DeepMelody: Personalized Music Recommendations Based on Instrumental Composition

Overview

DeepMelody is an innovative project designed to revolutionize music recommendation systems by focusing on the prominence of individual instruments within audio tracks. Leveraging advanced machine learning techniques, DeepMelody enables users to receive personalized music suggestions tailored to their instrumental preferences, moving beyond traditional genre-based recommendations.

At the heart of DeepMelody lies the Slakh2100 dataset, a comprehensive collection of labeled multitrack audio files spanning various genres. We employ a U-Net-based model for music source separation, effectively isolating spectrograms for individual instruments. From these spectrograms, we derive instrument embeddings, which quantify the prominence of each instrument in a track. By measuring cosine similarity between these embeddings, the system identifies tracks that align with the user’s instrumental preferences.

This project empowers users to explore and discover music based on the instruments they love, creating a highly personalized and immersive listening experience.

Key Features

Instrument Isolation:
Leveraging a U-Net-based model, the system effectively isolates individual instrument spectrograms, enabling detailed analysis of each instrument's prominence within a music track.

Instrument Embeddings:
Advanced embeddings are generated to represent the significance and characteristics of each instrument, forming the foundation for highly personalized music recommendations.

Personalized Music Recommendations:
Moves beyond traditional genre-based algorithms by tailoring recommendations based on a user’s specific instrumental preferences, offering a more unique and engaging music discovery experience.

Ongoing Research and Innovation:
Focused on enhancing the quality of instrument embeddings to further refine the understanding of instrumental prominence, paving the way for even more precise and relevant music recommendations.

Instrument Isolation:
Leveraging a U-Net-based model, the system effectively isolates individual instrument spectrograms, enabling detailed analysis of each instrument's prominence within a music track.

Instrument Embeddings:
Advanced embeddings are generated to represent the significance and characteristics of each instrument, forming the foundation for highly personalized music recommendations.

Personalized Music Recommendations:
Moves beyond traditional genre-based algorithms by tailoring recommendations based on a user’s specific instrumental preferences, offering a more unique and engaging music discovery experience.

Ongoing Research and Innovation:
Focused on enhancing the quality of instrument embeddings to further refine the understanding of instrumental prominence, paving the way for even more precise and relevant music recommendations.

Project Workflow

We are currently planning to Increase the number of Recommendations per song and increasing the number of instruments to select from 

Here's the Github to the Project