Nvidia Nemotron 3.5 Transcribe Explained: 40-Language Open Speech Model, Granary v2 and the India Medical Test

Image: Wikimedia Commons — studio condenser microphones (illustrative image; Nemotron 3.5 Transcribe is software, and its official visuals are company assets not reused here).

TL;DR: nvidia nemotron 3.5 transcribe explained: Nvidia released Nemotron 3.5 Transcribe on 28 October 2026, a free open-weight model that turns speech in 40 languages into text, labels up to eight speakers, translates 31 languages to and from English, and answers questions about a recording — all in a single checkpoint that runs on one GPU. Alongside it comes Granary v2, a 463,062-hour open speech dataset.

📌 KEY HIGHLIGHTS

• The release: Nvidia launched Nemotron 3.5 Transcribe on 28 October 2026, free and open-weight, with weights under the OpenMDW-1.1 licence — commercial use is permitted.
• Four jobs, one model: transcription, speaker labelling, translation and audio question-answering in a single promptable checkpoint, instead of a chain of separate tools.
• The scale: 40 transcription languages, 31 translation languages to and from English, speaker labels in 13 languages, up to eight speakers, recordings up to an hour long.
• The data: Granary v2, released the same day, is an open speech dataset and curation pipeline covering 463,062 hours across 48 languages.
• The India angle: Nvidia's medical case study used recordings from Indian doctors and cut transcription errors by 45.8% after fine-tuning — a concrete result for Indian English speech.

Nvidia Nemotron 3.5 Transcribe explained: what happened?

Most speech tools work like an assembly line: one step detects speech, another writes it down, another works out who spoke, and yet another summarises or translates. Nvidia's Nemotron 3.5 Transcribe folds all of that into one checkpoint that developers steer with plain-English instructions. Ask it to transcribe a meeting and label each speaker, translate the audio into English, or tell you what decisions the meeting reached — the same recording produces a different output depending on the prompt.

Underneath is a 30-billion-parameter mixture-of-experts language model with only about 4 billion parameters active per token, paired with a speech encoder. That sparse design is the point: the model runs on a single Nvidia GPU, from a Jetson AGX Thor robotics controller and a DGX Spark desktop chip up to Blackwell datacentre cards. Nvidia says one RTX PRO 6000 Blackwell Server Edition GPU can process 100 hours of audio in under three minutes — a company measurement under specific conditions (NVFP4 precision, 30-second clips, high concurrency), not a promise for every setup.

What it can do: one model, four speech jobs

Developers control the output with natural-language prompts: a verbatim transcript or a cleaned-up one, custom vocabulary for product names or medical terms, speaker tags with word-level timestamps, translation, or questions about what was said. The model also accepts a list of jargon or names to spell correctly — the kind of detail that matters in hospitals and boardrooms.

Capability	Nemotron 3.5 Transcribe
Transcription	40 languages, recordings up to 1 hour, word-level timestamps
Translation	31 languages to and from English
Speaker labelling	13 languages, up to 8 speakers per recording
Audio Q&A	Ask questions about a recording in plain English (trained in English)
Custom vocabulary	Prompt-supplied product and medical terms, formatting conventions
Hardware	Runs on a single GPU: Jetson AGX Thor, DGX Spark, Blackwell
Licence	OpenMDW-1.1 — free, commercial use permitted

Image: Wikimedia Commons — a handheld microphone (illustrative image for the speech section; not a Nemotron product photo).

On benchmarks, Nvidia reports first place among open-weight models on Hugging Face's Open ASR English leaderboard at 5.33% average word error rate (API models excluded), the lowest multilingual result at 4.62% mean word error rate across seven languages, and a tie for first among open-weight models on Artificial Analysis at 2.8% — while activating about 4 billion parameters against 24 billion for the model it tied. Treat these as company-reported figures until independent tests land; they are Nvidia's measurements, not outside verdicts.

Why it matters for India readers

The most concrete India result in the announcement is a medical fine-tuning case study. Nvidia tested the model on Eka Care's open benchmark of 3,619 recordings from Indian doctors and medical students. Simply naming English and the Latin alphabet in the prompt cut the error rate from 30.8% to 23.8%; fine-tuning brought it to 11.8%. On 2,764 held-out clips, fine-tuning reduced the word error rate from 22.44% to 12.17% — a 45.8% relative reduction — after roughly 14 GPU-hours of adaptation on a single GB300 chip.

