Supervised Machine Learning Model for Accent Recognition in English Speech Using Sequential MFCC Features

Abstract

Human–machine interfaces are rapidly evolving. They are moving from the traditional methods of input like keyboard and mouse to modern methods like gestures and voice. It is imperative to improve voice recognition and response since there is a growing market of technologies, world-wide, that use this interface. Majority of English speakers around the world have accents which are not exposed to speech recognition systems on a greater scale. In order to bridge the comprehension gap between these systems and the users, the systems need to be tuned according to the accent of the user. Accent classification is an important feature that can be used to increase the accuracy of comprehension of speech recognition systems. This paper recognizes Indian and American English speakers and distinguishes them based on their accents by constructing sequential MFCC features from the frames of the audio sample, oversampling the under-represented data and employing supervised learning techniques. The accuracies of these techniques reach a maximum of 95% with an average of 76%. Neural networks emerge as the top classifier and perform the best in terms of evaluation metrics. The results gleaned indicate that concatenating MFCC features sequentially and applying an apposite supervised learning technique on the data provide a good solution to the problem of detecting and classifying accents.