NVIDIA has released Nemotron 3.5 Transcribe, a free, open-weight model that turns recorded speech into text in up to 40 languages, alongside Granary v2, its open speech dataset. Announced on October 28, 2026, Transcribe combines transcription, translation, speaker attribution and spoken question answering in a single checkpoint, handling recordings up to an hour long and telling up to eight speakers apart. It runs on a single chip, from edge devices to data-centre hardware, and it arrives in the middle of a crowded month for voice AI.
One Checkpoint For Four Speech Jobs
Transcribe is a 30-billion-parameter mixture-of-experts model that activates about 4 billion parameters per token, pairing a speech encoder with a hybrid Mamba-Transformer language model. That sparse design lets it run on one GPU, from the Jetson AGX Thor robotics controller and DGX Spark laptops to Blackwell data-centre chips. In company benchmarks using NVFP4 precision, an RTX PRO 6000 Blackwell Server Edition GPU processed 100 hours of audio in under three minutes.
The practical pitch is consolidation. A developer building meeting tools or medical transcription no longer needs separate models for turning speech into text, translating it, labelling who said what, and answering questions about the recording. One prompt can produce a verbatim transcript, a clean summary, or answers about a recorded meeting, and prompts can include product terms and formatting conventions so specialised vocabulary in fields like healthcare and finance is handled correctly.
Language support differs by task, and the details matter. Transcribe transcribes 40 languages, translates 31 languages to and from English, and attributes speakers in 13 languages. Audio question answering was trained only in English, the model card warns, so teams planning multilingual question answering should test that gap before committing.
Benchmarks, With The Caveats Attached
NVIDIA claims Transcribe beats rivals on benchmarks, including Hugging Face’s Open ASR Leaderboard. Independent measurement adds useful context: Artificial Analysis puts Transcribe at 2.8 percent on its AA-WER index, tying Mistral’s Voxtral Small for first place among open-weight models, while activating about 4 billion parameters against 24 billion for the model it ties. That efficiency gap is the real headline for builders watching serving costs.
Speaker attribution is where the release looks strongest. Transcribe scored an 18.26 percent diarization error rate on David AI’s multilingual track, with NVIDIA’s own dedicated Nemotron 3 Diarization model placing second at 18.38 percent. VoiceArena’s English benchmark also ranked Transcribe first, at 32.4 percent cpWER, a word error score counted after matching speaker labels. As with all launch benchmarks, verify these on your own recordings, since names, accents and overlapping speech are where transcription usually breaks.
Granary v2: The Dataset Behind The Model
Alongside the model, NVIDIA released Granary v2, version 2 of its open speech dataset, whose public manifest covers 463,062 hours across 48 languages, up from 25 languages in version 1. The dataset and its curation pipeline are open, which matters for teams that want to fine-tune Transcribe on their own vocabulary rather than accept a generic model. NVIDIA says developers can use the model, the updated dataset and tools like its NeMo Speech model to customise apps to recognise an industry’s jargon in each language.
That openness is also a strategic move. By shipping weights, data and tooling together, NVIDIA makes its speech stack a platform other builders can extend, rather than a single model they merely download. Hosted endpoints are offered by Baseten, DeepInfra and Simplismart, with an interactive demo on Hugging Face Spaces, and the weights are released under the OpenMDW-1.1 licence for commercial or non-commercial use.
The One Big Limitation
Transcribe works with recorded audio, not livestreams. That single constraint draws the line between this release and the realtime voice agents dominating the rest of the week’s news. A team building a live call assistant needs a streaming model, and Transcribe is not that model. What it offers instead is depth on recordings: longer files, more speakers, and question answering over audio that has already been captured, which fits meetings, interviews, dictation and archived calls.
The timing is deliberate. Mistral shipped Voxtral 3 earlier in October, ElevenLabs launched its Realtime models for instant speech in voice agents, and Salesforce released Launa-Omni, its voice model. Transcribe joins that wave from a different angle: not the fastest live voice, but the most complete open tool for recorded speech, in the same family as the Nemotron 3.5 models NVIDIA shipped for always-on agents.
Why It Matters
Transcribe matters because recorded speech is still where most workplace audio lives. Meetings, hearings and interviews produce hours of audio that someone must turn into reliable text, and a free open model that does that in 40 languages, on one GPU, lowers the cost of building those tools dramatically. The risk is over-trusting it: speaker labels and translated quotes still need a human check before they enter a record, a chart or a court filing, and the livestream gap means live products need a different model entirely.
Background
NVIDIA’s open-weight voice line has grown quickly, from the small Parakeet model at 1.1 billion parameters to the Nemotron 3.5 family spanning server, laptop and edge tiers. Transcribe extends that line into speech understanding, pairing a model builders can run anywhere with a dataset large enough to adapt it, and it lands as voice becomes one of the most contested surfaces in AI.
Source: TechCrunch report
0 Comments