NVIDIA has released Nemotron 3.5 Transcribe, an open speech model that folds transcription, translation, speaker labeling and audio question answering into one checkpoint. It has 30 billion parameters in total, about 4 billion of them active per step, and NVIDIA says it runs on a single GPU, from a DGX Spark up to data center Blackwell chips. Alongside it, the company published Granary v2, an open speech dataset roughly 844,000 hours long.
The catch is in the fine print: it handles recordings, not live streams. That makes it a tool for meeting notes, call archives and medical dictation rather than voice agents. This post covers what it does, the benchmark claims and who made them, and how it fits next to the other speech models we have covered this month, such as Voxtral 3 and Microsoft's MAI-Transcribe-2.
TL;DR
| Question | Answer |
|---|---|
| What is it? | Open speech model from NVIDIA that follows text instructions |
| Size | 30B total parameters, about 4B active per step |
| Languages | Transcribes 40, translates 31 to or from English, labels speakers in 13 |
| Limits | Recordings up to one hour, up to eight speakers, no real-time mode |
| Speed | 100 hours of audio in under three minutes on an RTX PRO 6000 Blackwell Server Edition at NVFP4 (NVIDIA figure) |
| Headline accuracy | 5.33% average WER on the Open ASR Leaderboard, first among open-weight models (NVIDIA figure) |
| Dataset | Granary v2: about 844,000 hours, 48 languages |
| License | Not stated in the coverage we could read; check the model card |
What does Nemotron 3.5 Transcribe actually do?
Most speech recognition models take audio in and give text out. This one takes audio plus an instruction. According to The Decoder's summary of NVIDIA's release, the same prompt interface can yield a speaker-labeled transcript, a translation, a summary, or an answer to a question about the recording. You can also pass term hints, such as drug names, which the model uses only if those words actually occur in the audio.
That is a real shift from the pipeline most teams run today. A typical meeting-notes app chains a transcription model, a separate diarization model that decides who spoke when, a translation model and a language model for the summary. NVIDIA's own Nemotron 3 Diarization, released in September, was a dedicated 100M-parameter speaker model that needed a separate transcriber. The new model absorbs that job for 13 languages.
Language coverage and hard limits
Two speech bubbles joined by a bridge, showing how Nemotron 3.5 Transcribe translates between languages
The numbers reported for the release:
- 40 languages transcribed.
- 31 languages translated to or from English.
- 13 languages with speaker identification.
- Up to one hour per recording and up to eight speakers.
- No real-time processing. TechCrunch's report on the launch states it works with recorded audio, not livestreams.
NVIDIA's own caveats, as relayed by The Decoder: feature support varies by language, the audio question-answering feature was trained in English, and users should spot-check names, speaker changes and answers. That last point matters. An instruction-following model that summarizes a call can produce a fluent summary of something that was not said, so any workflow that feeds legal, medical or financial decisions needs a human spot-check step.
Benchmarks: what NVIDIA reports
All of the figures below come from NVIDIA's release as summarized by The Decoder. TechCrunch notes that the company claims it beats rivals on benchmarks including the Hugging Face Open ASR Leaderboard but gives no scores. Treat them as vendor-reported until you replicate them.
| Benchmark | Result | Comparison |
|---|---|---|
| Open ASR Leaderboard (English, average WER) | 5.33% | First among open-weight models |
| Artificial Analysis WER | 2.81% | Mistral Voxtral Small 2.77%, with roughly six times the active parameters |
| VoiceArena, overlapping speech (14 systems), cpWER | 32.4% | Nvidia Multitalker Parakeet 34.2%, Deepgram Nova-3 71.9% |
| David AI DAI-ASR-I18N, 21 languages, diarization error | 18.26% | Nemotron 3 Diarization 18.38% |
Two readings are worth keeping apart. On plain transcription accuracy, the model is essentially tied with Voxtral Small on the Artificial Analysis test (2.81% against 2.77%), so the story there is efficiency: roughly 4B active parameters against a much larger active footprint. On overlapping speech the margin is larger, and NVIDIA says it missed 8.32% of words where the median system missed about 31%. Overlapping speech is where meeting transcripts usually fall apart, so that is the number to test first with your own recordings.
For other open options, compare against IBM Granite Speech 5.0 and the tiny Bark 3.0 recognizer, which sit at very different size points.
