Native and Non-Native English Speech Classification: A premise to Accent Conversion

This research has been carried out with an intent of classifying if the English speech given as input is either a native or non-native accent with the help of Machine Learning classification algorithms on Mel Frequency Cepstral Coefficients. As the world has been evolving, so is the technology around us which has given rise to numerous voice-based tools. But the efficiency of these tools depends greatly on various unpredictable and pesky factors such as audio tempo and pronunciation which substantially differ throughout different accents and dialects. Hence, we chose to move ahead with working towards finding a method to best curb the problem of reduced accuracy in these technologies due to the said differences in accents and dialects. This has been carried out by applying five different machine learning classification models namely Gaussian Mixture Model, K-Nearest Neighbor, Logistic Regression, Support Vector Machine and Neural Network with an observed accuracy of 53%, 95%, 95%, 98% and 98% respectively.