NVIDIA Nemotron 3.5 Transcribe: the open-weight transcription model covering 40 languages, plus Granary v2 (463,000 hours of speech)
🔎 Speech is becoming a battlefield again — and NVIDIA has just opened fire
While everyone is fixated on LLMs, NVIDIA has just delivered what looks like the biggest open-weight offensive of the year in audio. On October 28, 2026, the company released Nemotron 3.5 Transcribe, a family of transcription models covering 40 languages, accompanied by Granary v2, a dataset of 463,062 hours of speech. TechCrunch reported the announcement with the angle "voice AI in vogue" — the timing is no coincidence, as demand is exploding on the voice agent and compliance side.
Since Whisper in 2022, no open model had seriously threatened proprietary APIs on speech recognition. NVIDIA now claims first place among open-weight models on the Open ASR Leaderboard, both in English and multilingual. That's no small detail: it's a shift in the balance of power.
The business logic is crystal clear — give away the weights, give away the data, and sell the silicon that runs it all. And it works, because the numbers are there.
The Essentials
- First open-weight model to top the Open ASR Leaderboard: 4.71% WER in English, 4.62% multilingual (NVIDIA announcements, October 2026).
- SALM 34B architecture: only about 4.3B active parameters, MoE-style hybrid Mamba-2/Transformer backbone (Lightning 30B family).
- 40 languages transcribed, including 31 translated to and from English; speaker diarization for up to 8 speakers; word-level timestamps; files up to 1 hour.
- OpenMDW-1.1 license: commercial use permitted, weights freely downloadable.
- From Jetson AGX Thor to Blackwell: fits on a single embedded chip, 2,238 seconds of audio processed per second on RTX PRO 6000.
- Granary v2: 463,062 hours of speech in 48 languages (vs. 25 for v1), free of charge.
- Acknowledged limitation: offline processing only, no real-time streaming.
Recommended Tools
| Tool | Main use | Price (October 2026) | Ideal for |
|---|---|---|---|
| Nemotron 3.5 Transcribe | ASR + translation + diarization | Free (open-weight, OpenMDW-1.1) | Confidential production, large local volumes |
| Whisper | Batch ASR | Free (MIT) | Rapid prototyping, simple scripts |
| Qwen3-ASR-1.7B | Embedded ASR | Free | Edge, mobile, constrained devices |
| Gemini 3.5 Transcribe | Cloud ASR, 85 languages, streaming | Usage-based API (pricing on ai.google.dev, October 2026) | Live use, maximum language coverage, zero infra |
| Granary v2 | Training dataset | Free | ASR fine-tuning, research |
A 34B SALM that only activates 4.3B parameters
Nemotron 3.5 Transcribe is built on a 34-billion-parameter SALM (Speech-Aware Language Model) architecture, of which only about 4.3 billion are activated per token. It's a mixture-of-experts model: the capacity of a large model, the inference cost of a mid-sized one.
The backbone is a Mamba-2/Transformer hybrid derived from NVIDIA's Lightning 30B MoE architecture. We break down this recipe in our article on Nemotron 3.5 Super Lightning, open-weight hybrid Mamba-Transformer MoEs — speech applies the same mechanics to the audio signal.
Why the Mamba hybrid changes everything for audio
Transformer attention scales quadratically with sequence length. And audio is hungry: one hour of sound represents tens of thousands of frames. Mamba-2 layers, which process sequences at linear cost, make it possible to ingest recordings up to 1 hour in one go, without manual chunking.
The second key piece: a CTC adapter that produces word-level timestamps. You know precisely who said what, and when. For subtitling or video editing, it's the kind of detail that saves entire days of post-production.
4.71% WER: first open-weight model on the Open ASR Leaderboard
On Hugging Face's Open ASR Leaderboard, Nemotron 3.5 Transcribe becomes the first open-weight model to claim the top spot in both the English and multilingual rankings, with a WER (word error rate) of 4.71% in English and 4.62% on average across multilingual benchmarks (NVIDIA announcements, October 2026).
For context: Whisper large-v3, the long-standing open-source reference, sits above 6% average English WER on this same type of benchmark. The gap may seem small; over a corpus of 10,000 words, it amounts to more than a hundred fewer errors.
One honest journalistic caveat: leaderboard benchmarks use clean, often read-out audio. Your noisy Zoom meetings, your podcasts recorded in a coffee shop, and your phone interviews belong to a different world. Test on your own data before promising anything to a client.
