🔍 Read the full analysis: NVIDIA Nemotron 3.5 Transcribe And Granary V2: AI Transcription You Can Tailor 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 reports strong results on named benchmarks and a 45.8% held-out-speaker WER reduction in a medical case study, but those results do not establish performance across other settings.
NVIDIA has released Nemotron 3.5 Transcribe, an open-weight speech model available on Hugging Face and ModelScope, alongside Granary v2 speech data and tools for adapting the model. The checkpoint combines transcription, speaker attribution, translation and audio question answering, giving developers one model for several speech tasks; the original analysis covers the release, while NVIDIA’s benchmark and case-study results remain company-reported and may not predict performance on other recordings.
NVIDIA says Nemotron 3.5 Transcribe can transcribe 40 languages, translate 31 languages to or from English and attribute speakers in 13. It supports recordings of up to one hour, labels as many as eight speakers and produces word-level timestamps. The company says users can guide output with natural-language prompts, including instructions about terminology and written formats.
The release pairs the model with Granary v2, a multilingual speech-data resource and curation pipeline, and NVIDIA NeMo Speech tools for fine-tuning. Developers can use prompts to guide outputs or train the model on vocabulary and speech suited to an application. NVIDIA’s 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, 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. These are specific benchmark results, not guarantees for other audio or evaluation methods.
One Checkpoint, Several Speech Tasks
Combining transcription, speaker labels and translation may reduce the need to route a recording through separate models for each output. That could help developers building meeting notes, classroom replays, multilingual call tools or specialized dictation, where a plain transcript may not identify speakers or provide text in the required language or format.
The accompanying data and fine-tuning tools make adaptation part of the announcement, rather than limiting the release to a downloadable checkpoint. Developers can test prompts first, then use Granary v2 and NeMo Speech to tailor a model to domain vocabulary and speech. This offers a route to customization, but its practical value depends on the quality and suitability of training data and on performance in the application’s actual conditions.
For users choosing transcription systems, the announcement is a reason to test an open-weight option, not evidence that it will outperform a service on every task. Accuracy can vary with noise, accents, overlapping speech and specialized terms; speaker labeling and translation also need evaluation separate from basic transcription.
Products Worth Considering
Bose QuietComfort Wireless Bluetooth Headphones with Active Noise Cancellation, Mic, USB‑C Charging, Deep Bass, 24‑Hour Battery – White Smok
Logitech MX Master 4 for Mac, Ergonomic Bluetooth Mouse with Haptic Feedback, Ultra-Fast Scrolling, USB-C Charging, for MacOS, iPadOS, Windo
Sony WH-1000XM6 Wireless Noise-Canceling Over-Ear Headphones, Blue, Bundle
As an affiliate, we earn on qualifying purchases.
From Benchmarks to Adaptation
The release describes separate comparisons for transcription and diarization. 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. These placements refer to the cited evaluations and do not establish results on every dataset or recording type.
NVIDIA also describes a medical case study in which fine-tuning reduced word error rate on held-out speakers by 45.8% relative to the best prompt. That is a result from a particular case study, not a general estimate for medical transcription or other specialties. The supplied information does not detail enough about the data and test conditions to independently assess how widely the result transfers.
The release’s examples include overlapping dialogue from the 1940 film His Girl Friday and a classroom replay developed by Agora using word-level timestamps. They illustrate possible uses, but are demonstrations rather than independent accuracy tests. NVIDIA advises users to check names, speaker changes and generated 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 Across Real Recordings
The published figures do not show how the model performs across every supported language, accent, domain or recording condition. The benchmark summaries provided do not include enough detail to independently assess each test set, comparison method or condition. Leaderboard rankings are limited to the cited evaluations, and results may differ with background noise, overlapping voices, mixed-language speech or specialized vocabulary.
The language counts apply to different tasks: transcription, translation and speaker attribution do not share identical coverage. NVIDIA says audio question answering is trained in English. The supplied material does not establish how well that feature handles other languages, or how consistently it answers questions about recordings in less controlled settings.
The medical case study’s 45.8% reduction is relative to the best prompt in that specific study. It does not establish the same gain across medical specialties, datasets or speakers. More independent testing would be needed to determine how the reported results transfer to routine use.
Testing and Fine-Tuning the Release
Developers can access Nemotron 3.5 Transcribe and Granary v2 on Hugging Face and ModelScope, consult the model card for task and language support, and try the recommended prompts. NVIDIA’s NeMo Speech tools provide a route to fine-tuning for applications with suitable data.
The next useful evidence will come from testing across languages, domains and recording conditions, with evaluations that state their datasets and comparison methods. For teams considering deployment, the immediate step is to compare outputs with the original audio and measure performance on representative recordings before relying on transcripts, speaker labels or generated answers.
Key Questions
What is Nemotron 3.5 Transcribe?
It is an open-weight NVIDIA speech model that combines transcription, speaker attribution, translation and audio question answering in one checkpoint.
Where can developers get it?
NVIDIA has made the model available on Hugging Face and ModelScope. Granary v2 data and fine-tuning resources are also part of the release.
Which languages does it support?
NVIDIA says it transcribes 40 languages, translates 31 to or from English and attributes speakers in 13. The counts describe different tasks, and audio question answering is trained in English, according to the company.
Are the reported accuracy results guaranteed?
No. The reported word error rates and rankings come from specific benchmarks, while the medical fine-tuning result comes from one case study. Performance on other recordings may differ.
Can developers adapt the model to their own terminology?
NVIDIA says developers can use natural-language prompts and fine-tune the model with NeMo Speech tools and suitable speech data. The results will depend on the data, task and recording conditions.
Primary source: Hugging Face · via ThorstenMeyerAI.com
