HaessigDB: A Database of Irritable Speech with Intensity Grading

HaessigDB HaessigDB (human-annotated emotional speech snippets with intensity grading) is a corpus of acted, call-center-style English speech for fine-grained irritability detection. Four professional voice actors recorded synthetic banking-support dialogues scripted so the customer grows increasingly irritable as the request goes unresolved. Each snippet carries crowdsourced ordinal intensity ratings (1–10) on three irritability dimensions: annoyance, frustration, and aggression. Unlike categorical resources such as Emo-DB, IEMOCAP, or RAVDESS, the labels capture how intense an emotion is rather than only its presence. Snippets retain their position within each call, which supports trajectory modeling and early-escalation research. Subsets The CSVs represent the different subsets described in the paper. Each row links to an audio clip via the file_name column (see Audio files). Subset Contents all_ratings Full rated set with all three dimensions aggression_high_agreement High-agreement subset (Krippendorff's α > 0.80) for aggression annoyance_high_agreement High-agreement subset for annoyance frustration_high_agreement High-agreement subset for frustration inner_join_high_agreement Snippets meeting α > 0.80 across all three dimensions outer_join_high_agreement Snippets meeting α > 0.80 in at least one dimension Fields Column Description file_name Audio filename, encoding actor, call, and sentence position (e.g. actor1_call_1_sentence_1.wav). The clip itself is at audio/<file_name> actor Voice actor identifier (group on this for speaker-disjoint splits) call Call identifier sentence Snippet position within the call (temporal index) transcript Text spoken in the snippet aggression / frustration / annoyance Mean ordinal intensity rating across annotators (1–10) Citation Weller, N., Grau, M., & Blohm, I. (2026). HaessigDB: A Database of Irritable Speech with Intensity Grading. Interspeech 2026.