Diarisation, translation, word-level timestamps: the complete package
Beyond raw WER, it's the completeness of the package that commands respect: a single model handles transcription, speaker identification, fine-grained timestamping, and translation.
Concretely:
- 40 languages transcribed, 31 of which are available for translation to and from English — enough to produce bilingual subtitles in a single pass.
- Diarization for up to 8 speakers, with segment attribution. Six-person meetings are no longer a nightmare.
- Word-level timestamps via the CTC adapter, essential for editing and indexing.
- Files up to 1 hour processed in one go, thanks to the hybrid backbone.
It's this package that distinguishes a production model from a benchmark toy. Whisper does none of these things natively: you have to pair it with a diarization tool, an aligner for timestamps, and an LLM for translation. Here, everything is integrated — and the rest of the pipeline plugs in directly: summarization, action item extraction, agentification with an LLM like Claude Sonnet 4.6 or GPT-5.4. Our dossier on the best LLMs for AI agents covers this upper layer in detail.
Granary v2: 463,062 hours of speech in 48 languages, free
The dataset may be the real bombshell of the announcement. Granary v2 provides 463,062 hours of annotated speech covering 48 languages — compared to 25 for the first version of the corpus.
In ASR, data matters more than architecture. Anyone can finely reproduce the weights; what separates a good model from an excellent one is hundreds of thousands of hours of clean, segmented, aligned audio with identified speakers. Producing this data manually costs a fortune — Granary v2 was largely built through pseudo-labeling and automatic filtering pipelines, the standard method used by labs today.
NVIDIA is giving all of this away for free, and it's an unstoppable strategic move. Teams training or fine-tuning a speech model now start from an industrial-quality foundation: academic labs, vertical startups (medical, legal, call centers), and researchers working on low-resource languages are the first winners.
Deploying Nemotron 3.5 Transcribe: from Jetson AGX Thor to Blackwell
The model can be downloaded freely and runs on a surprisingly wide range of hardware. The remarkable part: NVIDIA validates operation on a single Jetson AGX Thor chip, its embedded module — enough to target robotics, automotive, or edge computing.
At the other end of the spectrum, on an RTX PRO 6000 (Blackwell), NVIDIA reports 2,238 seconds of audio processed per second of compute — roughly 37 minutes of sound analyzed for every second of clock time. At that rate, transcribing a night's worth of recordings becomes a matter of minutes.
pip install "nemo_toolkit[asr]"
# weights and model cards available on the Hugging Face Hub (nvidia organization)
Integration goes through the NeMo stack. If you've already set up a local LLM environment, our guide to installing an LLM locally and our selection of the best LLMs to run locally will give you the right habits: the logic is identical, only the runtime changes. And for the application surrounding the model — API, queue, web interface — a simple VPS is more than enough; Hostinger works great for getting started without breaking the bank.
Still, keep one hardware reality in mind: 34B total parameters is still a big animal. The 4.3B active parameters make inference fast, but fitting the full weights will require a pro-grade card or aggressive quantization.
The real limitation: no streaming, offline only
Nemotron 3.5 Transcribe doesn't do real time. It transcribes files, not streams: as it stands, it's impossible to subtitle a live broadcast, caption a live video call, or feed a voice agent that responds in a second.
This is the most important limitation of the announcement, and NVIDIA isn't hiding it. For live transcription, proprietary APIs retain the advantage — that's precisely the territory of Gemini 3.5 Transcribe, whose speed of execution Google highlights.
Workarounds exist: split the stream into 30- to 60-second segments and feed them to the model in parallel. But each split adds latency, breaks the acoustic context, and weakens diarization at the boundaries. For quasi-live with a few seconds of delay, it works. For true live, it doesn't.
Facing Off: Gemini 3.5 Transcribe, Qwen3-ASR, and Whisper
The choice boils down to three trade-offs: privacy versus simplicity, 40 versus 85 languages, and zero marginal cost versus an API bill.
| Model | Languages | English WER | License | Streaming | Price (October 2026) |
|---|---|---|---|---|---|
| Nemotron 3.5 Transcribe | 40 (+31 translation pairs) | 4.71% | OpenMDW-1.1 (commercial) | No | Free (self-hosted) |
| Gemini 3.5 Transcribe | 85 | ~2.6% (Google's claim) | Proprietary | Yes | Usage-based API (check ai.google.dev) |
| Whisper large-v3 | ~99 | > 6% | MIT | No | Free |
| Qwen3-ASR-1.7B | Multilingual | n/a (lightweight model) | Open | No | Free |
Google plays three cards that Nemotron can't touch: 85 languages, a claimed WER of around 2.6%, and execution that's 70% faster than the previous generation. We cover this offering in detail in our article on Gemini 3.5 Transcribe, Google's 85-language STT engine.
