Feature generator for speaker recognition using the fusion of cepstral and melcepstral parameters

The paper examines issues related to the determination of features distinctive to sound generators using a fusion of melcepstral and cepstral information for an Automatic Speaker Recognition (ASR) system. Parameterization of the speech signal is crucial to these systems, as the chosen parameterization dictates the effectiveness of the diagnosis and the speed of the system. The authors focus on the use of speech signal processing methods that consider the phenomena connected with the speech generation process while searching for features related to speech characteristics. A well-designed system should be able to extract speech characteristics independent of the linguistic content of the speech. The research presented in this paper focuses primarily on multicriteria optimization of the generator parameters based on a set of descriptors derived from the fusion of the melcepstra and cepstra and also considers the use of additional features. Finally, the evaluation of the results was based on the analysis of the Fisher coefficients and main components of a set of descriptors.