Pitch normalization in conversation analysis: a median-based approach

Pitch plays a central role in spoken interaction, yet its acoustic measurement is challenging because speakers differ widely in their voice ranges. Consequently, fundamental frequency (F0) measurements must be normalized relative to speaker-specific reference points. In Conversation Analysis this normalization is commonly based on pitch-range estimates derived from a speaker's lowest and highest observed fundamental frequency values. This study investigates the reliability of this approach using conversational speech from 45 speakers across two languages. Several statistical measures were compared under conditions typical of spontaneous interaction, including skewed pitch distributions, F0 extraction errors, and limited data availability. The results show that extreme values are highly unstable under these conditions, whereas central measures – particularly the median – provide reliable estimates even in relatively small samples. Crucially, the median remains stable in both manually corrected and uncorrected, noisy data, making it suitable for large-scale analyses in which manual inspection is impractical. Based on these findings, a normalization method is proposed in which F0 values are expressed as semitone deviations from the speaker's median pitch. Conceptually, this method represents pitch values as bidirectional departures from the speaker's typical pitch level in perceptually grounded units. In this way, it captures the linguistic markedness of both high and low pitch extremes. A case-study analysis further demonstrates that this approach captures subtle inter-speaker pitch relations while remaining consistent with the analytic commitments of the field. Taken together, these findings provide a statistically grounded method for pitch normalization that is compatible with both large-scale analyses and fine-grained studies of talk-in-interaction.