Nvidia released Nemotron 3.5 Transcribe on Wednesday, a free, open-weight model capable of transcribing audio in up to 40 languages. Alongside it, the company published version 2 of Granary, its open-source speech dataset. According to the source, the target audience is developers building applications for corporate meetings and for transcription in medical settings.
Among the capabilities the company announced, the model translates 31 languages to and from English, processes recordings of up to one hour, and tells apart as many as eight different speakers in the same recording. Nvidia says it outperforms competitors on tests, including Hugging Face's Open ASR Leaderboard. These figures come from the company itself and have not yet been independently verified.
The model has 30 billion parameters and, according to Nvidia, runs on a single chip. The supported hardware ranges from the Jetson AGX Thor robotics controller and the DGX Spark, which powers new notebooks, to Blackwell datacenter chips. It joins the family of open voice models the company has already been publishing.
For teams running production systems, the change is about cost and architecture: an open model that runs locally allows transcription without sending audio to third parties, which matters in sectors bound by confidentiality, such as healthcare and legal. It is worth testing the model on your own recordings before any migration, since benchmarks published by the vendor do not replace measurements with your own content and accents.