audio in up to 40 languages. The model also supports translation between 31 languages and English, and can distinguish up to eight speakers. It outperforms competitors on benchmarks like Hugging Face’s Open ASR Leaderboard. Nvidia also released version 2 of its Granary speech dataset, which developers can use to customize applications for industry-specific jargon. Transcribe, with 30 billion parameters, runs on a single chip, including the Jetson AGX Thor and DGX Spark. It joins Nvidia’s open-weight voice models, such as Parakeet, which has 1.1 billion parameters. The release comes amid a surge in AI voice models, with Mistral launching Voxtral 3 earlier this month, ElevenLabs introducing Realtime for voice agents, and Salesforce unveiling Launa-Omni. Despite its capabilities, Transcribe currently works with recorded audio, not live streams, which is a limitation for real-time applications.