🔍 Read the full analysis: A Guide To Promptable Multilingual 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 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 results on named speech benchmarks and a 45.8% WER reduction in one medical case study, but those findings do not guarantee performance across languages or recording conditions.
NVIDIA has released Nemotron 3.5 Transcribe, an open-weight speech model for multilingual transcription, speaker attribution, translation and audio question answering. The release also includes Granary v2 speech data and tools for adapting the model, giving developers a package to test and customize rather than a transcription checkpoint alone, as detailed in the original analysis.
NVIDIA says the model transcribes 40 languages, translates 31 to or from English, and attributes speakers in 13. It can process recordings of up to one hour, label as many as eight speakers and return word-level timestamps. The company describes the model as promptable, allowing users to request different outputs or provide terminology and formatting instructions in natural language.
The release combines those tasks in one model checkpoint. A developer could request a speaker-labeled transcript, an English translation or answers to questions about a recording. NVIDIA says audio question answering is trained in English, and language coverage varies by task; the language counts should not be treated as applying equally to every capability.
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. The company describes these as the best results among evaluated open-weight models in the cited comparisons. 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 assurances of accuracy on an individual recording.
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 on David AI and VoiceArena comparisons. According to the release, the model uses Sortformer speaker cues during generation. The source material does not provide enough detail to independently compare all test sets or recording conditions.
The accompanying Granary v2 provides multilingual speech data and a reusable curation pipeline. NVIDIA’s NeMo Speech tools can be used to fine-tune the model. In a medical case study, NVIDIA says training lowered word error rate on held-out speakers by 45.8% compared with the best prompt. That figure describes one reported case, not a result established across medical specialties or other domains.
One Model, Several Speech Outputs
For developers, putting transcription, speaker labels, translation and audio questions in one open-weight model may reduce the need to route a recording through multiple separate systems. That could help applications such as meeting notes, classroom replays and multilingual calls, where a plain transcript may not identify who spoke or provide text in the needed language.
The release also links prompting with model adaptation. Users can first provide vocabulary and output instructions, then use Granary v2 and NeMo Speech tools to tune the system to a domain. NVIDIA’s medical case study suggests that tuning improved results in that particular evaluation, but the reported reduction does not establish what users will see with different datasets, specialties, languages or speakers.
Open weights and accompanying data tools give technical teams a route to evaluate and adapt the system within their own workflows. The practical value will depend on how it performs on their recordings, and on the resources required to run and maintain it.
Products Worth Considering
Bose QuietComfort Wireless Bluetooth Headphones with Active Noise Cancellation, Mic, USB‑C Charging, Deep Bass, 24‑Hour Battery – White Smok
Sonos Era 100 SL Compact Microphone-Free Speaker, Black
Bose QuietComfort Wireless Noise-Cancelling Over-Ear Headphones, Black
As an affiliate, we earn on qualifying purchases.
Prompts, Data and Evaluation
NVIDIA presents natural-language prompting as a way to guide terminology and output format. Its guidance recommends supplying domain terms with context and telling the model to use them only when supported by the audio. This matters because prompting can influence what appears in a transcript, while a supplied term should not be treated as evidence that a speaker actually said it.
The release includes demonstrations, among them 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 illustrate possible applications but are not independent accuracy evaluations. NVIDIA also advises users to check names, speaker changes and generated answers against the original recording.
The reported figures use different measures and comparisons: WER for transcription, AA-WER in Artificial Analysis, and separate comparisons for diarization. Those results should be read within their stated benchmarks. Performance can vary with accents, background noise, overlapping speech and specialized vocabulary, while the supplied material does not establish how consistently benchmark rankings transfer to routine use.
“One open model for multilingual transcription, speaker attribution, translation, and audio Q&A.”
— NVIDIA, in its Hugging Face release
Performance Across Real Recordings
It is not clear from the supplied release material how Nemotron 3.5 performs across every supported language, accent, domain or type of noisy and mixed-language audio. The language counts refer to different tasks, and audio question answering is trained in English, according to NVIDIA.
The benchmark summaries do not provide enough information here to independently assess every test set, recording condition or competing system. The reported 45.8% medical improvement is relative to the best prompt in that case study and applies to held-out speakers in that evaluation. Whether similar gains occur elsewhere remains unconfirmed.
As with other speech systems, generated transcripts, speaker changes and answers may need human review. The demonstrations cited in the release show examples, but do not establish performance in all real-world settings.
Testing and Fine-Tuning Options
Developers can access Nemotron 3.5 Transcribe and Granary v2 on Hugging Face and ModelScope. NVIDIA’s release points to a model card with presets, task-by-language support and prompt guidance. Teams can test the prompts, check outputs against their own recordings and use NeMo Speech tools to fine-tune the model where appropriate.
The next useful evidence will come from independent evaluations across languages, domains and recording conditions, as well as tests of how well tuned models generalize to speakers and recordings they have not seen. Until then, the published benchmarks and medical case study offer starting points for evaluation, not a guarantee of results in deployment.
Key Questions
What is Nemotron 3.5 Transcribe?
It is an open-weight NVIDIA speech model designed for 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 attributes speakers in 13. Coverage differs by task, and the source says audio question answering is trained in English.
What did NVIDIA report about accuracy?
NVIDIA reports leaderboard results including 5.33% average English WER and a 4.62% mean WER across seven languages on the cited Hugging Face comparison. These are benchmark measurements; performance on a particular recording may differ.
What are Granary v2 and NeMo Speech for?
Granary v2 provides multilingual speech data and a curation pipeline. NVIDIA’s NeMo Speech tools support fine-tuning Nemotron 3.5 for a developer’s data or domain.
Does the medical case study mean the model will be more accurate in every clinic?
No. NVIDIA reports a 45.8% WER reduction against the best prompt on held-out speakers in one medical case study. The source does not establish the same improvement across other specialties, datasets or languages.
Primary source: Hugging Face · via ThorstenMeyerAI.com
