Assamese Dialect Identification System Using Deep Learning

Abstract The goal of a dialect Identification is to label speech in an audio file with dialect labels. This paper presents a method for automatically identifying four Assamese dialects: Central Assamese, Eastern Assamese dialect, Kamrupi dialect, and Goaplari dialect, using Convolution Neu-ral Networks (CNN). In this study, utterances of four major regional dialects of the Assamese language, namely Central Assamese spoken in and around Nagaon district, Eastern Assamese dialect spoken in the Sibsagar and its neighboring districts, Kamrupi dialect spoken in Kam-rup, Nalbari, Barpeta, Kokarajhar and some parts of Bongaigaon district and Goaplari dialect spoken in the Goaplara, Dhuburi and part of Bon-gaigaon district were used. The classifier was trained on audio samples from each of the four dialects that lasted 2 hours. The CNN uses Mel spectrogram images created from two to four seconds divisions of raw audio input with varied audio quality. The performance of the system is also examined as a function of train and test audio sample durations. When compared to machine learning models, the suggested CNN model obtains an accuracy of 90.82 percent, which may be considered the best.