Music Generation with Machine Learning and Deep Neural Networks
Tudor-Constantin Pricop, Adrian Iftene · Procedia Computer Science · 2024
This paper explores advanced music generation through hybrid models combining deep neural networks, machine learning algorithms, variational autoencoders (VAEs), long short-term memory (LSTM) networks, and Transformers to create diverse and engaging musical experiences. Our research aims to advance the understanding of music’s impact on our lives and develop methodologies to create diverse and engaging musical experiences tailored to individual preferences. We begin by extracting relevant features from a large and diverse collection of music samples from different genres. These features, encompassing spectral properties, rhythmic patterns, and tonal characteristics, serve as the foundation for our generation models. To generate music, we explore the potential of VAEs, LSTMs, and Transformers, each offering unique capabilities for handling different aspects of the task. VAEs are employed to learn a continuous latent space representation of the music samples, enabling the generation of novel compositions within a specified genre. LSTMs and Transformers, on the other hand, are used to model the temporal dependencies and intricate patterns inherent in music. While not claiming state-of-the-art performance, our approach demonstrates promising outcomes in generation tasks, showcasing its potential to enhance music-related applications such as recommendation systems and creative tools for composers.