Application of deep learning approach for recognition of voiced Odia digits

Automatic speech recognition in a regional language like Odia is a challenging field of research. Voiced Odia digit recognition helps in designing automatic voice dialler systems. In this study, a deep learning approach is used for the recognition of voiced Odia digits. The spectrogram representation of voiced samples is given as the input to the deep learning models after considering the feature extraction using MFCC. Various performance metrics are obtained by considering several experiments with different epoch sizes and variation in the dataset using the train-validate-test ratio. Experimental outcomes reveal that the CNN model provides improved accuracy of 91.72% in epoch size of 500 with a split ratio of 80-10-10 as compared to the other two models that use VSL and DNN. From the reported outcome it unravels that, the proposed CNN model has better average recognition accuracy as compared with contemporary models like HMM and SVM.