What Nvidia released
Nvidia Nemotron Transcribe — the full release name is Nemotron 3.5 Transcribe — is a free, open-weight speech-to-text model that transcribes recorded audio in 40 languages, and it arrives with an open dataset for customizing it. Released on 28 October 2026 alongside Granary v2, Nvidia’s updated open speech dataset, it ranks first among open-weight models on the Hugging Face Open ASR Leaderboard for English, and it gives developers the option of running transcription on their own hardware instead of paying per-minute cloud APIs.
The model is published on Hugging Face and ModelScope, with weights released under the OpenMDW-1.1 license and cleared for commercial use. Hosted endpoints are offered by Baseten, DeepInfra, and Simplismart, and there is an interactive demo on Hugging Face Spaces. What makes this release more than a model drop is the rest of the stack that comes with it: the Granary v2 dataset, the NeMo Speech fine-tuning tools, and a prompt-based interface that lets developers steer the model’s output without retraining it.
What Nvidia Nemotron Transcribe can do
Transcribe turns recorded speech into text in 40 languages, translates 31 of them to and from English, and attributes speakers in 13 languages. It handles recordings up to one hour long and labels up to eight speakers. One checkpoint does four jobs — transcription, speaker attribution (diarization), translation, and audio question answering — and the output is chosen with a natural-language prompt. The same recording can be turned into different outputs depending on the instruction:
Transcribe this recording and label each speaker.
Translate this audio into English.
What decisions were made in this meeting?Language support differs by task, and audio question answering is trained in English, so the model is not equally capable in every language for every job.
Under the hood, Transcribe is a 30-billion-parameter model built on a hybrid Mamba-Transformer mixture-of-experts architecture, with only about 4 billion parameters active per token. Nvidia says it can run on a single chip, from the Jetson AGX Thor robotics controller and DGX Spark systems up to Blackwell datacenter GPUs. On throughput, Nvidia reports that the model processed 100 hours of audio in under three minutes on one RTX PRO 6000 Blackwell Server Edition GPU, using the forthcoming NVFP4 quantized version and measured on 30-second clips at high concurrency. The available checkpoint can be served with vLLM today; local deployment via NeMo-Speech.cpp and the quantized versions are coming soon.
How it compares with existing options
Transcribe is not Nvidia’s first open voice model. It joins Parakeet, the company’s earlier open model at 1.1 billion parameters — described at launch as super small yet powerful. The established open baseline for most developers is still OpenAI’s Whisper, and the research available for this article does not establish an independent, like-for-like comparison with Whisper, so any claim that one beats the other on real recordings is premature.
What Nvidia has published is leaderboard results on public test sets, and they are strong. On the Hugging Face Open ASR Leaderboard, Transcribe ranks first among open-weight models in English at 5.33% average word error rate, with 4.62% mean WER across seven languages in the multilingual track and 10.98% in the long-form track. Artificial Analysis independently ranks speech-to-text models by its own AA-WER index, where Transcribe ties Mistral’s Voxtral Small for first among open-weight models at 2.8% AA-WER — with about 4 billion active parameters against 24 billion for the model it ties. On speaker attribution, Transcribe posts the lowest diarization error rate of 14 systems on David AI’s 21-language benchmark at 18.26%, just ahead of Nvidia’s own dedicated Nemotron 3 Diarization model at 18.38%, and ranks first in VoiceArena’s 14-system comparison at 32.4% cpWER, dropping only 8.32% of reference words against a median of about 31%.
These are benchmark results on public test sets, not a guarantee of accuracy on any particular recording — Nvidia’s own announcement advises spot-checking names, speaker changes, and answers against the recording. And one boundary matters more than any benchmark: Transcribe is built for recorded audio, not live streaming. ElevenLabs’ Realtime and other live-speech tools occupy the real-time ground this model does not cover, and there is no streaming version announced. Readers should also not confuse Transcribe with a different, smaller Nvidia model — the 600-million-parameter Nemotron 3.5 ASR streaming model reported in June 2026 is a separate model for a different job.
The release lands in a busy month for voice AI: Mistral shipped Voxtral 3 earlier in October, ElevenLabs launched Realtime for instant speech in voice agents, and Salesforce released Launa-Omni. Transcribe is the recorded-audio entry in that wave.
The open dataset and customization
Granary v2 is the open multilingual speech dataset and reusable curation pipeline behind Transcribe, built with NVIDIA NeMo Curator. Its public manifest covers 463,062 hours across 48 languages and 16 source corpora, up from 25 languages in Granary v1. Nvidia notes these figures describe the dataset release, not Transcribe’s full training mixture or its sampling weights. A high-fidelity filter keeps only the best recordings: for the VoxPopuli corpus, it reduced 354,000 hours to 111,000 hours across 22 European languages.
