On Wednesday, Nvidia introduced Nemotron 3.5 Transcribe, which is a free and open-weight model capable of handling audio transcription across as many as 40 languages. Alongside this release, the firm also published version 2 of Granary, which serves as its open-source speech dataset.
According to Nvidia, Nemotron 3.5 is capable of translating 31 languages to and from English, processing recordings lasting up to one hour, and distinguishing between a maximum of eight separate voices. The corporation states that the system outperforms competitors on evaluation metrics such as the Open ASR Leaderboard from Hugging Face.
Such offerings are directed at software creators developing applications for corporate meetings and medical dictation, representing two prominent workplace areas for artificial intelligence. Nvidia notes that creators can combine the model, the revised Granary dataset, and utilities like the NeMo Speech model to tailor programs for sector-specific terminology in various languages. However, a limitation is that Transcribe only processes recorded sound rather than live audio feeds.
The organization guarantees that Transcribe operates on a single processor despite containing 30 billion parameters, fitting hardware ranging from the Jetson AGX Thor robotics controller and DGX Spark chip found in modern laptops to Blackwell datacenter hardware. This addition joins existing open-weight voice technologies from Nvidia, such as the compact 1.1 billion parameter Parakeet model.
This introduction arrives during a crowded timeframe for voice artificial intelligence announcements, following Mistral releasing Voxtral 3 earlier in the month, ElevenLabs introducing Realtime for real-time speech agents, and Salesforce publishing its Launa-Omni voice system.
