Persian Accents Identification Using Modeling of Speech Articulatory Features

One of the most commonly used areas of speech processing is the accent recognition system, that recognizes a speaker’s accent from speaking it. In this study, using hierarchical neural networks, speech articulatory features in Persian speech such as "Attribute Manner" and "Attribute Place "are appropriately modeled and extracted. Subsequently, using native modelers such as GMM_UBM and I_vector, 5 native accents were identified in FarsDat data. In this study, it has been shown that speech articulatory features usually perform better in distinguishing and identifying native Persian accents than spectral features such as MFCC and SDC. The results show that by applying a dimensionality reduction such as PCA to the Attribute Manner of phoneme of the long-term speech articulatory features, the average error rate is 8.37% and the average accent detection cost is 6.68. Compared to conventional features like MFCC + SDC, they can reduce relative error rates and average error costs by 45.12% and 29.28%, respectively. In addition, the best average accuracy for the 5 native accents is 75.29%. These features also exhibit better resistance to conventional spectral features such as MFCC and SDC when there is a shortage of educational data.