Nvidia releases open transcription model

NVIDIA’s Nemotron 3.5 Transcribe is an open-weight speech recognition model that turns recorded audio into text. It supports transcription in 40 languages, and can also label speakers, translate some languages to or from English, and answer questions about audio. The model is available on Hugging Face. NVIDIA says it’s designed for recorded audio, not live streaming, and advises users to check names, speaker changes and answers against the recording. Read NVIDIA’s announcement.

For a business, the useful question isn’t just whether a transcript is accurate. It’s whether a team can turn calls, meetings or other recordings into the format it needs, such as speaker-labelled notes or a translation. The model can be guided with natural-language instructions, though language support varies by task. It also supports adapting recognition to specific vocabulary and accents, using NVIDIA’s Granary v2 data and fine-tuning tools.

The model drew 61 points and 7 comments on Hacker News, which makes it worth watching for teams evaluating speech workflows. That interest isn’t proof it’ll fit your process. Test it on your own recordings and check the output before relying on it. This is the kind of capability we build into an AI command centre, so teams can keep track of tools that may matter to their workflows.