Speaker-oriented Feature Optimization Approach

Focusing on the popular feature parameters applied in speaker recognition, this paper produces the scheme of the effective parameter optimization. In particular, optimized mel-frequency cepstrum coefficients obtained by adjusting the parameters of the filterbank using unidimensional search reveal their improved ability to represent the vital identification information embedded in the utterances of the different speakers. To evaluate the performance of the optimized parameters, the recognition experiments addressing the speech data of Chinese digit words have been conducted and the average recognition accuracy increases by 0.94% consistently