NL-Vox: How New Lantern Built a Reporter with State-of-the-Art Dictation by Listening to Radiologists

Most dictation software still forces radiologists to work around its errors. As PowerScribe 360 nears its end of life, practices have an opportunity to replace it with a reporter designed around the radiologist. New Lantern built its reporter and its state-of-the-art speech model, NL-Vox, to be exactly that, by listening to radiologists.

Where it started

Radiologists know the errors of legacy dictation all too well. “Swan-Ganz catheter” becomes “swollen glands catheter,” “atelectasis” becomes “anorexia,” and “pons” becomes “pounds.” Over time, the frustration turns into resignation, and radiologists come to accept the workarounds and lost time as part of the job.

New Lantern’s reporter originally ran on a leading third-party speech model, and our radiologists faced the same issues. Radiologists coming from PowerScribe and Fluency found the experience comparable, including the inefficient workarounds they had been forced to resort to.

For radiologists who have to change how they speak to accommodate their software, a reporter performing on par with existing systems is not good enough.

“Watching radiologists dictate, we kept seeing clever but painful workarounds, like slowing down or pronouncing words unnaturally just so the software would catch them.

“Rather than accept this status quo, we built our own speech model, NL-Vox, around the way radiologists actually want to dictate.”

Christian Bundy, Software Engineer, New Lantern

Radiologists make it better every day

“NL-Vox is the most accurate speech-to-text model I have ever used. It knows the clinical context of what I’m saying, so it rarely mixes up words that sound alike, and I can dictate without pausing and trust that it all got picked up.”

Samir Mehta, MD, Senior Medical Director, New Lantern

That accuracy comes from the radiologists who use it. When NL-Vox misses a word, phrase, or command, the radiologist flags it. Those flags, together with the real-world dictation in every signed report, become targeted training that improves the model continuously. As radiologist trust grows, they stop working around the software and dictate the way they naturally speak, which makes the model even more accurate. Each new site adds new voices, and every radiologist on the platform benefits.

“When it misses something, I flag it right in the report and keep reading. Medial and mesial used to trip it up, and now I don’t think about it. Every radiologist flagging a word is making it better for the rest of us.”

Samir Mehta, MD, Senior Medical Director, New Lantern

This also holds up on public data, and the results are striking. On Eye Gaze, a public dataset of about 1,000 chest X-ray dictations, NL-Vox made 36% fewer errors than Google’s MedASR and 51% fewer than Google Medical Dictation, models in the same class as the one radiologists once found comparable to PowerScribe and Fluency. With NL-Vox, radiologists no longer have to adapt to their software.

Source: New Lantern evaluation on Eye Gaze, a public PhysioNet dataset.

Built into the reporter

Beyond dictation, New Lantern’s reporter takes over much of the repetitive reporting work. It learns how each radiologist prefers their reports written. Radiologists save their macros once and call them up by voice. When they dictate only the positive findings, Curie, New Lantern’s AI Resident, pre-drafts the rest of the report, including the impression. The radiologist still reviews and signs every report.

“Legacy systems made me dictate everything: every normal finding, every measurement and image number, all in a forced order. With New Lantern, I dictate the way I read, calling out findings as I see them, and it does the rest. It organizes everything into the report, and measurements and image numbers go in with a click. I haven’t seen that in any other system.”

Varun Chowdhary, MD, MBA, Radiologist at Radiology Associates at Main Street

Practices replacing PowerScribe 360 can request a demo to see a reporter that learns from every radiologist.