A Comparative Study of Machine Learning Models for Identifying Nigerian Languages from Audio Clips

Language is a vital aspect of human culture and communication, with over 7,000 languages spoken worldwide. Despite this diversity, many languages, particularly those spoken in Africa, lack extensive research and support for speech processing technologies. This study focuses on identifying five Nigerian languages, which are Yoruba, Nigerian Pidgin, Nigerian English, Hausa, and Igbo using machine learning techniques. We utilized Mel-Frequency Cepstral Coefficients (MFCCs) to extract acoustic features from audio clips and trained several models, including 1-Dimensional Convolutional Neural Network (1D CNN), 2-Dimensional Convolutional Neural Network (2D CNN), Deep Neural Network (DNN), Long Short-Term Memory (LSTM), Random Forest, and Support Vector Machine (SVM), to evaluate their effectiveness in language identification. The dataset comprised audio from OpenSLR, Common Voice, local recordings, and NWU, totaling 37.8 hours across all languages. Results indicated that the LSTM model achieved the highest accuracy (0.830) and precision (0.856), while the DNN model showed the best F1 score (0.797) and recall (0.852). Neural network-based models consistently outperformed traditional machine learning approaches. We observed varying levels of classification difficulty among the languages, with some presenting more challenges than others. The dataset's uneven distribution across languages posed additional complexities in model training and evaluation. This research underscores the importance of diverse data sources and advanced neural network architectures in improving language classification for underrepresented African languages.