Extraction of characteristics for the recognition of isolated words using the wavelet packet method

Several methods for digit recognition are affected by the speed of pronunciation. A method based on the decomposition of digits into several segments is presented. Each digit is divided into an equal number of overlapping segments. These segments are multiplied by a Hamming window. The signal is then split into 24 frequency bands similar to the Mel scale. The energy value for each of these 24 bands is calculated and normalised by the number of samples of the signal of the band concerned. The discrete cosine transform is then applied to the logarithms of each of the 24 band energy values. New parameters are thus obtained of which only some are used for recognition. The performance of the proposed method has been evaluated using the TIDIGITS database.