Nvidia has presented Nemotron 3.5 Transcribe, an audio transcription model that can be steered with natural-language instructions. It does more than transcribe: speaker attribution, translation, contextual prompts and audio understanding live in the same checkpoint, so the user can ask the model to do several things with the same file.
The model is a Mixture-of-Experts architecture with 30 billion total parameters, of which about 4 billion are active for each request. Even though it is large on paper, the actual computation per token stays contained. The weights are open and available on Hugging Face, together with the blog post in which Nvidia describes the project.
Image: NVIDIA Developer — https://developer.nvidia.com
In practice, the user uploads a recording and states the request: add speaker labels to a transcript, translate it into Spanish and then Japanese, or find and redact personal data in the text. The model handles recordings up to an hour long and distinguishes up to eight speakers, even when they talk at the same time. Color-coded captions and speaker labels help follow who said what.
On numbers, Nvidia says Nemotron 3.5 Transcribe ranks first among open-weight models on the Hugging Face Open ASR English leaderboard, with a 5.33% average word error rate. That is the company's figure, measured on a public benchmark. Language support varies by task: transcription covers 40 languages, while translation and other functions have a narrower scope.
Image: nvidia — https://huggingface.co/nvidia/nemotron-speech-streaming-en-0.6b
The release follows a series of recent Nvidia moves in audio and voice. The model is the speech-side counterpart of the Nemotron 3.5 line already presented for text and images, which includes the large Super VL and the smaller Bolt E4B. Nvidia has chosen to publish weights, recipes and datasets, a strategy that ties it to the Hugging Face platform, recently acquired for 12.9 billion dollars.
The practical question for developers is how well the model holds up beyond benchmarks. A system that transcribes, identifies who is speaking and rewrites the text in another language in a single pass reduces the pieces to assemble, but it still needs testing on your own kind of audio, where noise, accents and overlapping speech change the result.
No comments yet. Be the first!