My honest take: if your data can leave your infrastructure and you want live transcription or maximum language coverage, Gemini is objectively stronger. If you work in healthcare, legal, or defense — in short, anything that can't leave the premises — Nemotron 3.5 Transcribe is the best open option in recent memory, and it costs nothing in licensing. Qwen3-ASR-1.7B plays a different tune: the tiny model for embedded systems, when every megabyte counts. And Whisper? It blazed the trail in 2022 and remains the easiest to deploy, but it has been technically outclassed on every front that matters.
NVIDIA's strategy: give away the software, sell the silicon
None of this is philanthropy. NVIDIA is methodically building a Nemotron ecosystem that now covers text, vision, and speech — and every free model sells GPUs.
The family is growing at an impressive pace: Nemotron 3 Ultra 550B, the largest American open-weight model on the LLM side, Nemotron 3.5 Super VL 120B for multimodal agents on the vision side, and now speech. The open-weight wave even spills beyond text: on the image side, FLUX 3 Image Klein runs an open-weight model on a simple 8 GB card.
The end of the pipeline is the voice agent: transcription by Nemotron, reasoning by an LLM — Claude Sonnet 4.6 for responsiveness, Kimi K2.6 self-hosted for sovereignty — then speech synthesis. A fully local variant exists, which we detail in our guide on open source AI agents with Ollama. TechCrunch noted in late October 2026 investors' appetite for voice AI; NVIDIA has just made sure that the majority of these agents will run on its chips.
❌ Common Mistakes
Mistake 1: confusing open-weight and open-source
The weights are downloadable and the OpenMDW-1.1 license allows commercial use, but everything else — full training code, source data, warranties — follows rules you need to read. Check the terms on the model card before committing a product to it.
Mistake 2: expecting real-time
No native streaming, period. If your use case is live, go with a proprietary API or accept the latency of segment-based chunking. Better to warn the client now than to see the POC fail in a demo.
Mistake 3: extrapolating the English WER to your language
The 4.71% is an English score obtained on clean benchmarks; the 4.62% multilingual figure is an average that masks significant disparities across languages. Test on your own recordings — accents, noise, industry jargon — before any contractual commitment.
Mistake 4: neglecting audio preparation
A model with 4.71% WER can drop to 15% on mono audio crushed down to 8 kHz. Normalize, denoise if necessary, segment intelligently — and remember that diarization caps out at 8 speakers: beyond that, attributions degrade quickly.
❓ Frequently Asked Questions
Is Nemotron 3.5 Transcribe free for commercial use?
Yes. The OpenMDW-1.1 license explicitly permits commercial use of the weights; you only pay for the hardware or cloud that runs the model. Still, read the full terms on the Hugging Face model card before any production deployment.
What hardware is required at minimum?
NVIDIA validates the Jetson AGX Thor as the minimum for embedded use. On the workstation side, plan for a pro card to house the 34 billion parameters, or use quantization. On an RTX PRO 6000, the model swallows 2,238 seconds of audio per second — enough to process entire archives in minutes.
How does it differ from Whisper?
Whisper (OpenAI, 2022) is a dense model with a 30-second window, without native diarization or timestamps. Nemotron 3.5 Transcribe handles 1 hour in a single pass, identifies up to 8 speakers, timestamps at the word level, translates 31 languages, and posts a WER more than one point lower.
Can it transcribe and translate at the same time?
Yes: 31 languages can be translated to and from English in a single pass. Enough to generate bilingual subtitles, prepare international reports, or feed a localization pipeline without chaining two separate models.
What is Granary v2 actually for?
Training or fine-tuning your own speech model: 463,062 annotated hours, 48 languages, downloadable for free. Useful for specializing an ASR on a specific domain (medical, legal), an under-resourced language, or particular acoustic conditions.
Is there a simpler cloud alternative?
Yes: Gemini 3.5 Transcribe covers 85 languages with streaming and a claimed WER of around 2.6%, billed on a usage basis. Our dedicated article details its performance. For confidential material or large volumes, stick with self-hosted Nemotron.
✅ Conclusion
By releasing a first-of-its-kind transcription model, with 40 languages and 463,000 hours of training data, NVIDIA has just made industrial-grade transcription accessible to anyone with a graphics card — the rest is now merely a matter of integration. To choose the LLM that will process these transcriptions, check out our monthly comparison of the best LLMs.