Modeling segmentation precision and inter-segmenter variability.

Abstract

Inter-segmenter differences have been reported in the literature to affect the measurements of segment durations [Allen (1975)] and more specifically rhythm metrics [Mairano and Romano (2007)]. Segmenters may differ in criteria used for segmentation as well as resolution of uncertainty due to gradual change between segments (such as between nasalized vowel and coda nasal). In addition, small variations in the location of segmentation border (one period of the waveform, or 5–10 ms for adult speakers) are likely both between two segmenters and between different segmentation sessions of the same segmenter. This study investigates the effect of segmentation errors on the classification power of rhythm metrics. Errors are modeled to represent (1) small random variation and (2) resolution of larger interval uncertainties (20–30 ms). Simulations using the two scenarios were performed on samples of manually segmented speech. Results show that errors in scenario-1 do not affect the separation of languages in (%V, stdC), (rPVI-C, nPVI-V), and (%V, VarcoV) space, which implies that precision of less than 5 ms is not required in studies that use these measures. Errors in scenario-2 may be large enough to shift a language closer to the group it does not belong, depending on the frequency of such errors.