Prosodic Modelling based Speaker Identification

Abstract

The use of prosodic characteristics, mainly pitch and intensity, for speaker identification in noisy environments is examined in this work. To make the acoustic models more resistant to the variability in the speech signal in noisy situations, these features are supplemented with cepstral parameters. As a consequence, two systems for Automatic Speaker Identification (ASI) in the independent mode of text are implemented. The first based on Hidden Markov Models (HMM), whereas Support Vector Machines (SVM) are employed in the second. The addition of prosodic features improves recognition, especially in high-noise environments. The performance of SVM-based systems is better than HMM-based systems