Parameterization of speech signals for robust voice recognition

In this paper, we propose a speech parameterization technique based on a compressive Gammachirp filterbank. This filterbank represents a reliable model of the cochlear auditory filter and provides a good approximation of their spectral and selective behaviour. The recognition performance of our technique is tested on isolated-words extracted from the TIMIT database. The adopted speech recognition system is the HTK.3.4.1 platform based on Hidden Markov Models with Gaussian-Mixture densities. The evaluation results showed that the proposed technique gives better recognition rate compared to conventional techniques: PLP (Perceptual Linear Prediction) and LPCC (Linear Prediction Cepstral Coefficient).