lyh710/deep_piano

Piano music generation using deep learning

deep_piano

Piano music generation using deep learning.

Aim

We believe there are largely 2 ways to generate music:

• Generate the sound waves, which we tried and failed (https://github.com/lyh710/deep_piano/tree/music_as_time_series)
• Generate notation (sheet music), which we focused on here.

Dev env

1. Windows 10 Home, i7 Core, 16GB ram (most subsequent steps should still hold true if with Mac/Linux, but some will need to be modified, such as the conda env setup batch)
2. NVIDIA GeForce GTX 1080 with Max-Q, 8GB (Google CoLab would be a good alternative)
3. conda env setup: dnn_gpu_setup_test\conda_dnn_gpu_setup.bat (Ananconda3)

High level idea

By converting music files into music-sheets (notations), we can empploy Deep-Learning technique that has been widely used in the NLP domain to generate music. The idea is largely similar to text generation with Deep-Learning, in which LSTM or other more sophisticated architecture can be employed.

Music files (midi) handling

This is mainly done by the music21 package: https://web.mit.edu/music21/doc/index.html

Ref

• https://www.midiworld.com/classic.htm
• https://towardsdatascience.com/how-to-generate-music-using-a-lstm-neural-network-in-keras-68786834d4c5
• https://machinelearningmastery.com/text-generation-lstm-recurrent-neural-networks-python-keras/