Nvidia releases free open speech model that transcribes 40 languages

Nvidia released Nemotron 3.5 Transcribe on October 28, 2026, a free, open-weight speech model that transcribes 40 languages and handles translation, speaker attribution and questions about a recording in a single checkpoint. Alongside it, Nvidia published Granary v2, an open speech dataset and curation pipeline whose public manifest covers 463,062 hours of audio across 48 languages.

The pitch to developers is consolidation. Traditional transcription stacks run separate steps to detect speech, transcribe it, identify and label speakers, and then answer questions about the result; Transcribe does all of this in a single model pass. It targets builders of business-meeting and medical transcription apps, and it comes with two caveats worth knowing up front: it works on recorded audio only, and most of the headline numbers come from Nvidia’s own reporting.

What Nemotron 3.5 Transcribe does in one model

Nvidia says the model transcribes 40 languages, translates 31 of them to or from English, and attributes speakers in 13. It handles recordings up to one hour long and labels up to eight speakers.

Output is controlled with plain-language prompts. Developers can request verbatim or clean written transcripts, supply context vocabulary such as product or medical terms, ask for speaker tags and word-level timestamps, translate speech to or from English, or ask questions about the audio. SiliconANGLE gives examples of prompts the same checkpoint can take, including:

“Transcribe the recording and label each speaker,”

or “What decisions were made in this meeting?”. Context hints such as drug names are meant to be used only if they actually appear in the recording, according to The Decoder.

Speaker labelling is built into the transcript rather than bolted on. SQ Magazine reports that speaker cues from Nvidia’s Sortformer reach the language model alongside the acoustic features, so the model tags each turn as it writes.

Architecture, size and the parameter count sources disagree on

Unite.AI, citing the model card, describes Transcribe as a speech encoder paired with a language backbone in one 34-billion-parameter checkpoint, with about 4.3 billion parameters active per token and a 65,536-token context. The speech side combines a 600-million-parameter Transformer ASR encoder (ASR is automatic speech recognition), the roughly 99-million-parameter Nemotron-3-Diarization speaker encoder, and a connectionist temporal classification head for word timestamps. The language side is Nemotron-3.5 Lightning 30B-A3B, a hybrid Mamba-2–Transformer mixture-of-experts model. A mixture-of-experts model routes each token through only a subset of its parameters, which is why the active count is far below the total.

Other outlets describe the model as 30 billion parameters with about 4 billion active, including Nvidia’s own launch post quoted by SQ Magazine, which calls it “an open-weight, promptable 30B MoE model with ~4B active parameters”. TNW uses the 34B figure. The gap appears to reflect whether the speech components are counted on top of the 30B language backbone, but the sources do not settle it.

According to Unite.AI, weights ship in BF16, with FP8 and NVFP4 quantized variants and an optional DFlash2 draft model for speculative decoding, a technique that uses a small model to propose tokens the large model then checks.

Spec	As reported
Transcription languages	40
Translation	31 languages to or from English
Speaker attribution	13 languages, up to eight speakers
Max recording length	One hour
Parameters	34B total, ~4.3B active; or 30B, ~4B active
Context	65,536 tokens
Licence	OpenMDW-1.1
Mode	Offline, recorded audio only

Benchmarks: Nvidia’s numbers and the third-party leaderboards

Nvidia reports that Transcribe ranks first among open-weight models on Hugging Face’s Open ASR English leaderboard at a 5.33% average word error rate (WER, the share of words a model gets wrong), with API models excluded. SQ Magazine flags that exclusion as a detail worth reading twice. In the model card’s comparison, Nvidia also reports the lowest multilingual result among open-weight models, 4.62% mean WER across seven languages, and the lowest long-form result, 10.98%.

In the card’s own tables, captured October 12, 2026, the nine-dataset English mean is 4.71% WER against 4.94% for Qwen3-ASR-1.7B, the next-best open-weight entry; on multi-speaker audio, the mean cpWER is 15.80 against 16.83 for DiCoW v3.3 paired with DiariZen. cpWER is a word error score counted after matching speaker labels.

On Artificial Analysis’s speech-to-text index, Transcribe scores 2.8% AA-WER, which Nvidia describes as a tie for first among open-weight models while activating about 4 billion parameters per token against 24 billion for the model it ties. The Decoder reads the same comparison slightly differently: it puts Transcribe at 2.81% against 2.77% for Mistral’s Voxtral Small, calling that close to the best open model rather than level with it.

Speaker attribution is where the release looks strongest. On David AI’s 21-language DAI-ASR-I18N diarization track, Nvidia reports the lowest diarization error rate among 14 systems at 18.26%, just ahead of its own Nemotron 3 Diarization at 18.38%. On VoiceArena’s English Diarization-Bench, Transcribe placed first among 14 systems at 32.4% cpWER, deleting 8.32% of reference words against a median of about 31%.

