Open speech model labels who said what in hour-long recordings, NVIDIA says

NVIDIA has released Nemotron 3.5 Transcribe, a free, downloadable model that turns recordings into text from written instructions, plus a 463,062-hour speech dataset. The accuracy figures are NVIDIA's own.

Share
Open speech model labels who said what in hour-long recordings, NVIDIA says
Illustration: AI-generated

One meeting recording can become a transcript with speaker labels, an English translation, or an answer to what decisions were made, depending on the instruction you give the model, according to NVIDIA's post. The model is Nemotron 3.5 Transcribe, published on 28 October 2026.

It is an open-weight model, meaning its trained numbers can be downloaded. It turns recorded speech into text and follows written instructions, called prompts. NVIDIA says it transcribes 40 languages, translates 31 of them to or from English, and can tell speakers apart in 13. Users can also list medical terms like HbA1c and metformin and tell it to use them only when heard.

NVIDIA says it handles recordings up to an hour long and labels up to eight speakers. In one example, it captions a scene from the 1940 film His Girl Friday, known for fast, overlapping dialogue, and labels who is speaking. It works on recorded audio only, not live streams, NVIDIA and TechCrunch both say.

A speech encoder, the part that listens, feeds a language model built from many specialist sub-networks. Only about 4 billion of its 30 billion parameters, the adjustable numbers inside a model, are used for each piece of text it writes.

NVIDIA's accuracy figures come from its own tests and from public leaderboards. On Hugging Face's Open ASR Leaderboard, which measures the share of words transcribed wrongly on public test sets, NVIDIA says the model ranks first among open-weight models for English, at 5.33% average word error rate.

NVIDIA also released Granary v2, a speech dataset of 463,062 hours across 48 languages. NVIDIA says the 463,062 hours describe the dataset release, not the full data used to train the model. The audio is downloaded separately under each original collection's licence.

In a medical case study, NVIDIA tested on Eka Care's open benchmark of 3,619 clips from Indian doctors and medical students. Further training on medical speech cut the error rate from 22.44% to 12.17% on speakers held back from training. That is NVIDIA's own study.

The study also names limits. The error rate on general English rose slightly, from 1.54% to 1.69%, and an existing medical-specialist model still did better on exact brand-name drugs.

NVIDIA also says language support differs by task and audio question answering is trained in English only. It advises checking names, speaker changes and answers against the recording.

People can try the model in a Hugging Face demo or on ModelScope Studios, and download it from Hugging Face or ModelScope.

Sources