Speaker Identification System Enhanced By Optimized Neural Networks And Feature Fusion Techniques Evaluated By Cochlear Implant-Like Spectrally Reduced Speech
Abstract
Efficient feature extraction techniques are available in the literature in speaker identification system; yet, their combined influences on each other as a fusion feature have not been fully investigated. Much research has been conducted in terms of enhancing only one of either the feature extraction or classification techniques. Past research has not succeeded to focus in enough detail on determining more appropriate methods of evaluating parameters. In order to enhance the features extraction of any individual’s speech, we will concatenate two feature extraction techniques, along with the respective averages; Linear predictive, Mel-frequency, and the respective normalization Averages Cepstral Coefficients (LMACC) that would reflect positively on system performance. Optimized Radial Basis Function (RBF) neural network based on Bacteria Foraging Optimization algorithm is used as a classifier to improve system performance. Cochlear implant-like spectrally reduced speech proposed in the literature will be used alongside accuracy, sensitivity, and specificity to evaluate the system.