The Impact of Audio Quality Filtering on Demographic and Expressive Diversity in Dyadic Conversational Data

Abstract

Audio quality filtering is a standard preprocessing step in affective speech corpus curation, yet while downstream demographic bias in speech models is well-documented, the representational consequences of filtering strategies are rarely examined. We analyse DaiKon, a multilingual dual-channel dyadic conversation corpus, examining how six quality metrics—signal-to-noise ratio (SNR), voice activity detection (VAD), and four Meta AudioBox dimensions—interact with speaker demographics, emotional expressivity, and personality across 1,202 participants (605 rooms, 478 hours). Applying progressively stricter thresholds, we find minority language groups are disproportionately removed under SNR filtering, while non-English speakers, older participants, and female speakers are disproportionately removed across perceptual metrics; younger participants’ lower speech activity (VAD 0.754 for 18–24 vs. 0.823 for 45–54) further increases their vulnerability to VAD-based filtering. Higher perceptual quality correlates negatively with emotional expressivity (AES Content Enjoyment: Cohen’s d = −1.19 for confusion) and with neuroticism scores. These findings demonstrate that quality filtering is not demographically neutral: it operates as a diversity tax on who and what emotional content remains for training.