Speech parameterization based on AM-FM model and its application in speaker identification using AANN

This paper presents the parameterization of speech based on amplitude and frequency modul ation (AM - FM) model and its application to speaker identification. Speech parameterization is based on three different bandwidths viz 400Hz, 266mel, 106mel. The feature obtained by this parameterization is termed as PYKFEC which is not directly used as a f eature instead its average of each filter is used as the feature and termed as FAP. The speaker identification is done using auto associative neural network and Gaussian mixture model. The AANN/GMM is trained using the SOLO speaking style from CHAINS CORPU S database and a network/model is created for each speaker. The created model is tested using different speaking style like FAST and WHSP of the speaker. The identification rate of FAP is better than PYKFEC, and AANN performs well with these features.