Nvidia's open Nemotron 3.5 Transcribe tops speech benchmarks on one GPU

Nvidia's new open speech model, Nemotron 3.5 Transcribe, ranks first among open-weight models on Hugging Face's English Open ASR Leaderboard with an average word error rate of 5.33 percent — and it runs on a single GPU. The release, published on Hugging Face alongside a demo and a fine-tuning toolkit, bundles transcription, translation, and speaker recognition into one instruction-following model. It also arrives with Granary v2, an 844,000-hour speech dataset, in a direct push to make specialized, self-hosted speech AI cheap enough for developers to build themselves.

What does the model actually do?

A plain transcript doesn't tell you who said what. It won't translate, and it won't spell technical terms correctly. Nemotron 3.5 Transcribe is built to handle all three in a single model, and it can answer questions about a recording's content as well.

The stated feature set:

• Transcription in 40 languages
• Translation to or from English in 31 of them
• Speaker identification in 13 languages
• Recordings up to one hour long, with up to eight speakers
• Audio question-answering about the recording's content

Users steer the model with text instructions. Depending on the prompt, the same meeting recording can become a transcript with speaker labels, an English translation, or a summary of the decisions made. Users can pass along technical terms, such as drug names, as hints — and the model is only supposed to use them if they actually appear in the recording.

The instruction-following design is the point: one model replaces a pipeline of separate transcription, translation, and diarization (speaker attribution) tools. The trade-off, Nvidia says, is that the audio question-answering feature was trained in English only, and features vary by language.

How fast is it, and on what hardware?

The model has 30 billion parameters, but only about 4 billion are active per processing step — a mixture-of-experts-style design (only a fraction of the network fires for each input). Nvidia says that keeps the whole system on a single GPU, from the compact DGX Spark desktop unit up to data center hardware.

The company's throughput claim: 100 hours of audio processed in under three minutes on an RTX PRO 6000 Blackwell Server Edition at NVFP4 precision (a four-bit number format that trades a little accuracy for speed). Nvidia did not publish comparable figures for other hardware configurations, so the headline number applies to one specific, high-end setup.

On Hugging Face's English Open ASR Leaderboard, Transcribe ranks first among open-weight models at 5.33 percent average word error rate. In Artificial Analysis' comparison it scores 2.81 percent — close to the best open model, Mistral's Voxtral Small at 2.77 percent, but with roughly a sixth of the active parameters. According to Nvidia, several widely used commercial speech services trail behind, though the company did not name them.

The lead is largest in speaker recognition. VoiceArena tested 14 systems on conversations where speakers sometimes talk over each other; Transcribe took first place. It also dropped the fewest words, missing 8.32 percent against a median of about 31 percent. On David AI's benchmark, which measures speaker attribution across 21 languages, it posted the lowest error rate — finishing just ahead of Nvidia's own dedicated diarization model.

Takeaway: the benchmark claims are strong, but all of them are Nvidia's own reporting of third-party leaderboards, and the unnamed "commercial services" comparison should be read as marketing until independently verified.

What is Granary v2?

Alongside the model, Nvidia is releasing Granary v2, a training dataset of about 844,000 hours of speech in 48 languages — up from roughly 643,000 hours across 25 languages in the previous version. Each recording ships in several transcription styles: one verbatim, with filler words like "uh" and spelled-out numbers; another cleaned up, with no fillers and digits instead of words.

Nvidia is also sharing the tools it used to build the dataset, so developers can run them on their own recordings. One caveat comes from the company itself: the dataset does not fully match the model's actual training data. That limits how precisely outsiders can reproduce or extend the base model.

The stakes are practical. Speech data for specialized fields — medicine, law, engineering — is scarce and expensive, and Nvidia's bet is that cheap fine-tuning on open infrastructure will pull developers away from closed commercial APIs.

Can it be specialized, and at what cost?

Nvidia tested targeted fine-tuning on recordings of Indian doctors and medical students. On the full benchmark, a more precise prompt cut the error rate from 30.8 to 23.8 percent. After fine-tuning, the rate fell to 11.8 percent. The training data was mostly synthetic speech.

For speakers the model hadn't heard before, Nvidia separately reports a drop from 22.44 to 12.17 percent — a relative improvement of 45.8 percent. That put the adapted model ahead of an open medical speech model, which scored 13.79 percent. Nvidia says the whole process took about 14 GPU hours on a single GB300.

The specialization comes at a cost. The error rate rose slightly on general English speech, and the medical model stayed more accurate on brand-name drugs. Nvidia recommends setting aside at least 100 real, verified recordings for testing when building custom versions — a quiet admission that synthetic-training gains don't transfer automatically.

The base model has limits too. It is built only for recorded audio, not live streams, and Nvidia says users should spot-check names, speaker changes, and answers. In short: the model is a strong starting point, not an autonomous transcription clerk.

How does it fit Nvidia's roadmap and the market?

Nvidia only recently introduced Nemotron 3 Diarization, a dedicated speaker recognition model, in September. It can tell apart up to eight speakers, even in live conversations, but it needed a separate model, such as Nvidia's Parakeet, for the actual transcription. Nemotron 3.5 Transcribe combines both jobs and adds translation and audio understanding — but it does not support real-time processing, so the older diarization model keeps a use case.

Nvidia had already used the Granary dataset for its multimodal Nemotron 3 Nano Omni, so the dataset release extends an existing internal investment rather than starting a new one.

Competitors are setting different priorities. Mistral's Voxtral family also combines transcription and audio understanding, and with the recently announced Voxtral 3, Mistral is focusing on real-time voice agents — the interactive use case Nemotron 3.5 Transcribe explicitly does not cover. The model, dataset, fine-tuning guides, and an extension that lets an AI coding assistant run the training workflow are all available on Hugging Face now.

The open question is whether a one-GPU, open-weight model that beats commercial services on Nvidia-selected benchmarks will pull serious developers away from hosted APIs. The 14-GPU-hour fine-tuning result suggests the economics are on Nvidia's side; the unanswered question is whose benchmarks hold up outside the company's own testing.