Speaker identification based on Gammatone cepstral coefficients and general regression neural network

This paper presents a speaker identification method using Gammatone cepstral coefficients extracted by Gammatone filters and a group of general regression neural networks. The Gammatone cepstral coefficients are adapted to the characteristics of speech signals through adjusting the Gammatone filter and the filter bank settings. To reduce the training data and time cost of the general regression neural network used as the classifier for speaker identification, the non-linear partition algorithm is employed to divide the Gammatone cepstral coefficients used as the speech features. In this sense, the speaker identification task is partitioned into a number of small tasks which can be operated by a group of general regression neural networks. Each recognition rate of these general regression neural networks is integrated into the final recognition rate of the speech signals. The results indicate that the proposed method has an acceptable recognition rate with high accuracy.