yi-ting0129/music-composition

Music-Composition

Title

Music Composition/Generation with Deep Learning Networks

Team members

Yen-Sung Chen (chenpine), Yi-Ting Cheng (Yi-Ting0129)

Introduction

Interacting with music, people always play roles like performers, composers, and listeners. We commonly involve in several music activities including singing, playing instruments, improvising, and composing. Deep Learning Networks music system can identify different types of structure from composer's work and then use them to produce a new work in the similar style. Therefore, Deep Learning Networks can help people compose music, which introduces the masterpieces of musicians. It will change people's original inspiration and creativity in the music.

Goals

We will start from quickly picking up the basic music theory, which will subsequently help us understand how we manipulate audio/music data. After reviewing literatures, we need to decide the data format that we want to feed into the proposed neural network(s). We would like to test our neural network on as many as possible genres of music, and see if there is performance discrepancy between them. Lastly, we collect our result and compare with existing models and give conclusion to our project.

Steps Involved:

1. Quick review on basic music theory
2. Understand audio/music data manipulation
3. Process potential data formats and implement network(s)
4. Generate music pieces with different genres, e.g., classical, jazz, metal, post-rock
5. Result analysis and benchmark

The objective of this project: Analyze a large music collection (including pop music, classical music, jazz, etc.) to determine structure, pattern, and style of the songs; furthermore, learn and imitate the melody, harmonious, timbre, and rhythmic pattern to create unique music.

Challenges

1. Unlike classical music, other genres may not have existent post-processed data.
2. Different genres may require unique data preprocessing.
3. Learn to implement advanced deep learning methods like Long Short Term Memory (LSTM) and Recurrent Neural Network (RNN).
4. At this moment, we are not sure if our laptops without GPU are capable of training and generating large amount of music data.
5. Build a neural network architectures that effectively extract the notions of harmony and melody
6. Create a model capable of learning music structure and possessing the ability to build a melody piece
7. Create music with musical rhythm, complex struc