Methods of estimation of the fundamental frequency and the formant frequencies as input parameters in a system of Automatic Speech Recognition

Abstract

[EN] The fundamental frequency (pitch) and the formant frequencies constitute essential parameters in automatic speech recognition systems, as they capture both the excitation source and the resonant characteristics of the vocal tract. In this paper, we propose and analyze methods for the estimation of these parameters from speech signals. The fundamental frequency is estimated in the time domain using a modified short-time autocorrelation technique that incorporates nonlinear cutting functions, designed to reduce the influence of formant structure and enhance periodicity detection. For the estimation of formant frequencies, an alternative spectral approach is introduced, based on the identification of successive maximum peaks in the spectrum while excluding frequency intervals around previously detected formants to avoid spurious detections. The proposed algorithms are implemented efficiently in MATLAB and tested on speech data corresponding to the five Spanish vowels, recorded from multiple speakers. The experimental results show good agreement with standard reference values for vowel formant frequencies, demonstrating the suitability of the proposed methods as reliable input parameters for automatic speech recognition systems.