Researchers built a publicly available corpus of about 20 hours of clinical speech, manually annotated to mark sensitive health information at millisecond precision. It combines English medical speech with some Mandarin dialogue, covering 7,830 labeled entities across 26 privacy categories, so AI systems can learn to find and remove private details.
Voice tools that transcribe doctor visits are becoming common, but recordings often contain names, diagnoses, and other private details that are hard to strip out automatically. Privacy rules have kept most training data locked away, slowing progress. By releasing a shared, carefully labeled dataset, researchers give anyone building medical speech tools a common benchmark for testing de-identification systems.
✦ Shared privacy-labeled data could make voice-recorded medicine safer for patients.