Speechmatics Launches Oak Voice Model

Speechmatics has released Oak, a medical speech recognition model that transcribes clinical conversations as speakers switch languages mid-sentence, keeping drug names, doses, and diagnoses highly accurate.

Instead of word erroer rate, Speechmatics measures Oak on medical term recall specifically, which focuses on the share of correctly transcribed drug names, conditions, procedures, and side effects.

In an internal Speechmatics study, the Oak model accurately transcribed 91 percent of all medical terms in English and 10 additional commonly used languages, tested on both clean audio and recordings with realistic background noise. The eight other providers missed or mistranscribed an average of 25.7 percent of medical terms.

In Spanish, Oak transcribed 96.3 percent of medical terms, against 86.8 percent on average for the other eight providers. For every 100 drug names, conditions, and side effects spoken in a consultation in Spanish, Oak misses about four, while the others miss 13 on average.

Oak transcribes a conversation across languages and speakers, switching mid-sentence with no language set in advance. It supports 55 languages and can transcribe an hour of audio accurately within 18 seconds, allowing for more optimized workflows.

Oak can be deployed in the cloud or self-hosted within healthcare organizations' own environments. For self-hosted deployments, Speechmatics uses NVIDIA Dynamo-Triton to serve Oak. Speechmatics also uses NVIDIA Nsight Systems for performance profiling and optimization, alongside quantized NVIDIA CUDA kernels to improve inference efficiency.

"Our mission is to understand every voice and the risk of getting that wrong in a medical environment is huge. A consultation can move from Spanish to English and back in one sentence, with a drug name in the middle. That switch is the precise moment a clinician can least afford an error," said Katy Wigdahl, CEO of Speechmatics, in a statement. "Getting clinical terms right in every language is one of the hardest problems in speech recognition. Oak shows it can be solved when the model is built for how clinicians and patients actually speak, and we're proud to see it set a new bar for accuracy where it matters most."