Automatic speech/speaker recognition in noisy environments using wavelet transform

Feature extraction represents a crucial step in pattern recognition in general and in speech/speaker recognition in particular. Robustness to most of the common types of noise is essential. This paper presents a discrete wavelet transform-based feature extraction technique for multi-band automatic speech/speaker recognition. Experimental results have shown that this technique is of comparable performance with a full-band (conventional) technique, under matched conditions (clean speech for both training and testing). It has been found that both techniques are complementary under mismatched conditions (clean speech for training and noisy speech for testing), in that if the features extracted using each of them are combined, better recognition rates are attainable especially at low signal-to-noise ratios.