Granary v2: the dataset behind it
The companion release may matter longer than the model. Granary v2 grows NVIDIA's open speech corpus from about 643,000 hours in 25 languages to about 844,000 hours in 48 languages. Each recording comes with several transcription styles: verbatim, with filler words and spelled-out numbers, and cleaned, with fillers removed and digits instead of words. NVIDIA is also sharing the tools used to build the dataset, which lets other teams reproduce the pipeline for languages it does not cover.
One honest note from NVIDIA: the dataset does not fully match the model's own training data, so it is not a recipe for reproducing the checkpoint exactly.
The medical fine-tuning example
NVIDIA's most practical demonstration is domain adaptation. It tested the model on recordings of Indian doctors and medical students:
| Setup | Error rate (full benchmark) |
|---|---|
| Base model, simple prompt | 30.8% |
| Base model, more precise prompt | 23.8% |
| After fine-tuning | 11.8% |
On unseen speakers, error fell from 22.44% to 12.17%, a 45.8% relative improvement, against 13.79% for an open medical speech model. The run took about 14 GPU hours on a single GB300 and used mostly synthetic speech. The trade-offs, per NVIDIA: slightly higher error on general English, and the open medical model was more accurate on brand-name drugs. It recommends at least 100 real, verified recordings to evaluate any custom version, which is sensible advice for anyone tempted to trust synthetic-data gains.
The prompt-only step is the cheap lesson here: going from a vague prompt to a precise one cut error by seven points before any training. If you only have an afternoon, start there.
How to try it
NVIDIA points to three places, all on Hugging Face: the Nemotron speech model collection, the Granary v2 dataset, and a hosted demo Space. Fine-tuning guides are published, and NVIDIA also offers an extension that lets an AI coding assistant run the training workflow. The tooling is the NeMo Speech library, which NVIDIA says supports adapting apps to industry jargon in each language.
A sensible first test, in order:
- Run the demo on a recording you already have a human transcript for.
- Compare word error rate on your hardest segment, usually overlapping speakers or names.
- Try the same file with a term-hint prompt listing your domain words.
- Only then consider fine-tuning, with 100 or more verified clips held out for testing.
Where it fits in the voice-model landscape
October has been crowded. Mistral shipped Voxtral 3 aimed at real-time voice agents, ElevenLabs launched a realtime speech-to-speech model, and Microsoft pushed streaming MAI-Transcribe-2. Those are live-conversation tools. Nemotron 3.5 Transcribe is the opposite end: batch processing of finished audio, where accuracy, speaker labels and cost per hour matter more than latency. NVIDIA's own Nemotron 3.5 Bolt E4B targets small edge devices, so the Nemotron 3.5 family now spans from a 4GB model to a workstation-class speech system.
The single-GPU claim is worth testing too. A model with 30B total and 4B active parameters keeps compute low, but all the weights still need to sit in memory, so check the memory requirement on the model card against your card before planning a deployment.
Who should care, and who can ignore it
The release is most useful for three groups. Teams with large archives of recorded meetings, interviews or support calls get a self-hostable option that keeps audio off third-party APIs, which matters in healthcare, legal and finance settings where recordings are sensitive. Researchers working on low-resource languages get Granary v2 and the pipeline tools, which lower the cost of building a regional speech model. And anyone already running NVIDIA hardware gets a model tuned for it, with NVFP4 precision support on Blackwell cards.
It is less relevant if you build live voice assistants, call-center agents that must answer in under a second, or phone apps with no GPU. For those, a streaming model remains the right choice, and a small on-device recognizer will beat a 30B-parameter checkpoint on memory and battery. The general rule: pick this model when accuracy on finished audio is the product, and pick a streaming model when the conversation is still happening.
What to watch
- License terms. We could not confirm the license from the coverage, so verify before shipping it in a product.
- Independent benchmarks. The 5.33% Open ASR figure is NVIDIA's; the leaderboard itself is public and checkable.
- Non-English quality. Feature support varies by language, so the 13-language speaker labeling and 31-language translation lists matter more than the headline 40.
- Hallucinated summaries. Instruction-following adds a failure mode plain ASR never had.
Related reading
- NVIDIA Nemotron 3 Diarization
- Nemotron 3.5 Bolt E4B
- Voxtral 3 and the Mistral voice agent SDK
- Microsoft MAI-Transcribe-2
- IBM Granite Speech 5.0
- Grok Voice Transcribe 2
Figures and availability are as reported on October 28, 2026 and may change; check the Hugging Face model card for the current license and requirements.