Implementing Mood-Based Music Recommendations through Spectral Feature Analysis and Instrument Separation

Abstract

Mood-based music recommendation systems have the potential to significantly enhance the user experience by tailoring music selections to fit specific emotional states. This thesis presents a comprehensive approach to developing such a system by leveraging digital signal processing (DSP) and machine learning techniques. The proposed system collects user input via a web-based interface, where users report their current emotional states and music preferences for various situations. Utilizing the Demucs algorithm, the system decomposes MP3 files into individual instrumental tracks, allowing for detailed analysis of each component’s spectral features and emotional connotations. A hybrid model inspired by the U-Net architecture, incorporating both spectrogram and waveform separation, is used for this purpose. The mood assessment process, implemented with Streamlit, enables accurate capture of user emotions, which are then translated into mood weights influencing the recommendation process. This methodology ensures that the recommended tracks align with the user’s emotional state and context, providing a more personalized and engaging music experience. The results demonstrate the efficacy of the system in enhancing user satisfaction through contextually relevant music recommendations.