Robust and Optimum Features for Persian Accent Classification Using Artificial Neural Network

摘要

This paper presents a classification model for regional accents of Persian. The model is based on a combination of the conventional speech coding and pattern recognition techniques. In this study, the well-known multilayer perceptron plays the role of the classifier. Moreover, a wide variety of speech coding techniques is utilized for feature extraction. Among them, we determine the robust and optimum features for this task by comparing the classification performance. The method is validated on a corpus containing recordings from ten speakers, five males and five females, for each accent. Results show that perceptual linear predictive (PLP), relative spectral transform PLP (Rasta PLP), and linear predictive coefficient (LPC) perform well under both clean and noisy conditions.