Nvidia releases Nemotron 3.5 ASR, a free open-weight model for 40-language audio transcription

Nvidia has released Nemotron 3.5 Transcribe, a free, open-weight speech model that can transcribe audio in up to 40 languages.

Nvidia paired the launch with version 2 of Granary, its open-source speech dataset. One release gives developers a model. The other gives them the material to build their own.

What Nemotron 3.5 Transcribe actually does

The headline feature is breadth. The model transcribes speech across 40 language locales, according to research findings on the release.

It also works as a translator. Nemotron 3.5 Transcribe can translate 31 languages to and from English.

Length is less of a problem than it used to be. The model handles recordings up to an hour long.

Then there is speaker separation, known in the field as diarization. Nemotron 3.5 Transcribe can distinguish between up to eight speakers.

Under the hood, the research findings describe a model with 600 million parameters.

Built for real-time use

The model, also referenced in research materials as Nemotron 3.5 ASR (automatic speech recognition), supports real-time streaming. That means it can transcribe speech as it happens rather than waiting for a finished file.

Nvidia built that streaming capability on what the research findings call a cache-aware FastConformer-RNNT architecture. The model remembers what it has already processed so it does not have to reread the whole conversation every time a new word arrives.

Latency reaches as low as 0.07 seconds to the final transcript, according to the research findings. Developers can also configure chunk sizes, trading a bit of speed for accuracy or the reverse depending on the use case.

Punctuation and capitalization come built in.

The model is available on Hugging Face under the OpenMDW-1.1 license, per the research findings. Those findings place the release in early June 2026, around June 4 to 6.

The Granary factor

Nvidia first launched Granary in August 2025. That initial version included roughly one million hours of audio data covering 25 European languages, according to the research findings.

Version 2 now supports the expanded multilingual model. Releasing the dataset openly means outside researchers can study, reuse or extend the training material.

The previous Nemotron 3 model was English-only, per the research findings, so moving to 40 language locales is less an update and more a change of category.