Renyi entropy and deep learning-based approach for accent classification

Accent classification has been gained more attention, due to improving demands for better speaker recognition with the accented speech. An accent is defined by the pronunciation of the language, particularly with the locality, social class, or particular nation. Thus, this paper proposes a model for accent classification using feature extraction and classifier. At first, the input signals are pre-processed. Here, the spectral skewness, spectral centroid, tonal power ratio, spectral kurtosis, spectral flux, and Renyi entropy-based Multi kernel Mel Frequency Cepstral Coefficient features (ReMKMFCC) features is adapted for the feature extraction, and the ReMKMFCC feature is derived by the combination of Renyi entropy, Multiple Kernel Weighted Mel Frequency Cepstral Coefficient (MKMFCC). After the extraction of the features, the features are fed as input to the feature selection module. The feature selection is carried out using Tanimoto, and then the selected features are forwarded to the classification module, where the features are classified using DBN, and the classifier is trained by the proposed Dragonfly-Bird swarm optimization (DBSO), which is the combination of Dragonfly Algorithm (DA) and Bird swarm optimization algorithm (BSO). Thus, the DBSO-based DBN aims at classifying the accent. The analysis proves that the proposed method acquired a maximal accuracy of 96.96% by considering dataset-1, maximum F-Measure of 96.97% by considering dataset-2, and minimal FAR of 3.04%, by considering the dataset -1.