Abstract
Objectives: Publicly available datasets for clinical speech de-identification remain scarce due to privacy constraints and annotation complexity. We address this gap with the 2025 SREDH/AI Cup SHI Speech Corpus, a time-aligned clinical speech dataset for detecting sensitive health information (SHI). Methods: The dataset integrates three sources: the OpenDeID-v2 corpus, the dataset for automated medical transcription and medical dialogues derived from Taiwanese television dramas, combining English medical speech with Mandarin Chinese utterances. All audio was manually annotated with millisecond-level, time-aligned SHI segments using a schema of eight primary categories and 38 subcategories of which 26 were attested. Results: The corpus comprises approximately 20 hours of annotated audio (19.36 hours English, 0.89 hours Chinese) and 7830 SHI entities, split into training (10 hours, 1539 files), validation (5 hours, 775 files) and test (5 hours, 710 files) sets, supporting evaluation of both automatic speech recognition (ASR) and SHI recognition. Interannotator agreement (Fleiss’s kappa) reached 0.907. In the 2025 SREDH/AI Cup competition, top-performing systems achieved a mixed error rate of 0.1147 in ASR and a macro F1 of 0.7103 in SHI recognition. Discussion: SHI category distribution is long-tailed, and English–Chinese data volume is highly imbalanced, underscoring the scarcity of Chinese clinical speech resources and the difficulty of time-aligned annotation in low-resource settings. Conclusion: This corpus approximates dialogue-based clinical scenarios and supports research on medical speech de-identification, offering a unique reference for low-resource clinical speech privacy protection. Key message: What is already known on this topic: - Publicly available clinical speech datasets containing time-aligned sensitive health information (SHI) annotations for clinical speech de-identification research are limited. What this study adds: - The 2025 SREDH AI Cup SHI Speech Corpus introduces a clinically oriented speech dataset with millisecond-level time-aligned SHI annotations covering 26 of the 38 SHI subcategories defined in the annotation schema. - The corpus integrates structured annotation protocols and standardized data processing to support the reproducible benchmarking of speech-based de-identification models. How this study might affect research, practice or policy: - The availability of time-aligned SHI annotations may facilitate research on real-time or streaming de-identification systems beyond conventional, transcription-focused approaches. - This dataset may support the development of privacy-preserving technologies for clinical speech in bilingual settings involving English and low-resource languages.
Funding Information
  • Ministry of Education, ASUSTeK Computer Inc and the National Science and Technology Council