Nvidia has released Nemotron 3.5 Transcribe, an open-weight model for multilingual audio transcription and translation, together with an open data pipeline for fine-tuning it. Both sit on Hugging Face and ModelScope.
Four steps in a single model pass
Conventional transcription systems work in stages: detect speech, transcribe it, label the speakers, then answer questions about the result. Transcribe does all of that in one pass and digests up to 100 hours of audio. It transcribes 40 languages and translates 31 of them.
Under the hood it is a mixture-of-experts model with 30 billion parameters, 4 billion of them active. That fits on modest hardware. A single Nvidia GPU is enough, from a Jetson AGX Thor through a DGX Spark up to a Blackwell card in the data center.
Control happens through prompts. The same checkpoint turns one meeting recording into different outputs depending on the instruction: “Transcribe the recording and label each speaker”, “Translate this audio into English”, “What decisions were made in this meeting?” You can hand the model specialist terms up front.
Where it lands on the leaderboards
On Hugging Face’s Open ASR Leaderboard, Transcribe is first among open-weight models for English. For speaker attribution it posts the lowest error rate on David AI’s 21-language diarization leaderboard and ranks first in VoiceArena’s 14-system comparison.
Nvidia puts the word error rate at 30.8% out of the box and 20.8% with careful prompting. Fine-tuned for a domain it falls sharply: on Eka Care’s Indian medical speech benchmark it reaches 11.8%.
The dataset is the real lever
To build the model, Nvidia extended its open speech dataset Granary to version 2 using NeMo Curator. The public manifest covers 463,062 hours across 48 languages and 16 source corpora, up from 25 languages.
That is the part worth watching, rather than the error rates. An open model with good numbers is a product, and six months from now it is behind. An open dataset of this size with a reusable curation pipeline is a foundation other labs build on, and 48 languages is an offer aimed squarely at the regions speech models usually skip.