Soft-computation based speech recognition system for Sylheti language

The encouraging trend of usage of human machine interfaces in diverse areas has driven the evolution of Automatic Speech Recognition (ASR) systems during last two decades. Lately, the inclination has been towards the use of machine learning techniques for under-resourced human languages primarily to focus on designing of voice activated digital tool for a sizable portion of computer illiterate speakers. A vast majority of the works in this field have employed shallow models like conventional Artificial Neural Network and Hidden Markov Model in combination with Mel Frequency Cepstral Coefficients and other relevant features for the applications of speech recognition systems. Although these shallow models are found effective, but to minimize human intervention from the approach and also to yield the better system performance, recent research has focused to incorporate deep learning models for ASR applications especially for under-resourced languages. Sylheti language, a member of Indo-Aryan language group, is an under resourced language which has more than 10 million Sylheti speakers living across the world mostly in India and Bangladesh. Focusing on the need of an ASR model for Sylheti, this work aims to design a robust ASR model for an under resourced language Sylheti by employing state-of-the-art deep learning technique Convolutional Neural Network (CNN). To find out the best and suitable ASR model for Sylheti, certain ASR approaches are formulated and trained by Sylheti isolated and connected words. The specially configured ASR model based on CNN is trained with clean, and noisy speech data which are necessary for training and making the system robust. Thereafter, a comparative analysis is presented by configuring the ASR model by some shallow models like Feed-forward neural network, Recurrent neural Network, Hidden Markov model and Time Delay neural Network. Experimental results indicate that the proposed CNN based ASR system works well for Sylheti language and the performance accuracy obtained by the system is found to be satisfactory despite the system demonstrating certain training latency.