Persian Accent Identification: A Review

Accent identification and recognition based on speech signals is a significant and growing branch of speech processing that affects the performance of Automatic Speech Recognition (ASR) systems; if the input accent differs from the accent on which the ASR system was trained, its accuracy decreases noticeably. The comprehensive review has investigated advancements and various approaches employed in Persian accent identification. Studies in the field focus on the four main stages: pre-processing, feature extraction, classification, and evaluation. In terms of speech features, a significant amount of research has been conducted on conventional spectral features, such as Mel-Frequency Cepstral Coefficients (MFCC) and their derivatives; efficient methods of feature extraction, including Spectral Centroid Magnitude (SCM) and the Zak transform. Furthermore, the role of prosodic features, including general statistics regarding lower and upper contour, energy, and their derivatives, as well as the investigation of articulatory features such as "Attribute Manner" and "Attribute Place", in distinguishing accents has been evaluated. In the classification section, various models, for example, Gaussian Mixture Model (GMM), Multi-Layer Perceptron (MLP), Radial Basis Function (RBF), Probabilistic Neural Network (PNN), Support Vector Machine (SVM), Phoneme Recognition followed by a Language Model (PRLM), have been employed. A key process that significantly improves the identification rate is employing "Ensemble Methods", and the combination of classifiers that has led to higher performance than using single classifiers, for example, using binary combinations with majority vote. The results of the studies have shown that articulatory and prosodic features often yield better performance in identifying local accents and dialects than acoustic features.