Automatic Audio Replacement of Objectionable Content for Sri Lankan Locale

Abstract

Fake news, hate speech, crude language, ethnic and racial slurs and more have been spreading widely every day, yet in Sri Lanka, there is no definite solution to save our society from such profanities. The method we propose detects racist, sexist and cursing objectionable content of Sinhala, Tamil and English languages. To selectively filter out the potentially objectionable audio content, the input audio is first preprocessed, converted into text format, and then such objectionable content is detected with a machine learning filtering mechanism. In order to validate its offensive nature, a preliminary filtering model was created which takes the converted sentences as input and classifies them through a binary classification. When the text is classified as offensive, then secondary filtering is carried out with a separate multi-class text classification model which classifies each word in the sentence into sexist, racist, cursing, and non-offensive categories. The models in preliminary filtering involve the Term Frequency–Inverse Document Frequency (TF-IDF) vectorizer and Support Vector Machine algorithm with varying hyperparameters. As for the multi-class classification model for Sinhala language, the combination of Logistic Regression (LR) and Countvectorizer was used while the Multinomial Naive Bayes and TF-IDF vectorizer model was found suitable for Tamil. For English, LR with Countvectorizer model was chosen to proceed. The system has an 89% and 77% accuracy of detection for Sinhala and Tamil respectively. Finally, the detected objectionable content is replaced in the audio with a predetermined audio input.