NVIDIA releases Nemotron 3.5 Transcribe, an open-weight speech model that transcribes 40 languages

What NVIDIA released

NVIDIA's Hugging Face blog announced Nemotron 3.5 Transcribe on 28 October 2026, describing it as an open-weight, promptable model that turns recorded speech into useful text. It is available on Hugging Face and ModelScope, paired with the Granary v2 dataset and curation pipeline.

The model combines four capabilities in one checkpoint: transcription, speaker attribution (diarization), translation, and audio question answering. Developers control the output with natural language, choosing what to format, how to label speakers and what to extract from a recording.

Capabilities and scale

Nemotron 3.5 Transcribe transcribes 40 languages, translates 31 of them to or from English, and attributes speakers in 13. It handles recordings up to an hour long and labels up to eight speakers, with word-level timestamps supported. Audio question answering is trained in English.

The architecture is a hybrid Mamba-Transformer mixture-of-experts: a 30B-parameter model with about 4B active parameters per token. That efficiency puts it on Artificial Analysis's open-weight Pareto frontier of accuracy versus active parameters, and NVIDIA says it leads the Hugging Face Open ASR leaderboard in English at 5.33% average WER among open-weight models.

Granary v2 and fine-tuning

Granary v2 is the data story behind the model: NVIDIA published multilingual speech data with manifests covering 463,062 hours and a reusable curation pipeline, so teams can build their own training sets without starting from scratch.

The announcement highlights a medical case study in which fine-tuning with synthetic data cut word error rate on held-out speakers by 45.8% relative to the best prompt. The idea is that teams can adapt recognition to domain vocabulary, accents and the written form their applications need.

The model card includes recommended presets, a full prompt guide and a task-by-language matrix, with the caveat that users should spot-check names, speaker changes and answers against the recording, as with any speech model.

Why it matters

Open-weight speech models have lagged proprietary ASR APIs on quality, but NVIDIA's independent score of 2.8% AA-WER (tying first among open-weight models with about 4B active parameters) suggests the gap is closing. The promptable design also changes how transcription is done: instead of post-processing text with a separate model, the same model can be steered by instructions.

For developers, the model can run where they control the hardware, which matters for privacy-sensitive applications like medical, legal and meeting transcription. The combination of transcription plus speaker labels plus translation in one checkpoint simplifies pipelines that previously required several models.

What to watch

Watch how the model performs on long-form and accented audio in independent testing, beyond the vendor-reported leaderboard positions.

Watch whether the fine-tuning recipe proves practical for teams without speech expertise, and whether Granary v2 becomes a standard resource for speech data curation.

Watch where the open-weight ASR frontier moves next, since NVIDIA has now released three Nemotron 3.5 model families in quick succession.

How the model works

Speech-augmented language models pair a speech encoder with a large language model, which is what lets Nemotron 3.5 Transcribe follow natural-language instructions. Conventional speech recognition systems fix one output convention at training time; a promptable model can be told to change formatting, add speaker labels or translate on the fly.

NVIDIA's technical report describes the design as a speech encoder feeding a hybrid Mamba-Transformer mixture-of-experts language model. The mixture-of-experts design means the full 30-billion-parameter model routes each token through only about 4 billion active parameters, which keeps inference costs closer to a much smaller model.

Real-world use cases

The announcement cites Agora, a real-time voice and video platform, using word-level timestamps in a classroom demo: a recorded lecture becomes a transcript shown alongside the replay so students can see what was said and when. That is the kind of use case that makes the model useful beyond a raw ASR drop-in.

For call centers, medical dictation, media production and meetings, the combination of transcription, speaker labels and translation in one model removes the need to stitch together separate tools. Running it open-weight also means organizations can keep audio on their own hardware, which matters for regulated industries.

Media and education organizations that produce multilingual content are another natural fit: transcription in 40 languages and translation in 31 means a single pipeline can caption, subtitle and localize audio without chaining several proprietary services. The open weights also leave the door open to community fine-tunes for specific dialects, domains and output conventions that no vendor would support out of the box.

Benchmark caveats

The leaderboard positions are based on NVIDIA's own reported evaluations, so independent reproduction is still worthwhile. The Hugging Face Open ASR leaderboard figures, the Artificial Analysis scores, and the 45.8% medical fine-tuning result all come from NVIDIA's announcement rather than from third-party verification.

That said, Artificial Analysis is an independent outfit, and its inclusion of the model on the open-weight Pareto frontier of accuracy versus active parameters gives the 4B-active-parameter claim some external grounding. The practical test is how the model behaves on the messy audio your application actually encounters.