Granary v2 generates three transcript forms for each recording, the same verbatim and clean styles developers can request from Transcribe with a prompt:
I paid, uh, twenty-five euros.
I paid, uh, €25.
I paid €25.Beyond transcription, the pipeline produces short audio captions, code-switching restoration of loanwords, speech-grounded question-answer pairs, contextual biasing terms, and translations with automatic quality scores.
Customization starts with prompts, not training. For unfamiliar terms, developers can supply context terms before transcribing — in a healthcare setting, for example:
Transcribe this medical discussion in the original language. Possible terms: HbA1c, metformin. Use these terms only when supported by the audio. Preserve standard medical spelling and write numbers in conventional written form.When prompts are not enough, NeMo Speech fine-tuning pairs representative speech with explicit written targets. A medical case study on the Eka Care benchmark of Indian medical speech shows the progression: naming English and the Latin alphabet in the prompt cut full-benchmark word error rate from 30.8% to 23.8%, and fine-tuning cut word error rate on held-out speakers by 45.8% relative to the best prompt. That is the practical value of an open dataset — a team can adapt the model to the vocabulary its own recordings actually contain, rather than waiting for a vendor to support it.
Who it is for
The obvious audience is meeting transcription tools that process recorded calls after the fact — the use case Nvidia names first. A recorded meeting is exactly the input Transcribe is built for: up to an hour long, multiple speakers, and a transcript that needs to be accurate, searchable, and correctly attributed.
Medical transcription is the second named use case, and the one where the customization story matters most. Medical speech is full of drug names, abbreviations, and specialist terms that generic models handle poorly, and the work often has to happen on premises rather than through a cloud API. An open-weight model that can be fine-tuned on a clinic’s own vocabulary and run on its own hardware fits that constraint in a way a per-minute API does not.
The third audience is builders who want to run transcription on their own hardware instead of paying per-minute cloud APIs. For teams processing large volumes of recorded audio — compliance records, call archives, lecture libraries — avoiding per-minute billing is a reason to evaluate self-hosting, though the costs of hardware, utilization, and operations need to be assessed against a team’s own workload. Nvidia reports the model processed 100 hours of audio in under three minutes on one RTX PRO 6000 Blackwell Server Edition GPU, using the forthcoming NVFP4 quantized version and measured on 30-second clips at high concurrency — a vendor benchmark result, not a guarantee for every workload.
Limits and open questions
The biggest limit is the one Nvidia states plainly: recorded audio only. If your product needs live captions, a voice agent that responds in real time, or streaming transcription of an ongoing call, Transcribe is the wrong tool and ElevenLabs’ Realtime or a similar live-speech system is the right one. There is no streaming version announced, and no stated plan for one.
The benchmark results deserve the same care. They are vendor-reported results on public test sets, and Artificial Analysis is an independent ranking — but neither establishes how accurate the model will be on your recordings, in your accents, with your jargon. The honest test is your own audio: run your recordings through the demo or a hosted endpoint, and check names, speaker changes, and domain terms against the original.
Two things should not be read into this release. First, that the model runs on Nvidia silicon is compatibility, not a stated motive — Nvidia has not said why it released the weights for free, and a distribution reading is an inference, not a fact. Second, whether free open weights shift developers off paid cloud transcription APIs is unproven. Self-hosting avoids per-minute API billing, but setup effort, reliability, and support are factors each team should evaluate for its own workload rather than assume.
How to try it
The fastest way in is the interactive demo on Hugging Face Spaces, which needs no setup. For real use, download the checkpoint from Hugging Face or ModelScope and serve it with vLLM; hosted endpoints from Baseten, DeepInfra, and Simplismart are available if you would rather not run infrastructure. Local deployment via NeMo-Speech.cpp and the quantized versions are still coming, so check the model card for the current state before planning a local deployment.
Once it is running, test three things in this order: accuracy on your own recordings rather than demo audio, handling of the jargon your domain actually uses, and coverage of the languages you need — remembering that language support differs by task. Try the prompt-first approach with context terms before reaching for fine-tuning: in the medical case study, naming English and the Latin alphabet in the prompt cut full-benchmark word error rate from 30.8% to 23.8%. If the numbers hold up on your audio, you have a transcription stack you can run and customize under the OpenMDW-1.1 license — and that, more than any leaderboard position, is what this release is about.
