Published: IEEE Access

MGU-V: Lo-Fi Music Generation Using Variational Autoencoders

Overview

MGU-V is a deep learning-based framework for generating Lo-Fi music using a hybrid approach combining Long Short-Term Memory (LSTM) networks and Variational Autoencoders (VAEs). This model is capable of producing high-quality music by learning from a curated MIDI dataset and generating realistic-sounding Lo-Fi music, which is perfect for use in real-time applications like background scores, video games, or music streaming.

Features

• Hybrid Architecture: Combines the temporal modeling power of LSTMs with the generative capabilities of VAEs.
• Custom Datasets: Trained on a merged dataset of over 2300 MIDI files, including datasets like Nottingham Music Database, Maestro Piano Midi, and others.
• High Performance: Achieves state-of-the-art performance with an accuracy of 96.2% and minimal loss of 0.19.
• Lo-Fi Music Generation: Specifically designed to generate continuous, high-fidelity Lo-Fi music suitable for listening during concentration or background environments.