A Comparative Study of Machine Learning Classifiers for Speaker’s Accent Recognition
Hasibul Hasan Sabuj, Paul Richie Gomes, Kazi Ehsanul Mubin, Akram Hossain, Samin Yeasar Seaum, Md Tanzim Reza, Md. Golam Rabiul Alam · 2023
Recognizing a speaker’s accent is crucial in speech technology, natural language processing, and forensic linguistics, especially in English. With the English language having a diverse range of accents and being the most widely spoken language globally, it is vital to develop accurate accent recognition systems for speech recognition and language learning platforms. Our dataset comprises six different accents, including French English, Spanish English, German English, Italian English, UK English, and US English. Our proposed model utilizes Mel-Frequency Cepstral Coefficients (MFCCs) to generate numerical data from raw audio and predicts the speaker’s accent from our classification categories. This paper presents a comparative study of machine learning classifiers for speakers’ accent recognition. We conducted multi-class classification on our dataset and implemented necessary data preprocessing techniques to address the class imbalance and improve data quality. Eight different machine learning models, including Support Vector Machine, Random Forest, and XGBoost algorithm, were trained and compared for their performance. Random Forest achieved the best performance with an accuracy of 95.28%, and we provided a detailed analysis of each model’s strengths and weaknesses and their performance in each accent class.