Content-Based Tagging and Recommendation System for Tamil Songs Based on Text and Audio Input

Music recommendation systems play a crucial role in delivering personalized music suggestions by leveraging algorithms to analyze user data. However, the challenge of data sparsity arises when users have limited listening history or ratings. Content-based filtering involves extracting high-dimensional audio feature vectors, enabling comparisons with similar tracks based on audio characteristics unlike collaborative methods rely on user feedback, considering factors like listening duration, usage time and location, and genre/artist popularity. Concerns regarding user privacy and trust have driven a focus on content-based tagging and retrieval, incorporating music and lyrics. This work proposes a content-based approach for music tagging and retrieval through music segmentation and lyrics-based tagging. Particularly, this work focuses on Tamil songs and attempts to cater to the large and diverse audience of Tamil songs. The music segmentation is carried out through a model that utilizes a Gaussian Mixture Model along with an encoder–decoder architecture trained on MUSDB18 dataset which produces a high Source Artifact Ratio (SAR) of 11.004 in the vocal component. The model provides a significant improvement in the Signal-to-Distortion ratio compared to leading models. The system also allows the user to enter a text prompt and lists the five most relevant song, allowing the user to exercise more control in the recommendations provided, while ensuring the user’s privacy. In order to quantify the relevance of the recommendations, two measures are introduced, namely VibeMatch and ExploreMatch.