Automatic analysis of dialect/language sets
Mahnoosh Mehrabani, John H. L. Hansen · International Journal of Speech Technology · 2015
Dialect variations of a language represent considerable challenges for sustained performance of speech systems. In a given language space, estimation of similarity or diversity between multiple dialects provides valuable information for speech researchers. In the present study, fundamental differences between dialects or closely related languages are explored based on available speech data from those dialects/languages. First, a method is proposed to measure spectral acoustic differences between dialects based on a volume space analysis within a 3D model using log likelihood score distributions derived from traditional Mel Frequency Cepstral Coefficient features and Gaussian Mixture Models. Next, text-independent prosody features based on pitch and energy contour primitives are proposed to study excitation structure differences between dialects. The proposed dialect proximity measures are evaluated and compared on a corpus of Arabic dialects, as well as a corpus of South Indian languages, which are closely related languages. The presented measures are shown to be consistent and repeatable.