Nvidia has introduced the open-weight Nemotron 3.5 Transcribe model with 30 billion parameters to process audio recordings of up to one hour and distinguish 8 speakers across 40 languages, accompanied by version 2 of the Granary dataset.
Nemotron 3.5 Transcribe specifications and multilingual capabilities
Technical differences between transcription, speaker diarization, and translation
Open-weight model with 30 billion parameters and Granary 2 dataset
Focus on recorded audio and independent evaluation requirements
Audio accuracy challenges in real-world and professional environments
Hardware requirements and open-weight model licensing terms
Frequently asked questions
Nemotron 3.5 Transcribe specifications and multilingual capabilities
On October 28, 2026, Nvidia launched “Nemotron 3.5 Transcribe,” a model specialized in automatic speech recognition, alongside a new release of its “Granary” audio dataset via Hugging Face. According to TechCrunch, the new system supports transcription for audio recordings in up to 40 languages, as well as bidirectional translation between English and 31 other languages.
Nvidia states that the model can process recorded audio clips lasting up to a full hour at once, while identifying and separating the voices of up to 8 different speakers. These specifications remain subject to the vendor’s announcements, as practical quality varies depending on accents, background noise, overlapping voices, and specialized terminology.
Technical differences between transcription, speaker diarization, and translation
Speech-to-text systems rely on translating the audio signal into written language, while speaker diarization technology identifies time segments likely belonging to different individuals within a recording, without verifying their personal identities. Real-time translation adds another processing layer that may introduce changes to the original meaning.
These subtle distinctions are highly important for journalists, doctors, and enterprises seeking to automate the transcription of lectures and meetings; a transcript that reads smoothly may attribute an important statement to the wrong speaker or miss a precise medical term.
Open-weight model with 30 billion parameters and Granary 2 dataset
Nvidia classifies this system as an open-weight model featuring 30 billion parameters, with support for running across a spectrum of devices ranging from developer workstations to Blackwell-class data center processors. The availability of open weights provides access to model parameters under specific terms, but does not necessarily provide all original training data or exempt large-scale deployment from computational costs.
The “Granary 2” dataset represents a separate resource supporting model adaptation and evaluation, with developers required to comply with its accompanying licensing terms during use.
Focus on recorded audio and independent evaluation requirements
The TechCrunch report clarified that the product is designed to process pre-recorded audio files rather than live, real-time audio streams—a crucial detail for applications requiring an instant voice assistant with low latency and continuous processing.
Offering open-weight alternatives gives organizations greater control over internal deployments, particularly when handling sensitive audio recordings, while the need remains for independent testing to measure word error rates and speaker attribution accuracy under real-world recording conditions.
Audio accuracy challenges in real-world and professional environments
The release of the 30-billion-parameter open-weight Nemotron 3.5 Transcribe model alongside the second version of Granary data provides a powerful tool for organizations looking to process long audio recordings locally, especially given its stated capacity to handle audio files spanning a full hour and distinguish up to 8 speakers across 40 languages.
However, professional reliance on audio transcription and translation outputs in sensitive fields such as medicine, journalism, and law necessitates caution regarding subtle errors; speaker diarization relies on acoustic characteristics and does not establish a person’s legal identity, while overlapping voices and local accents can lead to transcription errors that require human review.
Hardware requirements and open-weight model licensing terms
Releasing the model in an open-weight format allows developers and organizations to run it on their own infrastructure, ranging from advanced workstations to Blackwell servers, enhancing the privacy of sensitive audio files compared to sending them to external cloud APIs.
Nevertheless, running a 30-billion-parameter model to process long audio recordings carries significant computational costs, and engineering teams must review the licensing terms for both the model and the Granary 2 dataset before integrating them into commercial products.
Frequently asked questions
Question: How many languages and speakers does the Nemotron 3.5 Transcribe model support?
Answer: The model supports audio transcription in 40 languages, translation between English and 31 languages, and the identification of up to 8 speakers in a recording of up to one hour.
Question: What is the size of Nvidia’s new Nemotron 3.5 Transcribe model?
Answer: The model size is 30 billion parameters, available as an open-weight model that runs on developer hardware and Blackwell processors.
Question: Is the new Nemotron model designed for live, real-time voice conversations?
Answer: No, the model is designed to process pre-recorded audio clips of up to one hour, rather than low-latency live audio streaming applications.
Question: What is the Granary 2 dataset released by Nvidia alongside the model?
Answer: It is an updated version of Nvidia’s audio database, used to train, adapt, and evaluate speech recognition and translation models.
