NVIDIA boosts multilingual transcription with a promptable model
Nemotron 3.5 Transcribe targets transcription, speaker attribution, translation, and audio question answering in a single open model. Granary v2 provides the data pipeline and fine-tuning to adapt it for business uses.
NVIDIA has introduced Nemotron 3.5 Transcribe, an open weight speech transcription model that also handles speaker attribution, translation, and audio question answering. It comes with Granary v2, a data curation pipeline designed to adapt the model to accents, industry vocabularies, and precise output formats. This is relevant for use cases in meetings, healthcare, or customer relations, where a good verbatim transcript is not always enough.
A model that does more than just "untangle" audio
Nemotron 3.5 Transcribe is a promptable model, meaning it can be given instructions in natural language to guide the output. NVIDIA states it transcribes 40 languages, translates 31 to or from English, attributes speakers in 13 languages, and handles recordings up to one hour with up to eight people. The idea is simple: the same recording can produce, depending on the instruction, a raw transcription, a translated version, a speaker-segmented transcript, or an answer to a question about the content.
The model is available on Hugging Face and ModelScope, and can run on a single NVIDIA GPU, from DGX Spark to Blackwell systems in datacenters. For developers, NVIDIA also highlights deployment guides with vLLM and NeMo Speech. No crystal ball needed to understand the benefit: the model aims to reduce patchwork between transcription, diarization, and post-processing.
Solid results on benchmarks, but [to be verified] depending on use cases
On paper, Transcribe shows good results on several public leaderboards. NVIDIA claims it is first among open weight models on the Hugging Face Open ASR Leaderboard in English, with an average WER of 5.33%, and places it at 4.62% average WER multilingual on seven languages in its open weight comparison. On the Artificial Analysis platform, it reaches 2.8% AA-WER, at about 4 billion active parameters, compared to 24 billion for the model it is compared to in that race.
The speaker attribution aspect is equally emphasized. On the David AI DAI-ASR-I18N benchmark, NVIDIA reports 18.26% DER on 21 languages, ahead of Nemotron 3 Diarization at 18.38%. On VoiceArena, the model shows 32.4% cpWER, placing it at the top of the comparison mentioned by NVIDIA. These figures remain benchmark measures, useful for comparing models, but not an automatic promise of perfect performance in a noisy open space on a Monday morning.
Granary v2, the unseen plumbing that matters a lot
Granary v2 is the data component associated with Nemotron 3.5 Transcribe. NVIDIA describes it as a reusable curation pipeline built with NeMo Curator, capable of producing multiple transcription variants, instruction pairs, provenance metadata, and quality indicators. The public manifest covers 463,062 hours of audio, 48 languages, and 16 source corpora, compared to 25 languages in Granary v1.
This pipeline is used to create consistent output formats, notably the verbatim and normalized variants that the model can reproduce on demand. It also produces datasets for contextual recognition, restoration of borrowed words, short audio summaries, translation, and audio-anchored question answering. In practice, this allows aligning the dataset with the expected output, avoiding asking a model to guess punctuation as if reading tea leaves.
Fine-tuning changes the game in specialized fields
NVIDIA highlights a medical use case to show the benefit of adaptation. On the Eka Care benchmark, the best instruction alone reduces the WER from 30.8% to 23.8%, then fine-tuning brings it down to 11.8% on the full set. For speakers held out of training, the WER drops from 22.44% to 12.17%, a relative decrease of 45.8% according to the published document.
The result illustrates a key point: “generic” performance is not always enough in professions where terms matter, such as healthcare or technical sectors. NVIDIA also specifies that developers can adjust recognition to domain-specific words, transcription conventions, or particular accents via its NeMo Speech tools. For companies, the challenge is not just to transcribe, but to transcribe accurately and usefully.
What this means for product and data teams
With Nemotron 3.5 Transcribe and Granary v2, NVIDIA pushes a clear logic: unify transcription, diarization, and business adaptation uses in a single block. Teams building voice assistants, reporting tools, or audio search functions find a more modular model than a simple dictation engine. It remains to be seen, depending on real cases and deployment costs, whether this flexibility simplifies teams’ work or adds another layer of tuning.
The published resources also cover local deployment, hosted endpoints, and fine-tuning recipes. In other words, NVIDIA is not just selling a model, but the whole workshop around it. And in this field, the toolbox often matters as much as the hammer.
Key points
- Nemotron 3.5 Transcribe handles 40 languages and 8 speakers.
- Granary v2 covers 463,062 hours and 48 languages.
- Medical WER: 12.17% on held-out speakers.
- 18.26% DER on 21 languages, according to David AI.
- 5.33% English WER on Hugging Face Open ASR.
- NVIDIA publishes the model on Hugging Face and ModelScope.
By the numbers
- 40 languages — transcription supported by Nemotron 3.5 Transcribe.
- 31 languages — translation to or from English, according to NVIDIA.
- 463,062 hours — public volume of the Granary v2 manifest.
- 18.26% — DER reported on 21 languages by NVIDIA via David AI.
- 12.17% — WER on the medical test set, according to NVIDIA.