🔍 Read the full analysis: 5 Ways To Use NVIDIA Nemotron 3.5 Transcribe And Granary V2 For Multilingual AI 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 model combines transcription, speaker attribution, translation and audio question answering; NVIDIA and benchmark providers report strong results, but independent performance across real-world settings remains to be established.
NVIDIA has released Nemotron 3.5 Transcribe, an open-weight model for speech tasks, on Hugging Face and ModelScope, together with Granary v2 speech data and tools for fine-tuning. The model combines transcription, speaker attribution, translation and audio question answering in one checkpoint, giving developers a single starting point for applications that turn recordings into several kinds of text output.
NVIDIA says the model transcribes 40 languages, translates 31 to or from English and attributes speakers in 13. It supports recordings up to one hour, can label as many as eight speakers and provides word-level timestamps. The company also says developers can use natural-language prompts to guide output and adapt it to domain vocabulary, accents and preferred written forms. These language counts refer to different tasks, rather than equal coverage across every capability.
The release is paired with Granary v2, which NVIDIA describes as multilingual speech data with a reusable curation pipeline, and NeMo Speech tools for fine-tuning. Developers can try prompts first, then use training tools and data to adapt the model to an application’s terminology or speech patterns. NVIDIA advises checking names, speaker changes and generated answers against the original recording.
NVIDIA reports a 5.33% average word error rate on English, 4.62% mean WER across seven languages and 10.98% mean WER on long-form audio on the Hugging Face Open ASR Leaderboard, describing each as the best result among evaluated open-weight models in the cited comparison. Artificial Analysis reports a tie for first among open-weight models at 2.8% AA-WER, with about four billion active parameters. These are benchmark results, not guarantees for every recording or deployment.
One Model, Several Speech Outputs
Combining transcription and speaker labels with translation and audio questions could reduce the need to route a recording through separate models for each task. That may be useful for meeting records, multilingual calls, classroom replays and specialized dictation, where a plain transcript can miss who spoke, leave speech untranslated or omit useful structure such as word-level timing.
The release also connects model access with an adaptation path: developers can prompt for terminology and formatting, then use Granary v2 and NeMo Speech to train for their own use case. NVIDIA reports a medical case study in which training lowered word error rate on held-out speakers by 45.8% relative to the best prompt. That result is specific to the reported study; it does not establish the same gain across specialties, languages or other datasets.
For teams evaluating speech systems, the practical question is not only whether a model performs well on a leaderboard, but whether it handles the languages, accents, noise and overlapping speech in their recordings. The release gives developers a package to test those questions, while leaving real-world performance dependent on local evaluation.
Products Worth Considering
Bose QuietComfort True Wireless Earbuds with Active Noise Cancellation, Black
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
As an affiliate, we earn on qualifying purchases.
From Benchmarks to Adaptation
NVIDIA reports separate results for transcription and speaker diarization. For speaker attribution, the company 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. According to NVIDIA, the model uses Sortformer speaker cues during generation. The rankings reflect the cited evaluations and should not be read as a universal assessment of performance.
The release’s demonstrations include a fast, overlapping-dialogue scene from the 1940 film His Girl Friday and a classroom replay built by Agora using word-level timestamps. These examples show possible workflows, but they are demonstrations, not independent accuracy tests. NVIDIA also says audio question answering is trained in English, so its stated support for transcription or translation should not be assumed to apply equally to audio questions.
“One open model for multilingual transcription, speaker attribution, translation, and audio Q&A.”
— NVIDIA, in its Hugging Face release
Performance Across Real Recordings
The supplied release material does not provide enough detail to independently assess every benchmark’s test sets, recording conditions and comparison methods. Performance may vary with accents, background noise, overlapping speech and specialized terminology. Leaderboard scores are measurements under particular evaluation setups and do not guarantee accuracy for a given organization’s recordings.
It is also unclear how performance varies across each supported language and how well the model handles noisy or mixed-language audio. The 45.8% medical result is relative to the best prompt in that specific case study, and the source does not establish that the improvement transfers to other settings. NVIDIA says audio question answering is trained in English; broader coverage for that task is not established in the supplied material.
Testing and Fine-Tuning the Model
Developers can access Nemotron 3.5 Transcribe and Granary v2 on Hugging Face and ModelScope, consult the model card for prompts and task-by-language information, and use NeMo Speech tools to fine-tune. The next useful evidence will come from testing across languages, domains and recording conditions, including comparisons against each team’s existing transcription workflow.
Until broader independent results are available, users should treat NVIDIA’s benchmark placements and medical case study as reported evidence for specific evaluations, not as a forecast for every deployment. Checking transcripts, speaker labels and generated answers against the source audio remains part of responsible use.
Key Questions
What is Nemotron 3.5 Transcribe?
It is an open-weight speech model from NVIDIA that combines transcription, speaker attribution, translation and audio question answering in one checkpoint.Which languages does the model support?
NVIDIA says it transcribes 40 languages, translates 31 to or from English and attributes speakers in 13. These counts apply to separate tasks, and support should not be assumed to be identical across them.What are Granary v2 and NeMo Speech for?
Granary v2 provides multilingual speech data and a reusable curation pipeline. NVIDIA’s NeMo Speech tools can be used to fine-tune the model for particular vocabulary, accents or application needs.Do the reported benchmark results guarantee accuracy?
No. The leaderboard results describe performance in particular evaluations. Results on actual recordings can vary with language, accents, noise, overlapping speech and domain vocabulary.Where can developers access the release?
The supplied source says Nemotron 3.5 Transcribe and Granary v2 are available on Hugging Face and ModelScope. NVIDIA also points developers to its model card for prompts and task-by-language details.Primary source: Hugging Face · via ThorstenMeyerAI.com
