Spoken Malay Profanity Classification Using Convolutional Neural Network

Abstract

Foul language exists in films, video-sharing platforms, and social media platforms, which increase the risk of a viewer to be exposed to large number of profane words that have negative personal and social impact. This work proposes a CNN-based spoken Malay foul words recognition to establish the base of spoken foul terms detection for monitoring and censorship purpose. A novel foul speech containing 1512 samples are collected, processed, and annotated. The dataset then has been converted into spectral representation of Mel-spectrogram images to be used as an input to CNN model. This research proposes a lightweight CNN model with only six convolutional layers and small size filters to minimize the computational cost. The proposed model’s performance affirms the viability of the proposed visual-based classification method using CNN by achieving an average Malay foul speech terms classification accuracy of 86.50%, precision of 88.68%, and F-score of 86.83. The class of normal conversational class outperformed the class of foul words due to data imbalance and rarity of foul speech samples compared to normal speech terms.