A comparative study of several approaches to short-term frequency analysis of a speech signal

Abstract

We study how the time-frequency representation of a speech signal depends on the chosen method of frequency analysis. We consider dynamical spectrograms obtained with a set of band-pass filters with different parameters and different order of their position along the frequency axis. We show that when a set of filters with parameters close to the filters of an audial analyzer is used, information on vowels and consonants in the speech signal is more uniformly distributed across the frequency axis, and spectral maxima related to the first and second formants of a vowel are more explicitly expressed, which is very important for speech recognition.