Speaker verification for security systems using artificial neural networks

This paper investigates automatic speaker recognition systems, which can be used for security purposes. The speech signal is compressed using linear prediction analysis and recognized by neural networks. This neural network technique is presented for the task of speech recognition and speaker verification. This technique first uses pattern recognition to identify the speech, then it is used to distinguish each user from all other speakers (impostors). With this method, unknown speech can be accurately classified as user or impostor speech. The approach used is based on the following steps: extraction of spectral features; training of an initial neural network to identify the speech; extraction of LPC-reflection coefficients for each user, training of a secondary neural-network to identify the user; and classification of unknown speech as either user or impostor.