New Lantern Releases State-of-the-art Radiology Speech Model

Why Radiology Needs Its Own Speech Model

Radiology dictation is structurally different from the kind of speech that general models are optimized for. Reports are dictated in short, fragmented bursts rather than flowing conversation. They're dense with measurements, anatomical terms, and classification systems (TI-RADS, BI-RADS, LI-RADS, Fleischner) that don't appear in typical training data. And they're full of dates, which even the best general-purpose models handle poorly.

The result is that radiologists spend real time every day cleaning up after their speech engine. It's not a catastrophic failure. It's a constant low-grade tax on attention and speed. As one diagnostician put it: morale sapped by a thousand tiny cuts.

What's Different

Our model is trained on radiology-specific language patterns. It understands anatomical terminology natively, handles laterality with significantly higher accuracy, and processes the short-phrase dictation style that's common in structured reporting without the errors that plague general engines.

Because the speech model lives inside New Lantern's unified platform, it also benefits from clinical context. The model knows what type of study you're reading, what template you're working in, and what findings are expected for that exam type. That context makes the difference between a generic transcription and one that's actually useful.

Part of a Bigger Picture

Better speech recognition on its own would be an incremental improvement. What makes this meaningful is how it fits into the broader AI-first-draft workflow. Curie uses your dictated findings as one of several inputs (alongside OCR from tech sheets and viewer interaction signals like measurements and prior comparisons) to generate a complete first draft of the report. The cleaner that speech input is, the better the draft. Fewer corrections at the dictation layer means fewer corrections downstream.

The speech model is available now to all New Lantern users. If you'd like to see the difference for yourself, reach out to our team.