A Clinical Speech Corpus with Temporally Aligned Sensitive Health Information

ABSTRACT Objectives Due to privacy constraints, the sensitivity of medical speech, and the complexity of speech-level annotation, publicly available datasets for clinical speech de-identification remain scarce. To address this gap, we constructed the SREDH-AI Cup Sensitive Health Information (SHI) speech corpus, a time-aligned clinical speech dataset annotated for 26 SHI categories. Methods We compiled the corpus by integrating two English medical speech resources with Mandarin Chinese utterances collected from Taiwanese television dramas. In addition to using an existing open dataset for automated medical transcription, we used the OpenDeID v2 corpus into spoken-style clinical scripts and generated corresponding audio records with 25 voice contributors. All audio data were manually annotated with millisecond-level, time-aligned SHI segments. The annotation schema covered 26 SHI subcategories. Results The final corpus comprises approximately 20 hours of annotated audio and supports evaluation of both automatic speech recognition (ASR) and SHI recognition. It is divided into training (10 hours, 1,539 files), validation (5 hours, 775 files), and test (5 hours, 710 files) subsets. The corpus contains 7,830 SHI entities. Its language distribution reflects the composition of the selected source materials, with 19.36 hours of English and 0.89 hours of Chinese speech. Inter-annotator agreement, measured using Fleiss’s kappa, reached 0.907. In the SREDH/AICUP competition, the top-performing team achieved a mixed error rate of 0.1147 for ASR and a macro F1-score of 0.7103 for SHI recognition. Discussion The corpus poses several important challenges for clinical speech de-identification. In particular, SHI categories are distributed in a long-tailed manner, and the amount of English and Mandarin Chinese data is highly imbalanced. The Mandarin Chinese subset should be understood as a supplementary multilingual extension rather than as a parallel monolingual benchmark. Our experience constructing this subset highlights both the scarcity of usable Chinese clinical speech resources and the practical difficulty of producing time-aligned SHI annotations for low-resource clinical speech settings. Conclusion The SREDH-AI Cup SHI speech corpus provides a time-aligned speech dataset that approximates dialogue-based clinical scenarios. It supports research on automated medical speech de-identification, ASR, and SHI recognition, while also offering an initial reference resource for future work on English and low-resource clinical speech privacy protection. Key message What is already known on this topic - There is a scarcity of publicly available clinical speech datasets containing time-aligned sensitive health information (SHI) annotations for clinical speech de-identification research. What this study adds - The SREDH-AI Cup SHI speech corpus introduces a clinically-oriented speech dataset with millisecond-level, time-aligned SHI annotations across 26 categories. - The corpus integrates structured annotation protocols and standardized data processing to support 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. - The dataset may support the development of English and low-resource bilingual conditions privacy-preserving technologies in clinical speech environments.