Performance Comparison of Multiple Speech Features for Speaker Recognition using Artifical Neural Network

Human speech is the natural way of communication between humans. Speech signal contains among others, information that conveys the message being spoken by a person, speaker characteristics and language. It also carries information regarding the speaker’s identity. The objective of Speaker Recognition Systems (SRS) is to automatically identify the speaker using the features extracted from their speech signals. Currently, SRS is one of the most popular biometric technique. It is used for surveillance, forensic speaker recognition, authentication, and such similar activities. In this work, Artifical Neural Network (ANN) based text-independent speaker identification is done. We compare the performance of Relative Spectral Amplitude-Perceptual Linear Prediction (RASTA-PLP), Mel-Frequency Cepstral Coefficient (MFCC) and Power Normalized Cepstral Coefficient (PNCC) features. To improve the performance accuracy of the system we extract features from the voiced frames of the signal. It was observed that MFCC outperforms the other two features namely PNCC and RASTA-PLP. TIMIT database was used for the experiments.