Speech Recognition of Isolated Odia Digits Using Convolutional Neural Network

Automatic Speech Recognition (ASR) is the process of recognition of voice by a computer or device. ASR has different applications such as Speaker Identification, voice based IVR system, automatic transcription system, etc. This paper proposes a Convolutional neural network (CNN) model for ASR of isolated digits spoken in the Odia language. Forty Mel Frequency Cepstral Coefficient (MFCC) features have been extracted for each isolated digit. The proposed acoustic model is implemented using CNN. The dataset is collected by recording the utterance of the digits (0-9). The model is tested with different dropout and learning rates. The proposed model gives better accuracy than other classification algorithms like DNN, HMM, SVN, and Random Forest. We have achieved an accuracy of 95.71%.