A Neural Network Phoneme Classification Based on Wavelet Features

Summary

This paper presents the use of discrete wavelet transform for feature extraction of phoneme. Instead of using the conventional wavelet coefficients, energy per sample is calculated in different frequency bands and used as features. Training and test samples of the phonemes were obtained from the TIMIT database from the dialect region DR1 and DR2. Features extracted were updated every 8ms to account for the non-stationary property of the speech signal. For the classification of the phonemes two different classifiers were used based on Linear Discriminant Analysis (LDA) and Multi-Layer Perceptron (MLP). The results obtained show high speaker independent recognition rate by both the classifiers. The recognition rates obtained by using MLP classifier were found to be about 3–10% higher than the LDA for different number offeatures.