Comparative Studies for Speech Analysis Based on Multiresolution Spectrograms

This paper presents an evaluation study of the Multiresolution Spectral Analysis (MRS) method which provides a higher temporal accuracy in the upper spectral region and a better frequency resolution in the lower spectral range. We showcase the importance of this tool by attempting an automatic transition zone detection and an automatic silence/sonorant/non-sonorant classification. Our approach is compared to existing methods based on the MRS and classical spectral analysis by the means of our Visual Assistance of Speech Processing (VASP) System and two corpora. Our approach appears to yield better results in the two tasks in question than the other methods.