Acoustic Features Based Accent Classification of Kashmiri Language using Deep Learning

Abstract

Automatic identification of accents is important in todays world, where we are souranded by ASR systems. Accent classification is the problem of knowing the native place of a person from the way He/She speaks the language into consideration. Accents are present in almost all the languages and it forms an important part of the language. Accents are produced from prosodic and articulation characteristics; in this research the aim is to classify accents of Kashmir Language. We have considered using the MFCC and Mel spectrograms for our research. A lot of research has been done for languages like English and is being done in this field and many models of machine learning and deep learning have shown state of the art results, but this problem is new for Kashmiri Language. The accents in Kashmir, vary from area to area and we have chosen 6 areas as our classes. We extracted the features from the audio data, converted those features into Images and then used the CNN architectures as our model. This research can be taken as base research for further researches in this language. Our custom models achieved the loss of 0.12 and accuracy of 98.66% on test data using Mel spectrograms, which is our best for our features.