NVIDIA Releases Nemotron 3.5 Transcribe, an Open Model for Speech

NVIDIA has released Nemotron 3.5 Transcribe, an open-weight model that turns recorded speech into text and can also translate it, label who is speaking and answer questions about it. NVIDIA’s announcement on Hugging Face, published on 28 October, says the model transcribes 40 languages, translates 31 of them to or from English, and attributes speakers in 13. It handles recordings up to an hour long with up to eight speakers.

NVIDIA has also released Granary v2, a multilingual speech dataset and data pipeline used in the model’s training, so developers can adapt it to their own vocabulary.

What the model does

Most speech-to-text tools produce one kind of output. Nemotron 3.5 Transcribe is “promptable”: the user tells it in plain language what to do. NVIDIA’s examples include “Transcribe this recording and label each speaker”, “Translate this audio into English” and “What decisions were made in this meeting?” Users can also give it a list of likely terms, such as drug names, and ask it to use them only when the audio supports them.

The model card describes it as “an open-weight, offline speech and audio model”, released under the OpenMDW-1.1 licence and “ready for commercial or non-commercial use”. It also supports word-level timestamps. NVIDIA says it is “built for recorded audio, not live streaming”, so it suits meetings, lectures and dictation files rather than live captions.

The model pairs a speech encoder with a hybrid Mamba–Transformer mixture-of-experts language model. NVIDIA’s blog says only about 4 billion of its 30 billion parameters are active for each token; the model card puts the figure at about 4.3 billion, including the audio encoder. That design keeps it small enough to run on one NVIDIA GPU, from DGX Spark and Jetson AGX Thor up to data-centre Blackwell systems, according to NVIDIA. It is available on Hugging Face and ModelScope, and through hosted services from Baseten, DeepInfra and Simplismart.

NVIDIA’s benchmark claims

NVIDIA says the model is first among open-weight models on the Hugging Face Open ASR Leaderboard for English, at 5.33% average word error rate (WER). The model card, using results captured on 12 October across nine public test sets, gives 4.71%. NVIDIA also says the model ties for first among open-weight models on Artificial Analysis’s speech-to-text ranking, at 2.8%, with about 4 billion active parameters against 24 billion for the model it ties.

For speaker labelling, NVIDIA says it has the lowest diarization error rate, 18.26% across 21 languages, on David AI’s multilingual leaderboard of 14 systems, and the lowest error rate, 32.4%, in VoiceArena’s 14-system comparison. On speed, NVIDIA says one RTX PRO 6000 Blackwell Server Edition GPU using NVFP4 precision “processes 100 hours of audio in under three minutes”, measured on 30-second clips at high concurrency.

Granary v2 and fine-tuning

The Granary v2 dataset covers 463,062 hours across 48 languages from 16 source corpora, up from 25 languages in the first Granary release. NVIDIA notes these figures describe the dataset release, not the model’s full training mix. For each recording, the pipeline produces three transcript styles, from word-for-word speech with “uh” and “um” to clean written text.

In a medical case study using recordings from Indian doctors and medical students, NVIDIA says fine-tuning cut WER on held-out speakers from 22.44% with the best prompt to 12.17%, “a 45.8% relative reduction”. The full adaptation took about 14 GPU-hours on one NVIDIA GB300. Accuracy on general English fell slightly, with LibriSpeech WER rising from 1.54% to 1.69%.

Why the open release matters more than the rankings

Speech models are often sold as cloud services, which means sending recordings of meetings, patients or customers to someone else’s servers. An open-weight model that runs on one GPU lets organisations keep that audio in-house. NVIDIA’s blog says developers can “choose where their audio is processed”. For hospitals, councils or law firms, that may matter more than a fraction of a point on a leaderboard. The model card rates English (United Kingdom) at its top “Leading” quality level, alongside US and Indian English.

The benchmark claims deserve some caution. They come from NVIDIA, and the blog and model card give different English figures, measured on different sets. The model card also says its training data “includes data licensed from David AI”, whose leaderboard it tops, while stating that the evaluation set was not part of it. And the fastest throughput figures in the model card come from optimised serving code in NVIDIA’s NIM product that is “not yet part of the open-source release”; NVIDIA expects the open vLLM integration to be “roughly two times slower”.

The open data is the more lasting contribution. Granary v2 publishes its curation pipeline as well as the data, so others can rebuild or extend it. The model card notes that some training sources are web-derived, and that its speaker labels are intended only to tell speakers apart within a recording, not to identify real people. Transcribing and labelling speakers in recordings raises privacy questions, and NVIDIA itself advises users to spot-check names, speaker changes and answers against the recording.

How it fits NVIDIA’s open model push

Transcribe is the latest in a run of open Nemotron 3.5 releases, after the Super VL multimodal model, the retrieval models for AI agents and the compact Bolt E4B agent model. It enters a busy field: IBM recently said Granite Speech 5.0 2B was the second-best open model for English speech recognition, and Mistral launched its Voxtral 3 speech models.

The real test will come in the messy audio outside benchmarks, such as noisy wards, regional accents and crosstalk, where NVIDIA’s own advice is to check the transcript against the tape.

Sources

• NVIDIA, “NVIDIA Nemotron 3.5 Transcribe and Granary v2: Intelligent Multilingual Transcription You Can Prompt and Adapt”, Hugging Face blog, 28 October 2026
• NVIDIA, NVIDIA-Nemotron-3.5-Transcribe-30B-A4B model card, Hugging Face
• NVIDIA, “Nemotron 3.5 Transcribe Technical Report”, NVIDIA Research
• NVIDIA, Granary v2 dataset, ModelScope

Related stories

• NVIDIA’s Nemotron 3.5 Bolt E4B Packs an Agent Model Into About 4GB
• NVIDIA Releases Five Open Nemotron 3.5 Retrieval Models for AI Agents
• NVIDIA Releases Nemotron 3.5 Super VL, an Open 120B Multimodal MoE for Agents
• IBM Says Granite Speech 5.0 2B Is Second-Best Open Model for English ASR