Nvidia Launches Nemotron 3.5 Transcribe as Open Source, Supporting 40 Languages

Nvidia claims that Nemotron 3.5 Transcribe can operate on a single chip, from the Jetson AGX Thor and DGX Spark to Blackwell for data centers. However, the model only processes recorded audio, with no support for live input. This announcement comes as Mistral, ElevenLabs, and Salesforce have recently launched their own voice models.

Works Offline and Announced on a Single Chip

Nemotron 3.5 Transcribe does not support live streams and is used solely with audio recordings. Nvidia states that the model, equipped with 30 billion parameters, can run on a single chip. The mentioned chips range from the Jetson AGX Thor controller and the DGX Spark chip for laptops to Blackwell chips intended for data centers. Nemotron 3.5 adds to Nvidia's lineup of open-source voice models, which also includes Parakeet, a model with 1.1 billion parameters.

Claimed Capabilities and Target Use Cases

Nemotron 3.5 Transcribe can transcribe audio in up to 40 languages. Nvidia specifies that the model allows for translation between 31 languages and English, manages recordings of up to one hour, and distinguishes between up to eight speakers. The company claims that the model outperforms its competitors on benchmarks, including Hugging Face’s Open ASR Leaderboard. These tools are designed for applications such as business meetings and medical transcription. Nvidia indicates that developers can pair Nemotron 3.5 with Granary v2 and tools like NeMo Speech to tailor jargon recognition specific to each sector and language. The model is offered for free as open source.

Granary v2 Available and a Very Active Voice Market

Nvidia accompanies this launch with version 2 of Granary, its open-source voice dataset. This announcement comes amid a surge of releases in the voice domain: Mistral recently delivered Voxtral 3, ElevenLabs launched Realtime for instant speech generation in voice agents, and Salesforce released its voice model Launa-Omni.