Speech Abusive Language Detection System Using MFCC Speech Feature Extraction and Convolutional Neural Network

As bullying rapidly spreads in schools, many parents are becoming increasingly concerned about their children being subjected to bullying by peers, which could negatively impact their physical and mental health. This paper proposes an abusive language detection system that combines Mel-frequency Cepstrum Coefficients (MFCC) speech features with a Convolutional Neural Network (CNN). Audio files which contain daily abusive phrases in the Minnan dialect are initially preprocesses to eliminate non-speech interference. Subsequently, the MFCC algorithm is applied to extract speech features that closely capture auditory characteristics. The CNN classification model is trained to learn the MFCC features and classify the phrases into their respective categories. Experimental results indicate that the CNN model achieves an average recognition accuracy of 91.9%. Finally, the CNN model was deployed on an embedded platform, where it achieves an average recognition time of 0.54 seconds per phrase and an average recognition accuracy of 86.25%.