How NVIDIA’s Open Transcription Model Could Change Voice AI

Image Source – NVIDIA.com

A transcript tells you what people said. It doesn’t tell you who said it, which language they switched into halfway through, or what the team actually decided. Most speech tools stop at the first part, and that gap is where NVIDIA’s new Nemotron 3.5 Transcribe aims to help.

NVIDIA released the open-weight speech model on October 28, 2026. It handles transcription, speaker attribution, translation, and audio question-answering in one model. NVIDIA also released Granary v2, an open speech dataset with a curation pipeline for adapting speech AI to specific jobs.

Speech recognition is clearly moving past plain transcripts. The goal now is to organise what people say so it’s useful afterwards.

What is NVIDIA Nemotron 3.5 Transcribe?

Nemotron 3.5 Transcribe is an open-weight AI model from NVIDIA that turns recorded audio into several kinds of text output, including transcripts, speaker-labelled records, translations, and answers to questions about the recording.

Developers choose the output with plain-language instructions. One model can transcribe a meeting, label who spoke, translate the result into English, or answer a question like “What deadline did the team agree on?”

That removes some of the usual assembly work. Teams no longer need to stitch together separate models for each task.

What Can Nemotron 3.5 Transcribe Do?

NVIDIA lists the following capabilities for the model:

Language support varies by task. A language that works for transcription may not work for speaker attribution. Check the exact language and feature you need before you plan around it.

How Does an Open-Weight Speech Model Work?

Automatic speech recognition (ASR) is software that converts spoken language into written text. An open-weight model is one whose trained parameters are published, so developers can download, run, and adapt it under the licence terms.

Nemotron 3.5 Transcribe uses a mixture-of-experts architecture. In plain terms, the model holds many specialised sub-networks and uses only some of them at each step. It has roughly 30 billion parameters in total, and about 4 billion are active per step. NVIDIA says it can run on a single NVIDIA GPU, from compact systems up to data-centre hardware.

Open weights also matter for privacy. Some organisations don’t want sensitive recordings leaving their own environment. With a self-hosted model, they may be able to keep audio in-house.

Open weights don’t make a system private, secure, or cheap, though. You still have to check the licence, size the hardware, and cover maintenance and data handling. Even so, you get more control than a closed service gives you.

What is Granary v2, and Why Does It Matter?

Granary v2 is NVIDIA’s open multilingual speech-data resource, paired with a reusable pipeline that prepares audio and matching text for training and fine-tuning speech models.

Training data shapes how well a model performs. Accents, background noise, overlapping voices, and specialist vocabulary all affect accuracy. A model that handles casual conversation well can still stumble on medical dictation or legal proceedings.

Granary v2 also supports different transcription conventions. Take the spoken line: “We paid, uh, twenty-five euros.”

• Verbatim: keeps the “uh.”
• Clean: drops the filler word.
• Normalised: writes the amount as “€25.”

No single version is right for everyone. A researcher studying speech patterns wants the verbatim text, while a team filing meeting notes probably wants the clean one. Granary v2 lets developers train toward the style they need instead of accepting the model’s default.

What are the Business Uses for Nemotron 3.5 Transcribe?

The best-fit uses are the ones where spoken information has to be searched, translated, or summarised after the fact: meetings, healthcare dictation, multilingual support, and interviews. How well it performs depends on the language, terminology, and recording quality in each setting.

Meeting Transcription

Companies produce a lot of spoken information through meetings, sales calls, and interviews. Transcription makes it searchable. Speaker labels show who said what, and audio question-answering helps people find decisions and deadlines quickly.

Verify any answer that feeds into a contract, a financial decision, or an employee’s responsibilities. A fluent AI summary can still drop context or misread a statement.

Healthcare Documentation

Clinical speech is full of drug names, abbreviations, and accents that general-purpose tools often miss. NVIDIA’s published medical case study shows what fine-tuning can do. On the Eka Care Indian medical speech benchmark, NVIDIA reports that adaptation cut word error rate from 22.44% to 12.17% on held-out speakers. That is a 45.8% relative reduction.

Treat that figure as a result from one evaluation setup. It doesn’t guarantee similar gains in your clinic. Dictated notes still need clinician review, and organisations must handle patient privacy, security, and regulatory duties.

Multilingual Customer Support

Support teams can use transcription and translation to review calls, spot recurring complaints, and share calls with colleagues who don’t speak the customer’s language. Test your actual language pairs, accents, and call conditions before launch.

Education, Research, and Journalism

Speaker labels and word-level timestamps tie text back to specific moments in a recording. That helps with interviews, lectures, and research discussions.

Keep the original audio, though. Check names, quotations, and ambiguous statements against it before you publish or cite anything.

How Accurate is Nemotron 3.5 Transcribe?

Accuracy in speech recognition is usually measured by word error rate (WER), which counts substitutions, deletions, and insertions against a reference transcript. A lower WER means fewer mistakes on the tested material.

NVIDIA reports that Nemotron 3.5 Transcribe ranks first among open-weight models on the English Hugging Face Open ASR Leaderboard, with an average WER of 5.33%. The company also reports strong results on multilingual and long-form benchmarks.

Benchmarks only go so far. A model that scores well on a clean test set can behave differently on noisy phone calls, regional accents, crosstalk, unusual names, or technical jargon.

The better question is how it performs on your audio. To find out:

1. Collect representative recordings from your real use case.
2. Create human-verified reference transcripts.
3. Measure WER and review the errors that matter most to you.

Informal meeting notes can tolerate an occasional slip. Medical and legal records can’t.

What Should Developers Check Before Deploying an Open-Weight Speech Model?

Before deploying, check five things: hardware and running costs, privacy and security, language and domain coverage, human verification, and customisation.

• Hardware and operating costs: The model can run on a single supported GPU, but real performance depends on your hardware, precision, workload, and setup. Don’t assume NVIDIA’s reported speeds will carry over to your system.
• Privacy and security: Local deployment keeps recordings under your control. You still need access controls, encryption, and retention policies.
• Language and domain coverage: Test the exact languages, accents, and terms you need. Support for transcription in a language doesn’t mean every other feature works in it.
• Human verification: Review names, numbers, speaker changes, and generated answers wherever an error would cause real harm.
• Customisation: Fine-tuning can sharpen recognition for a specialist field. It needs good training data and careful evaluation, and gains in one domain can reduce performance in another.

Where is Voice AI Heading?

Voice AI is shifting from transcribing speech to understanding and organising it. Nemotron 3.5 Transcribe fits that pattern by pairing recognition with translation, speaker labels, and question-answering.

Benchmark scores won’t decide what happens next. Businesses need models that work on their own data, cover their customers’ languages, fit existing software, and stay within budget and security limits. Open weights give developers more control over deployment. Granary v2 shows how much training data and transcript style matter for specialist work.

Limits remain. Transcribing a recording is not the same as handling a live conversation. An accurate transcript also doesn’t guarantee a correct reading of what someone meant. Human review stays important wherever a mistake could affect people or business decisions.

The real value of voice AI lies in making spoken information easy to find, search, translate, and use. How reliably it does that in real working conditions will decide how much it matters.