Giving Radiologists Their Voice Back — in 40 language-locales — with NVIDIA Nemotron Speech

I became a radiologist to read images and care for patients. I did not become one to spend my days narrating findings in a stilted, robotic cadence into a legacy dictation system — pausing to insert punctuation, fighting template fields, and bending my own language to fit what an aging speech engine could understand. That work is mind-numbing. It pulls attention away from the two actually important things: the patient and the images in front of you.

RADPAIR was born out of that frustration. I wanted to take back control. I used generative AI to offload the mechanical burden of reporting so that I could simply speak naturally, the way a physician thinks and talks, and let the system do the structuring, formatting, and heavy lifting. Instead of dictating to a machine, I could focus on the case and let the technology meet me where I was. That single shift, from serving the software to having the software serve me, is the entire idea behind RADPAIR.

From one frustrated radiologist to a global platform

Since we started in 2023, RADPAIR has grown far faster than I imagined. We’ve captured meaningful market share across the United States, and we are now expanding internationally across the EU, the UK, Australia, the Middle East, and beyond. Radiologists everywhere are wrestling with the same problem I was: too much administrative friction, not enough time for the work that requires their training. The response to a more natural way of reporting has been remarkable.

But international growth surfaced a challenge that we couldn’t engineer our way around with English alone. Radiology is a global profession, and our customers don’t all practice in English. To serve them well, we needed speech recognition that was highly accurate, low-latency, and genuinely multilingual at the level of clinical trust that a radiologist will stake a report on. Anything less reintroduces exactly the friction we set out to eliminate.

The English foundation: PAIR 3.0

Our newest speech engine, PAIR 3.0, performs exceptionally well in English. It captures natural, conversational dictation with the accuracy and responsiveness radiologists need, and it has become the backbone of the RADPAIR experience for our U.S. and English-speaking users. The question we kept coming back to was deceptively simple: how do we deliver that same quality in German, French, Spanish, and every other language our growing customer base speaks?

We didn’t want to compromise. We wanted German-speaking radiologists to feel the same “this just works” moment that our English-speaking users feel — accurate transcription of dense medical terminology, low enough latency to keep up with natural speech, and the freedom to dictate the way they actually think.

Validating the path with German and NVIDIA Nemotron Speech

Our first target language was German, and this is where the NVIDIA Nemotron Speech model family changed the equation for us.

We successfully fine-tuned the brand-new multilingual Nemotron 3.5 ASR to German for the realities of radiology — the specialized vocabulary, the phrasing, the way findings are actually spoken in a reading room. We then put it in front of German-speaking radiologists to validate it against real clinical use, not just synthetic benchmarks. Their verdict was clear: the model is very good. It held up to the standard we hold ourselves to in English, which is exactly the bar we needed to clear before bringing it anywhere near a patient’s report.

That validation is important because it does more than solve German. It proves out a repeatable methodology. We now have a clear path toward production-level deployment of a multilingual speech engine, and the same approach, the same fine-tuning and validation workflow, can be applied across the rest of the languages we support: French, Spanish, Portuguese, Swedish, and others still to come. What was an open-ended research problem is now a process we can run, language by language, with confidence.

Built in close collaboration with NVIDIA

I want to be direct about how much working with NVIDIA shaped this outcome. We worked closely with the Nemotron Speech team and others across NVIDIA to train this model the right way. They gave us guidance at every step — on the model, on the fine-tuning approach, on getting the most out of the technology — and they have been genuinely, consistently helpful throughout. This wasn’t a matter of downloading a model and hoping for the best. It was a real collaboration, and it accelerated us dramatically.

What comes next

For me, this is about returning radiologists everywhere to the work they trained for. Every minute a physician spends wrestling with a reporting system is a minute taken away from the patient and the images. Multilingual support isn’t a feature on a roadmap for us; it’s how we extend that mission to a global community of radiologists who deserve the same natural, human way of working, no matter what language they practice in.

German was the beginning With NVIDIA Nemotron Speech, the rest of the world’s languages are now a path we know how to walk.

RADPAIR is an AI-native radiology reporting platform that lets radiologists dictate naturally and focus on patients and images instead of paperwork. To learn more, visit radpair.com