The performance of a pattern recognition technique is usually determined by the ability of extracting useful features from the available data so as to effectively characterize and discriminate between patterns. We describe a novel method for feature extraction from speech signals. For this purpose, we generate spectrograms, which are time-frequency representations of the original signal. We show that, by considering this spectrogram as a textured image, a wavelet transform can be applied to generate useful features for recognizing the speech signal. This method is used for the classification of voice dysphonia. Its performance is compared with another technique taken from the literature. A recognition accuracy of 98% is achieved for the classification between normal an dysphonic voices.