Build Your Own AI Trust Center

Demonstrate your commitment to AI safety and transparency with a customized trust center

Banner

MedScribe AI

A clinician-led AI company pioneering ambient clinical documentation, processing hundreds of thousands of clinical encounters weekly across dozens of health systems.

Word Error Rate

Updated 5 min ago via Autoblocks

Overview

Test Description

Measures transcription accuracy by calculating minimum edits needed to match reference text.

Key Metrics

Overall WER
13.7%
Medical Term Accuracy
96.6%

Evaluation

Methodology

Compare AI-generated transcripts against human-written gold standard reference transcripts. Calculate Word Error Rate (WER) as: (number of edits) / (length of reference transcript). Validated through blinded head-to-head trials.

Interpretation of scores

Lower WER indicates better performance. Our model shows a 16% relative reduction in word error rate compared to other medical ASR systems, with a 45% relative reduction in error on new medications.

Dataset

Dataset Description

Over 10,000 hours of clinical conversations with associated audio, gold standard reference transcripts, and rich metadata on patient characteristics. Includes the public Librispeech benchmark dataset and internal medical conversation datasets.

Dimensions

Language

English13.7%
Spanish3.6%
Vietnamese2.5%
Tagalog3.1%

Clinical Setting

Primary Care13.9%
Specialty Care13.5%