Protecting Children from Harmful Audio Content: Automated Profanity Detection From English Audio in Songs and Social-Media

Abstract

A novel approach for the automated detection of profanity in English audio songs using machine learning techniques. One of the primary drawbacks of existing systems is only confined to textual data. The proposed method utilizes a combination of feature extraction techniques and machine learning algorithms to identify profanity in audio songs. Specifically, the approach employs the popular feature extraction techniques of Term frequency–inverse document frequency (TF-IDF), Bidirectional Encoder Representations from Transformers (BERT) and Doc2vec to extract relevant features from the audio songs. TF-IDF is used to capture the frequency and importance of each word in the song, while BERT is utilized to extract contextualized representations of words that can capture more nuanced meanings. To capture the semantic meaning of words in audio songs, also explored the use of the Doc2Vec model, which is a neural network-based approach that can extract relevant features from the audio songs. The study utilizes Open Whisper, an open-source machine learning library, to develop and implement the approach. A dataset of English audio songs was used to evaluate the performance of the proposed method. The results showed that both the TF-IDF and BERT models outperformed the Doc2Vec model in terms of accuracy in identifying profanity in English audio songs. The proposed approach has potential applications in identifying profanity in various forms of audio content, including songs, audio clips, social media, reels, and shorts.