Bangladeshi Dialect Recognition using MFCC, Delta, Delta-delta and GMM

Abstract

Automatic recognition systems are commonly used in speech processing to classify observed utterances by the speaker identity, dialect and language. A lot of research has been performed to detect speeches, dialects and languages of different region throughout the world. But the work on dialects of Bangladesh is infrequent to our research. These dialects, in turn, differ quite a bit from each other. In this paper, we propose a method to detect Bangladeshi different dialects which utilizes Mel-Frequency Cepstral Coefficients (MFCC), its Deltas and Delta-Deltas as main features and Gaussian Mixture Models (GMM) to classify characteristics of a specific dialect. Particularly we extract the MFCC, Deltas and Delta-Deltas features from the speech signal. Then they are merged together to form a feature vector for a specific dialect. GMM is trained using the iterative Expectation Maximization (EM) algorithm where feature vectors are served as input. This scheme is tested on 5 databases of 30 speech samples each. Speech samples are contained dialects of Borishal, Noakhali, Sylhet, Chittagong and Chapai Nobabgonj regions of Bangladesh. Experiments show that GMM adaptation gives comparable good performance.