That matters because Indian English — with its accents, medical vocabulary and code-switching — is exactly where generic transcription tools stumble. A free model that hospitals, clinics and meeting-tool builders can adapt on one GPU, with a dataset they can inspect, lowers the barrier for Indian developers building transcription for Indian speakers. The tools target business meetings and medical transcription first, two of voice AI's most established footholds, and Nvidia says developers can use the Granary v2 pipeline and its NeMo Speech tools to teach the model an industry's jargon in each language.

Granary v2 is the other half of the story: its public manifest spans 463,062 hours across 48 languages and 16 source corpora, up from 25 languages in version 1, built with Nvidia's NeMo Curator pipeline. The source audio itself is downloaded separately under each corpus's original licence — the dataset release is the manifests, curation pipeline and quality flags, not a single downloadable audio dump.

Should you care?

✅ Care if you build or buy transcription tools — a free, commercial-use model covering 40 languages with speaker labels and translation in one checkpoint is a serious new baseline. ⏳ Wait if you need live captions: this model works on recorded audio only, not real-time streams, and its fastest serving optimisations are not yet in the open-source release.

Availability, licence and limits

The model is available now across Hugging Face, GitHub via Nvidia NeMo Speech, NGC and Build.Nvidia.com, with hosted endpoints from Baseten, DeepInfra and Simplismart and an interactive demo on Hugging Face Spaces and ModelScope Studios. It arrives in a busy month for voice AI: Mistral shipped Voxtral 3 earlier in October, ElevenLabs launched Realtime for instant speech in voice agents, and Salesforce released Launa-Omni — while Nvidia's own tiny Parakeet model (1.1 billion parameters) already showed how small open speech models can go. It also follows a week of open-model releases on this blog: our Liquid AI LFM3 open-weight models explainer covered small specialist agents for phones and laptops, our Google DeepMind Genie 4 world model explainer covered playable AI worlds, and our Google Gemini 4 Carbon India explainer covered Google's cybersecurity-focused frontier model.

The honest limits: Transcribe is an offline model — its internal audio chunking does not make it a real-time streaming service, so live captioning needs a different tool. Audio question answering was trained in English, and the model card grades each of the 40 languages by task, from Leading to Supported. As with any open-weight release, the company provides the weights; running, tuning and supporting them is the adopter's job, unlike a managed API.

FAQs

What is Nvidia Nemotron 3.5 Transcribe? A free, open-weight speech model released by Nvidia on 28 October 2026 that transcribes audio in 40 languages, labels speakers, translates to and from English, and answers questions about a recording — all in one promptable checkpoint.

How many languages does Nemotron 3.5 Transcribe support? It transcribes 40 languages, translates 31 of them to and from English, and attributes speakers in 13 languages, with up to eight speakers per recording and recordings up to an hour long.

Is Nemotron 3.5 Transcribe free for commercial use? Yes — the weights are released under the OpenMDW-1.1 licence and Nvidia states the model is ready for commercial or non-commercial use, with hosted endpoints also offered by Baseten, DeepInfra and Simplismart.

Does Nemotron 3.5 Transcribe work for live transcription? No — it is built for recorded audio, not livestreams. The model card stresses it is an offline model, so real-time captions need a different tool.

Takeaway: Nvidia's Nemotron 3.5 Transcribe packs transcription, speaker labels, translation and audio Q&A into one free, single-GPU model — and its India medical test shows why open, adaptable speech models matter for Indian speakers. The benchmarks are Nvidia's own, but the open weights and the 463,062-hour Granary v2 dataset are real and available today.

Follow World Tech on TechByCountry for tomorrow's update.

📸 Get tech in the last 24 hours on Instagram — follow @_techinlast24hrs for daily posts (9 AM–9 PM) and reels at 12:30 PM & 8:30 PM.

Sources

TechCrunch — Nvidia releases new open transcription model
Unite.AI — NVIDIA launches open-weight speech model and 463,062-hour audio dataset
SiliconReport — Nvidia releases Nemotron 3.5 Transcribe for speech in 40 languages
AI Parabellum — Nvidia releases open Nemotron 3.5 Transcribe speech model