One context point on the David AI result: the model card says some training data was licensed from David AI, while stating that the DAI-ASR-I18N evaluation set was not part of training.

Where and how developers can get it

The model card lists an October 28 release across Hugging Face, GitHub via NVIDIA NeMo Speech, NGC and Build.Nvidia.com, with weights under the OpenMDW-1.1 licence for commercial or non-commercial use. Hosted endpoints are offered by Baseten, DeepInfra and Simplismart, and interactive demos run on Hugging Face Spaces and ModelScope Studios. SQ Magazine reports developers can serve the current checkpoint with vLLM today.

Nvidia says the model runs on a single GPU, from the Jetson AGX Thor robotics computer and DGX Spark to Blackwell data-centre chips. Its showcase throughput figure is 100 hours of audio processed in under three minutes on one RTX PRO 6000 Blackwell Server Edition GPU at NVFP4 precision, measured on 30-second clips at high concurrency.

Sources disagree on whether the compressed builds behind those claims are available yet. SQ Magazine reports that the single-GPU claim depends on quantized versions Nvidia lists as coming soon, that the same caveat covers the 100-hours figure, and that local deployment through NeMo-Speech.cpp waits on that release. Unite.AI, by contrast, reports that FP8 and NVFP4 quantized variants already ship alongside the BF16 weights. Teams planning edge or laptop deployment should check the model card directly before committing.

The model is also part of Nvidia’s consumer push: the company said its new RTX Spark laptops will tap software including Nemotron 3.5 Transcribe.

Granary v2 and the medical fine-tuning case study

Granary v2’s public manifest spans 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. The Decoder reports a different size, about 844,000 hours in 48 languages, up from roughly 643,000 hours in the previous version. The 463,062-hour figure is the one tied to the public manifest in two reports.

Built with NVIDIA NeMo Curator, the pipeline generates three transcript forms per recording, plus captions, code-switching restoration, speech-grounded question-answer pairs, contextual biasing terms and translations with automatic quality scores. A hallucination check gates transcription, and a high-fidelity filter cut the VoxPopuli corpus from 354,000 hours to 111,000 hours across 22 European languages. Source audio is downloaded separately under each corpus’s original licence. Transcribe itself was trained on approximately 1.72 million unique indexed audio hours across about 372 million examples, plus 5.6 billion unique text tokens.

To show adaptation, Nvidia ran a case study on Eka Care’s open benchmark of 3,619 recordings from Indian doctors and medical students. Naming English and the Latin alphabet in the prompt cut full-benchmark WER from 30.8% to 23.8%, and fine-tuning brought it to 11.8%. SiliconANGLE gives a different prompted figure, 20.8%. On 2,764 held-out clips, fine-tuning reduced WER from 22.44% to 12.17%, a 45.8% relative reduction, putting it ahead of an open medical-specialist model at 13.79%. The run took about 14 GPU-hours on one GB300.

The specialisation had costs: LibriSpeech test-clean WER rose from 1.54% to 1.69%, and the specialist model stayed more accurate on brand-name drugs. Nvidia recommends setting aside at least 100 real, verified recordings for testing custom versions.

What it means for builders, and the limits

For teams building meeting notes, call analytics or clinical dictation, the appeal is one self-hostable model with a commercial-use licence instead of a chain of separate ASR, diarization and summarisation services. Prompted vocabulary and the Granary v2 and NeMo Speech fine-tuning recipes are aimed at domain jargon in each language.

The limits are concrete:

• No streaming. The model card stresses Transcribe is offline; internal chunking does not make it a real-time service.
• Uneven language support. The card grades each of the 40 languages by task, from Leading to Supported, and audio understanding and spoken instruction following were evaluated only in English.
• Verification still needed. Nvidia’s own guidance is to spot-check names, speaker changes and answers against the original recording.
• Self-reported evidence. Leaderboard placements come largely from Nvidia’s announcement and model card, and the Open ASR ranking excludes API models.

What to watch next

The release lands in a busy month for voice AI: Mistral shipped Voxtral 3, ElevenLabs launched Realtime for voice agents, and Salesforce released Launa-Omni. It also extends Nvidia’s open model line after Nemotron 3 Nano Omni in April and Nemotron 3.5 Lightning in August, and its open voice models such as the 1.1-billion-parameter Parakeet.

Three things to watch: confirmation of whether the quantized single-GPU builds and NeMo-Speech.cpp support are fully available; independent tests that check Nvidia’s leaderboard claims, including against API-based services left out of its chart; and Nvidia’s stated next step for the medical study, real recordings of clinicians saying brand names.