NVIDIA released the Granary v2 dataset with 844 thousand hours of audio for Nemotron 3.5 Transcribe
Nemotron 3.5 Transcribe has 30 billion parameters, approximately 4 billion of which are active, and runs on a single GPU according to NVIDIA. The release is accompanied by the Granary v2 dataset with 844 thousand hours of audio and tools for domain adaptation.
New details about Nemotron 3.5 Transcribe describe 30 billion parameters, approximately 4 billion of which are active during processing. According to NVIDIA, the model can run on a single GPU. The release also includes the Granary v2 dataset with approximately 844 thousand hours of speech in 48 languages, along with tools for data preparation and domain adaptation of the model.
The model combines transcription, attribution of speech to individual speakers, translation and question answering about recordings. According to NVIDIA, it transcribes 40 languages, translates between English and 31 languages, and distinguishes speakers in 13 languages. It handles recordings up to an hour long with no more than eight speakers. It works with recorded audio, not live streams; question answering about audio was trained in English.
The model is free and has open weights available on Hugging Face and ModelScope. NVIDIA offers BF16, FP8 and NVFP4 formats and lists support for systems including DGX Spark and GeForce RTX 5090 GPUs. NVFP4 variants can be deployed through NeMo-Speech.cpp or vLLM.
NVIDIA demonstrated domain adaptation using medical recordings: according to its evaluation, the transcription error rate for speakers not included in training fell from 22.44 % to 12.17 %. However, the adaptation slightly worsened results on general English. The company recommends setting aside at least 100 real, verified recordings to validate your own adaptations.
Why it matters
Developers can combine transcription, speaker labeling and translation in one model and choose whether recordings are processed locally or in the cloud. Domain adaptation offers a route to more accurate specialized transcription, but the documented example also shows possible deterioration outside the target domain.
Release card
Nemotron 3.5 Transcribe
NVIDIA
- Specifications
- The model has 30 billion parameters, approximately 4 billion of which are active when generating a single token.
- Inputs
- Inputs are audio recordings and text instructions; the output is text. The model transcribes 40 languages, translates between English and 31 languages, and attributes speech to speakers in 13 languages. It can process recordings up to an hour long with no more than eight speakers.
- Price
- free
- Availability
- The weights are available for download on Hugging Face and ModelScope. NVIDIA offers BF16, FP8 and NVFP4 formats; NVFP4 variants can be deployed through NeMo-Speech.cpp or vLLM.
- Hugging Face Open ASR Leaderboard – English 5,33 % WER The value represents the average error rate for English speech transcription on public test sets.
- Open ASR – Multilingual 4,62 % WER The value represents the average transcription error rate across seven languages.
- Open ASR – Long-form 10,98 % WER The value represents the average transcription error rate for long recordings in this evaluation.
- Suitable for transcribing recorded meetings with speaker labels.
- Suitable for translating recordings between English and 31 supported languages.
- Allows speech recognition to be adapted to specialized terminology through instructions and further training.
- Does not support processing live audio streams.
- Question answering about recordings was trained in English.
- Language support varies by task.
According to the source, the model combines transcription and speaker attribution, which required Nemotron 3 Diarization to work with a separate transcription model, and adds translation and audio comprehension.
The card summarizes information from the article and any dated corrections, with a link to the original source. It is not our assessment of the model. It does not yet have a dedicated editorial profile. Model selection and other announcements →
What was added since the original report
Verified updates
-
The model has 30 billion parameters, with approximately 4 billion active in a single step.; The model can run on a single GPU.; The release includes the Granary v2 dataset with 844 000 hours of audio.; The release includes tools for adapting the model to specialized domains.
- The model has 30 billion parameters, with approximately 4 billion active in a single step.
- The model can run on a single GPU.
- The release includes the Granary v2 dataset with 844 000 hours of audio.
- The release includes tools for adapting the model to specialized domains.
-
The weights are available in BF16, FP8 and NVFP4 formats.; The model supports question answering about recordings.; The NVFP4 version can be deployed through NeMo-Speech.cpp or vLLM.; The model is available on RTX Spark.
- The weights are available in BF16, FP8 and NVFP4 formats.
- The model supports question answering about recordings.
- The NVFP4 version can be deployed through NeMo-Speech.cpp or vLLM.
- The model is available on RTX Spark.
-
The model processes recorded audio, not live streams.; The model has 30 billion parameters.; According to the company, the model can run on a single chip.; The second version of the open speech dataset Granary was released at the same time.
- The model processes recorded audio, not live streams.
- The model has 30 billion parameters.
- According to the company, the model can run on a single chip.
- The second version of the open speech dataset Granary was released at the same time.
Two audiences, two different impacts
What this means
For individuals
For a recorded meeting, text instructions let you choose between a transcript with speaker labels, a translation and an overview of decisions made. Available features vary by language.
For a business
Teams developing applications for specialized transcription gain data and tools for adapting to domain terminology. When evaluating quality, they must monitor both performance in the target domain and any deterioration in general transcription.
DevelopmentCheck the original
Event sources
independently confirmed · 3 publishers, 1 independent. We count feeds from the same owner only once.