Bangla speech isolation from noisy auditory environment using convolutional neural network

Abstract

In recent years, the primary solution to sound enhancement has gained popularity. There is a rich research contribution from academia and industry to remove noise and enhance sound quality. With the advance in machine learning and deep learn ing algorithms, well-performing audio enhancement models now exist. But such a sophisticated and well-researched model has not existed utilizing the language of Bangla. Although there have been models trained and tested to comprehend the language, no such model exists that can process real-time Bangla speech. Also, no such dataset exists that contains a substantial amount of speeches conducted in the Bangla language spanning over multiple hours. In this research, we stud ied the existing models that are working to separate noise in composite auditory environments, and on the basis of that study, we designed and implemented a U Net architecture model that has been trained in the Bangla language and is able to isolate and separate external noise from Bangla language speeches providing a clean feed to the listeners. Implementation of convolution neural networks in digital signal processing is a different approach and we achieved our desired results through it.