A novel approach to the speaker identification over telephone networks

The authors propose a channel equalization approach to overcome the channel variation in speaker identification over a telephone system. A set of inverse channel spectra is chosen to form a basis for generating the estimated channel equalization filter. When a short utterance goes through the basis filters, the coefficients of a convex combination are produced automatically so that a channel equalization filter is generated to compensate the channel variation embedded in the speech signal. A neural network is used to optimize the coefficients in convex combination. For the case of using a 100-speaker database, an average error rate of 5.2% can be achieved in simulation.<>