Nvidia has launched Nemotron 3.5 Transcribe, a new open-weight model capable of transcribing audio across up to 40 languages. Alongside this, the company rolled out version 2 of Granary, its open-source speech dataset. The tools are aimed at developers building enterprise applications in fields like business meetings and medical transcription.
Nemotron 3.5 supports translation between English and 31 other languages, recognizes up to eight distinct speakers in a single recording, and can process audio clips up to an hour long. According to benchmark comparisons, it outperforms rivals on the Open ASR leaderboard. Although its capabilities are impressive, the model handles only recorded audio—it does not support live audio streams.
Performance & Applications
Running at 30 billion parameters, Nemotron 3.5 is designed for versatile deployment. It’s optimized to function on everything from Nvidia’s Jetson AGX Thor robotics controller to the DGX Spark and Blackwell datacenter chips. It joins the company’s lineup of open-weight voice models, such as Parakeet, a much smaller model that nevertheless delivers strong performance with just 1.1 billion parameters.
Nvidia also emphasized that developers can fine-tune these tools using Granary and NeMo Speech to adapt them to industry-specific vocabularies across different languages. This customization is especially relevant for sectors requiring specialized terminology, like healthcare or legal fields.
Where It Fits in the Voice AI Landscape
Nemotron 3.5 arrives amid a surge of new voice AI tools. Earlier in October, Mistral released Voxtral 3, while other companies such as ElevenLabs and Salesforce have introduced their own speech and voice models—highlighting growing competition and rapid innovation in speech recognition and transcription.
Despite its strengths, Nemotron 3.5’s inability to process live audio is a limitation in comparison to some real-time speech tools entering the market. But for use cases like post-meeting summaries, medical records, and any situations where accuracy and multilingual transcription matter, it presents a substantial step forward.
As voice AI becomes more central to enterprise workflows, Nemotron 3.5 reflects both the promise and the trade-offs typical in this wave of innovation. What to watch next: whether Nvidia adds live audio support, how its accuracy holds up in noisy, real-world conditions, and how it stacks up against rivals’ offerings in latency, cost, and customization.
