Abstract:

Speech rhythm in terms of durational variability of different levels of phonetic inter vals can vary between speakers. The present article examines the role of syllabic intensity characteristics in rhythmic variability. Mean and peak intensity vari ability across syllables (stdevM, varcoM, stdevP, varcoP, rPVIm, nPVIm, rPVIp, nPVIp; henceforth: intensity measures) were investigated as a function of speaker in a database where within-speaker variability was strong (BonnTempo) and another database designed to examine between-speaker rhythmic variability (TEVOID). It was found that the intensity measures varied significantly between speakers in both databases. Semiautomatic speaker recognition based on duration measures (%V, ∆V(ln), ∆C(ln), ∆Peak(ln), ∆Syll(ln) and nPVISyll) and intensity measures using multinomial logistic regression and feedforward neural networks was carried out for the two databases. Results showed that intensity measures contained stronger speaker specific information compared to measures based on durational variability of phonetic intervals. In addition, effects of the recognition algorithms (speaker recognition using multinomial logistic regression was significantly better than neural networks for BonnTempo) and data normalisation procedures (z-score normalised data was significantly better than non-normalised data in TEVOID) were discovered. This means that syllable intensity characteristics play an important role in between-speaker rhythmic differences and possibly in speech rhythm variability in general.

Author: Lei He, Volker Dellwo