Identifying explicit language in songs typically requires listening to the track or reading the lyrics. However, Deezer is exploring the potential of AI to handle this task more efficiently. The streaming service is developing a machine learning technique that can detect explicit lyrics directly from audio recordings.
Rather than relying on an extensive database of annotated samples, Deezer’s approach involves isolating the vocals and analyzing word occurrences that could match entries in a dictionary of explicit terms. A binary classifier then determines whether a specific word is explicit. This "explainable" system provides insights into the AI's decision-making process.
Deezer aims to enhance accuracy and minimize bias by incorporating an equal number of explicit and clean songs across various genres. While the current method shows promise, the company acknowledges that it is not yet ready for practical application. Although this AI system significantly improves upon traditional approaches to profanity detection, it still lacks the efficiency of AI with access to song lyrics or human reviewers.
Nonetheless, this technology has the potential to aid human curators in determining whether a song should bear an explicit content label. Ultimately, the goal is to develop the AI to operate independently, easing the burden of song tagging and reducing the likelihood of children inadvertently hearing inappropriate language in tracks considered 'safe.'