Evaluating Machine Learning Models for Nigerian Languages Audio Classification Using MFCC Features

As communication among societies increasingly relies on speech-based technologies, the ability to accurately identify languages across linguistically diverse cultures is crucial for various applications. This project outlines a comparative study of three machine learning models: Support Vector Machine, Recurrent Neural Network, and Convolutional Neural Network, for the classification of three major Nigerian language (Hausa, Igbo and Yoruba) audio files. The goal is to evaluate the reliability of these models in accurately predicting spoken Nigerian language (Hausa, Igbo and Yoruba) based on extracted audio features. A fundamental component of the research work is the application of a pre-emphasis filter to the audio signals and the extraction of Mel-Frequency Cepstral Coefficients, which serve as the primary features for training the models. The Support Vector Machine model is configured to handle non-linear feature spaces, while the Recurrent Neural Network model leverages its inherent ability to capture temporal dependencies within audio sequences. The Convolutional Neural Network model is utilised for its ability to learn spatial hierarchies in the Mel-Frequency Cepstral Coefficient representations, effectively capturing local feature patterns. Performance evaluation is administered using standard model evaluation metrics to determine the strengths and limitations of each model in the context of language classification. This analysis also provides insights into the suitability of different machine learning approaches for language audio classification tasks, particularly in real life applications (such as speech based language recognition systems) requiring sensitivity to accent and voice patterns. The results indicate that the Convolutional Neural Network model achieved 83.33% accuracy, surpassing both Recurrent Neural Network (78%), and Support Vector Machine (68.17%) models.