Das Transkript im Zeitalter der Künstlichen Intelligenz Automatische Spracherkennung in der Oral History

Abstract

The contribution examines the potential and limitations of automatic speech recognition (ASR) for oral history interviews. Manual transcription of audiovisual recordings is time-consuming and costly; this “transcription bottleneck” hinders the indexing of extensive collections. The “ASR4Memory” project developed a data protection-compliant open-source transcription pipeline based on WhisperX, which is operated on local servers. A key challenge is the smoothing of spoken language caused by ASR models, which results in the loss of dialects, filler words, and nonverbal expressions – elements that are often essential for scientific analysis. Its use also raises ethical and legal questions, particularly with regard to data protection and potential distortions caused by training data. The ASR application generates various time-aligned transcript formats with speaker recognition and supports multimodal work and full-text search. Despite considerable time savings, manual post-processing is often still necessary. Domain-specific fine-tuning with 300 hours of anonymized interviews significantly improved both the word error rate and the recognition of historical terms. The study advocates for transparent, critically reflected ASR use that takes into account subject-specific requirements and ethical standards in order to sustainably develop audiovisual research data in accordance with FAIR principles.

Bibliography: Kilgus, Tobias/Kompiel, Peter: Das Transkript im Zeitalter der Künstlichen Intelligenz. Automatische Spracherkennung in der Oral History, BIOS – Zeitschrift für Biographieforschung, Oral History und Lebensverlaufsanalysen, 1+2-2025, pp. 68-82.