Predicting regional accents of Bengali language using deep learning

Abstract

Accent is a huge challenge in communication for all languages. Different people who speak the same language might pronounce the same word differently. In a conversation, if two people are from different regions and they have different accents, we can use our intuition to make sense of what the other person is saying. Sometimes, even our intuition cannot help determining the meaning of the words because of the difference in accent. Therefore, it is extremely difficult for an ASR (Automatic Speech Recognition) system to properly understand the words when the speaker uses different accent instead of the standard or formal accent as most of the time the ASR systems are trained on the formal or standard language. Now a days, most of these issues caused by accents are somewhat worked upon in most used languages like English, Mandarin and few other languages. However, the ASR systems used for Bengali Language is still at its infancy and different accents are a major issue. Finding audio features that differentiate the accents from one another and creating models to predict the accent using Deep Learning techniques will help to create a much better ASR System for Bengali Language. This paper will emphasize on creating few models which can determine the regional accent of Bengali language given an audio sample. Furthermore, after getting the accuracy of the individual models we can choose the model which results in the most accuracy. Further work can be done based on the models to create an ASR System for Bengali language which will be able to handle few more accents than the standard one.