Nvidia releases Nemotron 3.5 Transcribe for audio in 40 languages
Developers can download the checkpoint or try a demo, then use one model to transcribe, translate, label speakers and answer questions about recordings.
Loading… Բեռնվում է…
Developers can download the checkpoint or try a demo, then use one model to transcribe, translate, label speakers and answer questions about recordings.
Nvidia released Nemotron 3.5 Transcribe on 28 October, giving developers one downloadable speech model for transcription, translation, speaker labels and questions about recordings. Its accompanying Granary v2 release includes public lists covering roughly 463,062 hours of audio, alongside tools for adapting the model to specialist speech.
Nvidia says the model replaces a workflow that needs separate models for transcription, speaker labels and translation. Developers guide it with plain-language instructions, such as asking for a transcript with speaker labels or a list of decisions made in a meeting.
SiliconANGLE said a downloadable checkpoint—a saved copy of the model—and a demo are available on Hugging Face. Open weights give developers access to the model’s learned settings, allowing them to run and adapt it rather than only use a hosted service. Nvidia says developers can choose whether to process audio locally or in the cloud.
Granary v2 expands the speech data available for adaptation. The Decoder gives totals for the full dataset, while The New Stack describes the smaller set covered by the public manifests—lists identifying recordings in the release.
| Release or subset | Audio hours | Languages | Scope |
|---|
The W27 keeps the Galaxy Z Fold 8 Ultra’s core hardware, with dark-red or dark-blue finishes, golden accents and 16GB of working memory in both storage versions.
2 min
| Granary v1 | About 643,000 | 25 | Earlier dataset total |
| Granary v2 | About 844,000 | 48 | Dataset total |
| Granary v2 public manifests | Roughly 463,062 | 48 | Public lists spanning 16 source collections |
Neither the public-manifest tally nor the dataset total describes Transcribe’s complete training mixture or how heavily each source was sampled. Nvidia says the released dataset does not fully match the model’s training data.
Granary provides transcript styles that retain filler sounds such as “um”, format spoken numbers as digits or remove fillers. Nvidia also released a reusable data-preparation pipeline that developers can run on their own recordings before adapting the model.
Nvidia says Transcribe handles recordings up to one hour long, as TechCrunch detailed. It works with recorded audio, not live streams. Its 40-language transcription coverage does not apply equally to every feature.
| Task | Coverage |
|---|---|
| Transcription | 40 languages, including seven Indic languages |
| Translation | 31 languages to and from English |
| Speaker labels | 13 languages; up to eight speakers |
| Questions about audio | Trained in English |
Analysis
Transcribe is a 30-billion-parameter mixture-of-experts model: it activates only part of its learned numerical settings for each processing step. Nvidia specifies approximately 4.3 billion active parameters.
Nvidia confirms that the weights are downloadable on Hugging Face in BF16, FP8 and NVFP4 formats, which use different levels of numerical precision. Supported hardware includes RTX Spark and DGX Spark systems, Jetson AGX Thor devices and GeForce RTX 5090 graphics processors. Nvidia says the model can run on a single graphics processing unit, or GPU, with deployment options extending to Blackwell data-centre hardware, depending on compression.
Nvidia says a Blackwell Server Edition GPU using NVFP4 precision and one-quarter model compression can process 100 hours of audio in under three minutes. That processing-speed claim does not change the one-hour limit for an individual recording.
Nvidia also says Transcribe ranks first on Hugging Face’s Open ASR transcription leaderboard and second among open-weight models on Artificial Analysis. These rankings are the company’s stated performance comparisons.
Analysis
Developers can supply medical terms or other specialist vocabulary in a prompt. Fine-tuning goes further by training the model on examples from a particular field. Nvidia tested this approach on recordings of Indian doctors and medical students, using mostly synthetic speech for adaptation.
Nvidia’s case study separates full-benchmark results from tests on held-out speakers—people whose recordings were not used for adaptation. The specialist test is identified as the Eka Care Indian medical speech benchmark. Word-error rate measures missed, added or substituted words; lower is better.
| Test group | Condition | Word-error rate |
|---|---|---|
| Full medical benchmark | Baseline | 30.8% |
| Full medical benchmark | More precise prompt | 23.8% |
| Full medical benchmark | After specialist fine-tuning | 11.8% |
| Held-out speakers | Best prompt | 22.44% |
| Held-out speakers | After specialist fine-tuning | 12.17% |
For held-out speakers, Nvidia says the reduction was 45.8% relative to its best prompt. The complete adaptation workflow took about 14 GPU-hours on a single Nvidia GB300.
The Decoder said the adapted model’s error rate rose slightly on general English speech, while a separate medical speech model remained more accurate on brand-name drugs. Nvidia recommends setting aside at least 100 real, verified recordings to test a custom version, and spot-checking names, speaker changes and answers.
The satellite operator expects access to more than twice OneWeb’s current bandwidth in return for spending in 2027–2034, with initial IRIS² services expected from 2030.
3 min
The black, gold and green edition follows September’s sold-out preorders. Nintendo’s replenished launch-day stock also quickly sold out.
1 min