Nvidia’s Nemotron 3.5 Transcribe Is a Free Speech Model That Knows Who Said What
Nvidia has released Nemotron 3.5 Transcribe, an open-weight speech model you can download and run yourself.
What’s confirmed. The model card lists support for 40 languages, recordings up to one hour, and optional word timestamps. Nvidia says one model handles transcription, speaker labels, translation and questions about the audio. You pick the output style with a plain-English prompt, like “label each speaker.” It can label up to eight speakers, and it is built for recorded audio, not live streams.
What Nvidia claims. Per Nvidia’s launch post, no other open-weight model beats it on the English track of Hugging Face’s Open ASR Leaderboard, where it averages a 5.33% word error rate. It says the model translates 31 languages to or from English and labels speakers in 13. Its research page says it is the first model of its type to beat specialized multi-step systems at tagging overlapping speakers.
Tuning it yourself. Nvidia also released Granary v2, its open training data and cleanup pipeline, plus fine-tuning recipes. In a medical case study, fine-tuning cut word errors on unseen speakers by 45.8% compared with the best prompt alone.
Our take. For podcasters, researchers and anyone stuck with meeting recordings, one open model that knows who spoke is a real upgrade over gluing tools together. But “open” does not mean “runs on your laptop.” The model card points to an 80 GB H100 GPU for the full version, though smaller quantized builds exist. Nvidia itself advises spot-checking names and speaker changes against the recording.
What’s real here: a top open transcription model, with heavy hardware strings attached.
Sources
- Hugging Face blog: NVIDIA Nemotron 3.5 Transcribe and Granary v2
- Hugging Face model card: NVIDIA-Nemotron-3.5-Transcribe-30B-A4B
- NVIDIA Research: Nemotron 3.5 Transcribe
By Willa Ravenwire · AI Bandwagon