Estimating the Acoustic Variation of s via Principal Component Analysis

Abstract
Articulatory and acoustic variation has recently become one of the most prominent spheres in phonetics. The acoustics of fricative sounds, sibilants in particular, however, is known to be very difficult to study. As far as we are concerned, no algorithm has been created yet to estimate the degree of acoustic variation of fricatives, neither between the speakers of one language, nor between languages. In this article, we try to estimate the degree of acoustic variability of the alveolar sibilant s. We were interested in creating and evaluating algorithm for estimation interlanguage variability and make data from different languages comparable. We also were interested in estimation correlation between the phonological complexity of sibilant subsystem and variability of s sound. We analyze s in similar contexts pronounced by several female speakers of five unrelated languages of Russia: Adyghe, Nanai, Russian, Udmurt and Chukchi. All pronunciations were manually annotated, and then spectral information were automatically extracted and transformed via Linear Predicting Coding. The obtained spectral slices were analyzed and ten different features were extracted: frequency of the first peak (in Hertz and Bark), slope of the linear regression based on values from global minimum before peak (annotated manually) to peak itself, center of gravity (in Hertz and Bark), standard deviation (in Hertz and Bark), skewness and kurtosis. Since it is hard to analyze all these features separately, we used Principal Component Analysis transformation for reducing number of variables. Cumulative percentage of the dispersion explained by first and second Principal Components is equal to 65 %. At the end we show how it is possible to use obtained Principal Components for measuring variability and comparing different utterance of alveolar sibilants. As a result we achieve some goals we planned: 1) we developed the algorithm for variability analysis that could be used in any other field of acoustics; 2) our analysis shows that some speakers could be more variable then the whole languages; 3) the analysis of our data using this algorithm shows that Nanai and Chukchi is more variable comparing to other variables. This also corresponds to the least complex sibilant subsystems.

Keywords
acoustic variation, sibilants, Principal Component Analysis, phonological typology, consonants