Profanity Filtering in Speech Contents Using Deep Learning Algorithms

Abstract

The worldwide online exposure has significantly increased as a result of the Covid-19 pandemic and remote working, online learning, e-commerce have all become the norm. This has drastically increased the use of hate speech, swear words, racial slurs and many other inappropriate contents on the online platforms. These inappropriate contents are slowly degrading the quality of the online user experiences. Consequently, automatic detection and filtering of such inappropriate contents has grown to be a significant issue for enhancing the calibre of contents. Inappropriate contents may include profanity, violence, misleading information, sexually explicit material, extremism, and it may occur in textual, audio or video forms. In this study, a methodology for profanity filtering in speech contents is proposed. The proposed methodology focuses on identifying the audio segment ‘fuck’ which is the most frequently used swear word in the English Language. Audio segments related to swear words and non-swear words were collected, annotated, pre-processed, and analysed for the development of a RNN configuration by using Mel Frequency Cepstral Coefficients (MFCCs) as inputs to the model.