Emotion Through Breath in the ACT UP Oral History Project

This paper explores how computational analysis of paralinguistic cues, particularly breathing, can enrich the study of emotionality in Oral History Archives (OHA). While oral historians recognise the loss of non-verbal information in transcript-based research, large-scale audio analysis remains underdeveloped. Using the Voices of ACT UP (VO-ACTUP) dataset, which contains interviews with ACT UP activists about their experiences during the AIDS epidemic, we investigate how non-verbal audio features capture emotional nuances that transcripts cannot convey. To support this analysis, we implement and extend state-of-the-art speech-based breathing prediction models using the UCL-SBM dataset and propose two new WavLM-based architectures that improve current performance. We then assess robustness by aligning predicted breathing signals with annotated breathing events in VO-ACTUP. Finally, we examine how breathing-derived features, such as breathing rate, correlate with emotional expression across the corpus. Our work demonstrates the potential of breathing analysis to reveal affective dimensions of oral history at scale.