🔍 Read the full analysis: Can NVIDIA Nemotron 3.5 Transcribe And Granary V2 Adapt To Your Needs? 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 adapting the model. The company and benchmark providers report strong leaderboard results, but those measurements and a medical fine-tuning case study do not establish performance across all languages, recording conditions or applications.
NVIDIA has released Nemotron 3.5 Transcribe, an open-weight speech model that combines transcription, speaker attribution, translation and audio question answering, as described in the original analysis, and made it available on Hugging Face and ModelScope. The release also includes Granary v2 speech data and tools for fine-tuning, giving developers options to test the model as provided or adapt it to their own audio and terminology.
NVIDIA says the model transcribes 40 languages, translates 31 languages to or from English, and attributes speakers in 13 languages. It accepts recordings up to one hour, can label as many as eight speakers and provides word-level timestamps. The language counts refer to different tasks, rather than indicating that every feature supports the same set of languages. NVIDIA also says developers can use natural-language prompts to guide formatting and terminology.
The company reports a 5.33% average English word error rate, a 4.62% mean across seven languages and a 10.98% mean on long-form audio on the Hugging Face Open ASR Leaderboard. NVIDIA describes those as the best results among open-weight models in the cited comparison. Separately, Artificial Analysis reports a tie for first among open-weight models at 2.8% AA-WER, with about four billion active parameters. These are measurements on particular evaluations, not a promise of accuracy on other recordings.
For speaker attribution, NVIDIA cites first place among evaluated open-weight systems on David AI’s combined transcription and diarization track, along with top results in comparisons from David AI and VoiceArena. The release says the model uses Sortformer speaker cues during generation. NVIDIA also reports that, in a medical case study, fine-tuning reduced word error rate on held-out speakers by 45.8% relative to the best prompt. That result is specific to the case study and does not establish an equivalent gain in other settings.
One Checkpoint, Multiple Speech Tasks
Combining transcription, speaker labels, translation and audio question answering in one model may simplify workflows that otherwise require several speech systems. For meeting records, classroom audio or multilingual calls, a transcript with speaker attribution and timestamps can be more useful than plain text alone. Translation and questions about a recording may provide additional ways to work with the same audio.
The release also links those capabilities to a route for customization. Developers can start with prompts, then use Granary v2 and NeMo Speech tools to prepare data and fine-tune the model. This could help teams test whether domain vocabulary, accents or preferred written forms are handled better in their own use case. The reported medical result offers an example of adaptation, but its scope is limited to the evaluation NVIDIA describes.
For organizations choosing a speech system, the announcement is a reason to evaluate a new open-weight option—not evidence that it will outperform existing tools on every task. Practical results may depend on audio quality, language, overlap and specialist vocabulary, as well as how each user defines and measures accuracy.
Products Worth Considering
Bose S1 Pro+ Portable PA Speaker
Logitech MX Master 4 for Mac, Ergonomic Bluetooth Mouse with Haptic Feedback, Ultra-Fast Scrolling, USB-C Charging, for MacOS, iPadOS, Windo
Jackery HomePower 1000 Plus v2 Portable Power Station
As an affiliate, we earn on qualifying purchases.
Release Includes Data and Tools
NVIDIA’s announcement is broader than a model release: it pairs Nemotron 3.5 Transcribe with Granary v2 multilingual speech data and reusable data-curation resources. The company says developers can use NeMo Speech tools to fine-tune the model. The materials are presented as a way to move from testing general prompts toward training on speech and vocabulary closer to an application’s needs.
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 how the model may be used; they are not independent accuracy evaluations. NVIDIA’s guidance also recommends checking names, speaker changes and generated answers against the original audio.
Task coverage is not uniform. NVIDIA says audio question answering is trained in English, while transcription, translation and speaker attribution have separate language support. That distinction matters when comparing the model’s advertised language counts with a particular application’s requirements.
“One open model for multilingual transcription, speaker attribution, translation, and audio Q&A.”
— NVIDIA, in its Hugging Face release
Performance Outside Published Tests
The supplied release information does not provide enough detail to independently assess every benchmark’s test sets, recording conditions or comparison methods. Leaderboard scores are tied to specific evaluations; accuracy may differ with accents, background noise, overlapping speakers, mixed-language audio and specialized vocabulary. The reported figures should not be read as results guaranteed for a particular deployment.
The medical case study reports a 45.8% reduction in word error rate on held-out speakers relative to the best prompt, but the material does not establish that the result applies across medical specialties, languages or other datasets. The degree of performance variation across the model’s supported languages and recording conditions is also not specified. Developers will need to assess those questions on audio relevant to their own use.
Testing and Fine-Tuning Options
Developers can access Nemotron 3.5 Transcribe and Granary v2 through Hugging Face and ModelScope, then use the model card’s presets and prompt guidance to test supported tasks. Those who want to adapt it can explore NVIDIA’s NeMo Speech fine-tuning tools and the release’s data-curation resources.
The next useful evidence will come from testing across more languages, domains and recording conditions, including independent evaluations that explain their datasets and methods. For now, users should compare outputs with source audio and measure performance on their own recordings before relying on the model in a routine workflow.
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. NVIDIA says it is available on Hugging Face and ModelScope.
Which languages does the model support?
NVIDIA says it transcribes 40 languages, translates 31 languages to or from English, and attributes speakers in 13 languages. Support varies by task, and the company says audio question answering is trained in English.
Can developers adapt the model to specialist vocabulary?
NVIDIA says developers can guide output with prompts and fine-tune the model using Granary v2 data and NeMo Speech tools. The reported medical improvement is from one case study and does not show that the same result will occur in other fields.
Do the leaderboard results guarantee accuracy on my recordings?
No. The reported scores apply to specific benchmarks and comparison groups. Results on a particular recording may vary with noise, accents, overlapping speech, language and domain terms, so users should test the model on representative audio.
Primary source: Hugging Face · via ThorstenMeyerAI.com
