A novel technique of data compression for spoken word recognition systems

A technique is described for data compression which helps in reducing the storage of data per reference/test word without degrading the recognition accuracy. An effort has been made to solve this problem in two stages. Feature extraction of the voice sample is performed using a conventional filter bank speech spectrum analyzer to get a time domain representation of the speech signal into time averaged frequency domain descriptions known as frames. The second stage of data compression deals with the reduction of the encoded crude data matrix. The number of rows in the data matrix of a word is reduced to a uniform size of rows. This data compression technique is based on the variable time segment coding of individual frequency channels. The performance of this technique was evaluated for the recognition of spoken isolated words using a Z-80A microcomputer