Development of a Clinical Speech Corpus with Temporally Aligned Sensitive Health Information Annotations

The 2025 SREDH/AI Cup SHI Speech Corpus provides 19.36 hours of English and 0.89 hours of Mandarin clinical speech with millisecond-aligned SHI annotations and 7830 entities.

• Why it matters: Public clinical speech de-identification datasets are scarce because privacy constraints and time-aligned annotation are complex, especially for low-resource Chinese clinical speech.
• What they did: The corpus combines OpenDeID-v2 with Taiwanese-drama-derived medical transcription and dialogues, manually annotating SHI segments at millisecond level across eight primary categories and 38 subcategories.
• The result: It is split into 10/5/5 hours for training/validation/test with Fleiss’ kappa 0.907, and competition systems reached ASR mixed error rate 0.1147 and SHI macro F1 0.7103.