Genetic Optimization of Cepstrum Filterbank for Phoneme Classification

Some of the most commonly used speech representations, such as mel-frequency cepstral coefficients, incorporate biologically inspired characteristics into artificial systems. Recent advances have been introduced modifying the shape and distribution of the traditional perceptually scaled filterbank, commonly used for feature extraction. Some alternatives to the classic mel scaled filterbank have been proposed, improving the phoneme recognition performance in adverse conditions. In this work we propose an evolutionary strategy as a way to find an optimal filterbank. Filter parameters such as the central and side frequencies are optimized. A hidden Markov model classifier is used for the evaluation of the fitness for each possible solution. Experiments where conducted using a set of phonemes taken from the TIMIT database with different additive noise levels. Classification results show that the method accomplishes the task of finding an optimized filterbank for phoneme recognition.