🔍 Read the full analysis: NVIDIA Nemotron 3.5 Transcribe And Granary V2: Intelligent Multilingual Transcription You Can Prompt And Adapt on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
NVIDIA has made Nemotron 3.5 Transcribe available on Hugging Face and ModelScope, alongside Granary v2 speech data and tools for fine-tuning. The company reports strong leaderboard results and a 45.8% reduction in word error rate in one medical case study, but those findings do not establish performance across all languages, domains or recordings.
NVIDIA has released Nemotron 3.5 Transcribe, an open-weight speech model, alongside Granary v2 multilingual speech data and tools for fine-tuning. Available on Hugging Face and ModelScope, the model combines transcription, speaker attribution, translation and audio question answering in one checkpoint, giving developers a single system to test for several audio-to-text tasks.
NVIDIA says the model can transcribe 40 languages, translate 31 languages to or from English, and identify speakers in 13 languages. It supports recordings up to one hour, can label as many as eight speakers and produces word-level timestamps. The company also describes natural-language prompting for tasks such as requesting a speaker-labeled transcript, an English translation or answers to questions about a recording.
The release includes Granary v2, a multilingual speech-data resource with a reusable curation pipeline, and NeMo Speech tools for fine-tuning. NVIDIA presents these as ways for developers to adapt the model to domain vocabulary, accents and preferred written forms. Its guidance says audio question answering is trained in English, and language coverage differs by task.
NVIDIA reports a 5.33% average English word error rate on the Hugging Face Open ASR Leaderboard, a 4.62% mean across seven languages and a 10.98% mean on long-form audio. It describes these as the best results among evaluated open-weight models in the cited comparisons. Artificial Analysis separately reports a tie for first among open-weight models at 2.8% AA-WER, with about four billion active parameters. These are benchmark findings, not assurances of accuracy on an individual recording.
One Model, Several Audio Tasks
Combining transcription, speaker labels and translation could let developers produce multiple outputs from one recording without routing it through separate models for each task. That may be useful for meeting records, classroom replays, multilingual calls and specialized dictation, where a plain transcript can miss who spoke or retain terminology in an unsuitable language or format.
The release also connects model access with data curation and fine-tuning. Developers can start with prompts and then adapt the system using relevant speech data and vocabulary. NVIDIA reports that training in one medical case study reduced word error rate on held-out speakers by 45.8% compared with the best prompt. That result points to a possible benefit in that setting, but does not show the same improvement will occur in other medical specialties or datasets.
For organizations considering the model, published scores are a starting point rather than a substitute for testing. Results can vary with accent, noise, overlapping speech and specialist terms, as well as the language and evaluation method. The practical value will depend on whether it performs reliably on the recordings and tasks a developer actually needs.
Products Worth Considering
Logitech MX Master 4 for Mac, Ergonomic Bluetooth Mouse with Haptic Feedback, Ultra-Fast Scrolling, USB-C Charging, for MacOS, iPadOS, Windo
Sonos Era 100 SL Compact Microphone-Free Speaker, Black
Jackery HomePower 1000 Plus v2 Portable Power Station
As an affiliate, we earn on qualifying purchases.
Benchmarks and Adaptation Tools
The release reports results for transcription and diarization separately. For speaker attribution, NVIDIA cites first place among evaluated open-weight systems on David AI’s combined transcription-and-diarization track, as well as top results in comparisons from David AI and VoiceArena. The company says the model uses Sortformer speaker cues during generation. The supplied material does not include enough detail to independently assess every test set or comparison.
NVIDIA’s examples include a rapid, overlapping-dialogue scene from the 1940 film His Girl Friday and a classroom replay built by Agora using word-level timestamps. These demonstrations show possible applications, but are not independent evaluations of accuracy. NVIDIA advises users to check names, speaker changes and generated answers against the original audio.
The company also recommends supplying domain terms with context and instructing the model to use them only when supported by the recording. Developers can access the model and Granary v2 on Hugging Face and ModelScope, and use NeMo Speech tools for adaptation. The model card is described as containing presets, task-by-language support and prompt guidance.
“One open model for multilingual transcription, speaker attribution, translation, and audio Q&A.”
— NVIDIA, in its Hugging Face release
Limits of the Published Results
The supplied information does not establish how results vary across every supported language, accent, recording condition or domain. Language counts differ by task: transcription, translation and speaker attribution do not have identical coverage, and NVIDIA says audio question answering is trained in English. Performance on mixed-language or particularly noisy recordings is not specified.
The medical result is a comparison with the best prompt in one case study, not evidence of a general 45.8% reduction across medical use. The source material also does not provide enough test-set and methodology detail to independently verify all leaderboard comparisons. Real-world accuracy remains dependent on the recording and evaluation conditions, and benchmark placement does not guarantee results for a particular user.
Testing Across Real Recordings
Developers can try Nemotron 3.5 Transcribe and Granary v2 through Hugging Face or ModelScope, consult the model card and test the recommended prompts. Those working with specialized vocabulary can explore adaptation through NeMo Speech and relevant speech data.
The next useful evidence will come from testing across languages, domains and recording conditions, including independent evaluations that describe their datasets and comparison methods. Until then, organizations should check generated transcripts, speaker labels and audio-question answers against source recordings before relying on them in consequential settings.
Key Questions
What is Nemotron 3.5 Transcribe?
It is an open-weight speech model released by NVIDIA for transcription, speaker attribution, translation and audio question answering.
Which languages does it support?
NVIDIA says it transcribes 40 languages, translates 31 languages to or from English, and attributes speakers in 13 languages. Coverage differs by task; the company says audio question answering is trained in English.
Where can developers get the model and data?
NVIDIA says Nemotron 3.5 Transcribe and Granary v2 are available on Hugging Face and ModelScope. NeMo Speech tools are provided for fine-tuning.
Do the benchmark scores guarantee transcription accuracy?
No. The scores are measurements from specified leaderboard comparisons. Accuracy on a particular recording can differ with language, accent, noise, overlapping speech, vocabulary and evaluation conditions.
What does the medical case study show?
NVIDIA reports that training reduced word error rate on held-out speakers by 45.8% relative to the best prompt in that case study. The supplied material does not show that the same improvement applies to other specialties or datasets.
Primary source: Hugging Face · via ThorstenMeyerAI.com
