The study, conducted by Nadav Matalon of the Weizmann Institute of Science and the Hebrew University of Jerusalem, examines how pitch should be normalized in Conversation Analysis, a research tradition devoted to the fine-grained study of talk-in-interaction. In this field, even predominantly qualitative work often reports instrumental acoustic measurements to support analytic observations, and large-scale quantitative studies depend on them heavily. But for those measurements to mean anything, they must be interpreted relative to each speaker’s individual voice range, since absolute fundamental frequency (F0)—the acoustic correlate of pitch—is, as prosodic theory holds, meaningless unless placed in relation to a speaker’s range.

Two normalization approaches dominate current practice. The first, often called baseline-topline normalization, extracts F0 from a speech sample of roughly one minute and takes the lowest and highest observed values as estimates of the speaker’s pitch-range limits. The second uses only the lower bound, expressing pitch values as deviations from a speaker-specific baseline. Both rest on extreme values, and that is precisely where the trouble begins. Conversational pitch distributions are strongly skewed, extreme pitch events occur only rarely, and voice range can shift with physical state, emotional state, or even the time of day. Minima and maxima are therefore highly sensitive to how much data is available, and pitch-tracking algorithms are prone to errors that introduce spurious extremes—a problem that becomes acute in large-scale studies where manual correction is impractical.

To test how badly these problems bite, Matalon built a gold-standard dataset from naturally occurring everyday conversations in two languages: English, drawn from the CallHome and CallFriend corpora, and Hebrew, drawn from the HUJI Corpus of Spoken Hebrew. Forty-five speakers were analyzed—25 English speakers (15 female, 10 male) and 20 Hebrew speakers (14 female, 6 male). For each speaker, a representative sample of 3 to 5 minutes of continuous active speech was extracted after removing silent pauses, yielding on average about 26,000 ten-millisecond frames per speaker, of which roughly 14,000, or 55 percent, were voiced. F0 was extracted using Praat’s raw cross-correlation algorithm at 10-millisecond intervals, then manually inspected and corrected frame by frame, with particular care taken to verify that genuine extreme pitch events were retained and that values during bursts of laughter were excluded.

With this verified dataset in hand, the study ran three complementary stress tests. The first, incremental trimming, systematically removed between 1 and 5 percent of each speaker’s F0 distribution from either the upper or lower tail, simulating the chance absence of rare extreme events in a sample, and measured how far each summary statistic drifted from its gold-standard value in semitones. The second compared statistics computed from the cleaned data with those from a completely uncorrected extraction using a fixed pitch floor and ceiling of 75 to 600 Hz, which includes typical artifacts such as octave errors. The third, block-based subsampling, repeatedly drew progressively smaller portions of each speaker’s data—contiguous blocks of 20 voiced frames, approximating syllable-length windows—rather than random individual frames, to mimic realistic sampling conditions, repeating the procedure 200 times per speaker and sampling level.

The distributional analysis itself revealed why extremes are so treacherous. Speaker voice ranges spanned 18 to 36 semitones, roughly one and a half to three octaves, with female speakers averaging a median F0 of 217 Hz and male speakers 117 Hz, closely matching previously reported values for English speakers. Yet the average interquartile range was only 4.2 semitones, meaning half of all pitch values fell within a narrow band of the total range. Highlighting the outer 1 percent of values in each tail showed that much of the attested range is realized only rarely. Extreme pitch values, in other words, are both infrequent and widely dispersed—and, arguably, it is precisely their rarity that makes them interactionally meaningful, contributing to their linguistic markedness as deviations from what is expected.

The stress tests delivered a clear verdict. Under trimming, extreme-based measures showed high sensitivity, with near-linear displacement as trim levels increased: at just the 1 percent level, maxima shifted by nearly 6 semitones from the gold standard and minima by just over 2, and the divergence grew steadily with further trimming. Central-tendency measures, by contrast, stayed under 1 semitone of deviation even at 5 percent trimming, with the mean somewhat more sensitive than the median to upper-tail trimming because of high outliers. Under extraction noise, the mean shifted by an average of 0.4 semitones while the median shifted only 0.1, a difference confirmed as statistically significant by a Wilcoxon signed-rank test (p < 0.001). And under subsampling, the mean and median converged rapidly toward their gold-standard values even with only 1 to 2 percent of the data, while the maximum deviated by several semitones with 10 percent of the data and still by 1 semitone with 25 percent—roughly one minute of active speech.

Based on these findings, the study proposes expressing F0 values as semitone deviations from the speaker’s median pitch. The approach has two conceptual advantages beyond its statistical reliability. First, it uses the semitone scale, which best reflects human perception of spoken pitch and is already standard in the field. Second, unlike deviation from a speaker’s lower pitch boundary, deviation from the median represents pitch movements, jumps, and levels as bidirectional departures from a statistical approximation of the speaker’s habitual pitch—capturing the markedness of both high and low extremes rather than treating one end of the range as the natural zero point.

A case study illustrates the method in action. In an excerpt from an everyday English conversation between a university student, Beth, and her father, Dan, discussing whether Beth’s former roommate Rachel might return for graduate school, the two exchange a series of four escalating rise-fall pitch movements: Dan’s assessment wow, Beth’s upgraded yeah, Dan’s high-rise so-prefaced inference-checking question, and Beth’s second upgraded yeah. When the peaks of these four movements are expressed in semitones above each speaker’s median—164.3 Hz translating to 4 semitones for Dan’s wow, 367.3 Hz to 9.4 for Beth’s first yeah, 234.3 Hz to 10.2 for Dan’s question, and 474.3 Hz to 13.8 for Beth’s second yeah—the inter-speaker pitch relations emerge cleanly: Beth’s responses are prosodically upgraded relative to Dan’s preceding turns, and the second pair of movements is elevated relative to the first. The study is careful to note that this does not prove median-based normalization is superior to a carefully applied baseline-topline method, which remains informative for visualization; rather, it shows the median approach is an economical and robust alternative that yields accurate results in a real research setting.

The broader message extends beyond one methodological fix. When acoustic measurements are used to support interactional claims, the choice of summary statistics is not a neutral technical detail but a consequential decision that shapes how prosodic behavior is represented and interpreted. The median’s stability holds for both manually corrected and uncorrected F0 data, making it suitable for large-scale analyses where manual inspection is impractical, while remaining compatible with the fine-grained sequential analysis that characterizes Conversation Analysis. Limitations remain: the sample is not perfectly balanced across languages or sex, generalization to a broader typological range of languages and interactional settings awaits further work, and it is unknown whether the approach extends to other prosodic dimensions such as loudness and voice quality. There is also no established visualization convention yet for median-based normalization, unlike the intuitive baseline-topline plots. Still, the study reframes a long-running debate: the question is not whether quantification should play a role in Conversation Analysis, but how quantitative methods can be designed to remain consistent with the field’s empirical and analytic commitments.

Subject of Research: Statistical robustness of pitch normalization methods in Conversation Analysis

Article Title: Pitch normalization in conversation analysis: a median-based approach

Article References: Matalon, N. (2026). Pitch normalization in conversation analysis: a median-based approach. Frontiers in Communication, 11, 1833068. https://doi.org/10.3389/fcomm.2026.1833068

Image Credits: AI Generated

DOI: 10.3389/fcomm.2026.1833068

Keywords: pitch normalization, Conversation Analysis, fundamental frequency, prosody, median, semitones, speech corpora, F0 extraction, spontaneous conversation, acoustic measurement, linguistic markedness, voice range