Prosodic peak estimation under segmental perturbations

Despite the apparent simplicity, measuring the position of peaks in speech fundamental frequency (f(0)) can produce unexpected results in a model where f(0) is the superposition of a supersegmental component and a segmental component. In these models, the measured f(0) peak position can be as much as an entire syllable different from the peak of the intonation component. This difference can be large enough so that the measured peak positions could falsely suggest a phonological distinction in the intonation where none really exists. This paper then discusses measurement techniques that are less sensitive to segmental effects than directly measuring the position of the f(0) maximum. A algorithm, called the "bracketed maximum," is presented. The performance of these techniques is compared on a corpus of speech data where the intonation is expected to be in a stable position. The bracketed maximum can reduce the variance of peak position measurements by at least 15%, in the presence of changing segmental structure, thereby presumably yielding a more accurate measurement of the intonation peak position.