Nvidia bets on open transcription with Nemotron 3.5

Nvidia releases Nemotron 3.5 Transcribe, an open-access model to transcribe 40 languages and translate 31 into English. The company also releases Granary v2, a corpus of 463,062 hours of speech.

Nvidia launched Nemotron 3.5 Transcribe on Wednesday, an open-weight and free AI model capable of transcribing speech in 40 languages, then translating 31 of them to and from English. The system is designed for pre-recorded audio files, up to one hour, with possible separation of eight speakers. This release targets professional uses, as it aims at meeting summaries, documentation, and medical or technical transcription.

A model tailored for pre-recorded, not live

Nemotron 3.5 Transcribe is not intended for live streams. Nvidia designed it to process already captured recordings, which brings it closer to documentary production chains than real-time conversations. The model accepts natural language requests: a developer can ask for a verbatim transcription, a cleaned-up version, or increased sensitivity to industry jargon.

The architecture features about 30 to 34 billion parameters in total, but activates only about 4 billion at each computation step. This type of design, close to a partial activation model, helps reduce machine load without sacrificing broad capacity. Result: the system can run on a single chip, from Jetson cards to laptop processors, up to enterprise Blackwell GPUs. Not very shy for a model meant to save computation.

Granary v2 expands the training base

At the same time, Nvidia made public Granary v2, a training corpus of 463,062 hours of speech in 48 languages. The previous version covered 25 languages; the linguistic coverage has thus nearly doubled. Developers can access the source audio data under their respective licenses, offering an alternative to paid commercial offerings.

The company also claims that Nemotron 3.5 Transcribe ranks among the top on Hugging Face’s open speech-recognition benchmark, with an average word error rate of 5.33% in English. Nvidia says it adapted the model to speeches by Indian doctors, with a 45.8% error reduction on test sets after about 14 hours of training on a single processor. These figures remain to be read within the framework of the evaluation announced by the company, but they illustrate a key point: the model fine-tunes quickly.

A strategic move in voice and open tools

This launch is part of a broader sequence. Nvidia has already released its audio model Parakeet, with 1.1 billion parameters, and relies on the Nemotron family, with Nano Omni in April and Lightning in August. By opening these building blocks while linking them to its hardware, the group aims to make its computing infrastructure the base layer for voice, assistance, and transcription applications.

The competition is already well established. Mistral AI launched Voxtral 3, ElevenLabs deployed real-time conversational agents, and Salesforce presented Launa-Omni. In this context, Nvidia bets on a simple combination to explain to companies: open models, authorized commercial deployment, and direct compatibility with its chip range. It’s less spectacular than an assistant chatting at full speed, but often more useful when hours of audio need to be turned into usable text.

Key points

  • Nemotron 3.5 Transcribe covers 40 languages and translates 31 languages.
  • Granary v2 totals 463,062 hours of speech in 48 languages.
  • The model identifies up to eight speakers on the same file.
  • Nvidia announces a 5.33% average error rate in English on Hugging Face.
  • Granary v2 increases from 25 to 48 languages, according to Nvidia.
  • The model can run on a single chip, from Jetson to Blackwell.

By the numbers

  • 40 languages — transcription supported by Nemotron 3.5 Transcribe.
  • 31 languages — translation to and from English.
  • 463,062 hours — volume of the Granary v2 corpus.
  • 5.33% — average error rate in English cited by Nvidia.
  • 45.8% — error reduction announced after fine-tuning on Indian doctors.

To read