This paper presents a new method for studying within-speaker variability in fricative spectra. In this method, nonlinear regression is used to predict values of the spectral mean of a 10-ms window of frication noise from its position in the fricative. The resulting measure of model fit, R2, is used as an index of within-speaker variability. This method was applied to a corpus of /s/ and /∫/ tokens spoken by three groups of participants: 3-4-year-old children, 7-8-year-old children, and adults. Results indicated that fricatives were more variable when produced before the stop consonant /p/ than before /t/. No difference in variability was noted between adults and children, although a nonsignificant trend of decreased variability in adults was noted.