🔍 Read the full analysis: Explore Promptable, Adaptable Transcription With NVIDIA Nemotron 3.5 And Granary V2 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 released Nemotron 3.5 Transcribe, an open-weight model for transcription, speaker attribution, translation and audio question answering, alongside Granary v2 speech data and fine-tuning tools. The company reports strong leaderboard results and a medical-domain improvement after fine-tuning, but those results do not guarantee performance across languages, domains or recording conditions.
NVIDIA has released Nemotron 3.5 Transcribe, an open-weight speech model that combines transcription, speaker attribution, translation and audio question answering, alongside Granary v2 speech data and tools for adapting the model, as detailed in the original analysis. The release gives developers a single model checkpoint to test across several audio tasks, though NVIDIA’s reported benchmark results and case study are not a guarantee of accuracy in other settings.
NVIDIA says the model can transcribe 40 languages, translate 31 languages to or from English, and attribute speakers in 13 languages. It supports recordings up to one hour, can label as many as eight speakers and provides word-level timestamps. NVIDIA describes the system as promptable: developers can use natural-language instructions to request different outputs or guide formatting and terminology.
The release includes separate results for transcription and speaker attribution. NVIDIA reports a 5.33% average word error rate in English, a 4.62% mean across seven languages and a 10.98% mean on long-form audio on the Hugging Face Open ASR Leaderboard. It describes 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. For speaker attribution, NVIDIA cites top placements on comparisons from David AI and VoiceArena.
Granary v2 provides multilingual speech data and a reusable data-curation pipeline. NVIDIA also points developers to its NeMo Speech tools for fine-tuning. In a medical case study, the company says training reduced word error rate on held-out speakers by 45.8% relative to the best prompt. That result describes one case study, not an established improvement across medical specialties or other datasets.
One Model, Several Audio Tasks
Combining transcription, speaker labels, translation and audio questions in one checkpoint could simplify workflows that otherwise rely on separate tools for each output. Meeting notes, classroom recordings, multilingual calls and specialized dictation may need more than a plain transcript: users may also want to know who spoke, translate the discussion or retrieve information from the audio.
The release pairs that consolidated model with ways to adapt it. Developers can first try prompts for terminology and output format, then use speech data and fine-tuning tools to tailor behavior to a domain. That matters for vocabulary-heavy applications, but adaptation does not remove the need to check transcripts and generated answers against recordings. Results will depend on audio quality, speech patterns and the evaluation method.
Products Worth Considering
Bose QuietComfort Wireless Bluetooth Headphones with Active Noise Cancellation, Mic, USB‑C Charging, Deep Bass, 24‑Hour Battery – White Smok
Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation
Bose QuietComfort True Wireless Earbuds with Active Noise Cancellation, Black
As an affiliate, we earn on qualifying purchases.
Prompting and Adaptation Tools
NVIDIA presents prompting as a way to specify a task or guide how the model writes its output. Its recommendations include supplying domain terms with context and telling the model to use them only when supported by the recording. The source says audio question answering is trained in English, and that language coverage differs by task; the model’s language counts should not be read as identical coverage for every feature.
The release also includes demonstrations, such as a rapid, 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 rather than independent accuracy evaluations. NVIDIA advises users to verify names, speaker changes and answers against the original audio.
“One open model for multilingual transcription, speaker attribution, translation, and audio Q&A.”
— NVIDIA, in its Hugging Face release
Performance Outside Test Sets
The supplied material does not provide enough detail to independently assess every benchmark’s test sets, recording conditions or comparison methods. Leaderboard scores are measurements on particular evaluations, not a promise of equivalent accuracy for every user. Performance may vary with accents, background noise, overlapping speech and specialized vocabulary.
The medical case study reports an improvement relative to the best prompt in that specific setting; it does not establish that other specialties, languages or datasets will see the same result. The source also does not spell out performance for every supported language or how the model handles all forms of noisy, mixed-language audio. Audio question answering’s English training is a further limitation for users considering other languages.
Testing and Fine-Tuning
Developers can access Nemotron 3.5 Transcribe and Granary v2 on Hugging Face and ModelScope, then consult NVIDIA’s model card for presets, prompt guidance and task-by-language support. Those resources offer a starting point for testing the model on application-specific recordings and, where appropriate, adapting it with NeMo Speech.
The next useful evidence will come from independent evaluations across languages, domains and real recording conditions. Until those comparisons are available, teams should treat NVIDIA’s leaderboard and medical results as reported findings, test their own audio and review generated outputs before relying on them.
Key Questions
What is Nemotron 3.5 Transcribe?
It is an open-weight speech model from NVIDIA that supports transcription, speaker attribution, translation and audio question answering.
Which languages does it support?
NVIDIA says it transcribes 40 languages, translates 31 to or from English, and supports speaker attribution in 13. Coverage varies by task, and audio question answering is trained in English, according to the source.
Can developers adapt it to specialized vocabulary?
NVIDIA says developers can guide output with natural-language prompts and fine-tune the model using Granary v2 data and NeMo Speech tools. The reported medical improvement comes from one case study and should not be assumed to apply elsewhere.
Do the leaderboard results predict accuracy on every recording?
No. The reported scores apply to specific evaluations. Accuracy in other settings may differ with audio quality, accents, overlapping speech and domain vocabulary, so users should test and review outputs against the source recording.
Primary source: Hugging Face · via ThorstenMeyerAI.com
