Event Title

Using Long Short Term Memory network to train machine composing Baroque Fugue/Canon.

Presenter Information

Yihe Chen

Faculty Advisor

Edisanter Lo

Start Date

24-4-2018 12:00 PM

End Date

24-4-2018 1:00 PM

Description

The goal of this project is to train a machine to compose Baroque Fugue/Canon by using Long-Short Term Memory architecture (LSTM). LSTM is a type of artificial recursive neural network (RNN), which excels at learning patterns at both long and short time periods. The network takes one symbol at a time, and estimates the probability of next symbol based on previous readings. It is suited to learn temporal structures of the input and applicable to automated music composition since music is temporal and often rich in structures. Due to difficulty in quantitating human aesthetics, art creation is a challenging problem in the field of Artificial Intelligence and an important criterion on evaluating the capability of the future strong AI. The performance of existing techniques for general purpose music composition has been unsatisfactory. By limiting to particular “styles” of structures and patterns, the problem becomes more tractable. In our study, we focus on the Baroque Fugue/Canons as they are polyphonic music with standard rules and structures to regulate the composing process. A 2-layer bi-directional LSTM network has been designed. Approximately 16 hour long data, which are midi files of Fugue/Canon primarily composed by Johan Sebastian Bach, have been collected from internet and translated into a text file. The length, frequency, intensity and timing were considered as training features. The checkpoint file with best validation loss was used to generate output file, and converted back into midi file. The result was evaluated by using phonologic criterions, which are Polyphony, Scale consistency, Repetitions and Tone span. The result shows that the model gained relatively strong counterpoint capability. However, it can not generate apparent melody.

This document is currently not available here.

Share

COinS
 
Apr 24th, 12:00 PM Apr 24th, 1:00 PM

Using Long Short Term Memory network to train machine composing Baroque Fugue/Canon.

The goal of this project is to train a machine to compose Baroque Fugue/Canon by using Long-Short Term Memory architecture (LSTM). LSTM is a type of artificial recursive neural network (RNN), which excels at learning patterns at both long and short time periods. The network takes one symbol at a time, and estimates the probability of next symbol based on previous readings. It is suited to learn temporal structures of the input and applicable to automated music composition since music is temporal and often rich in structures. Due to difficulty in quantitating human aesthetics, art creation is a challenging problem in the field of Artificial Intelligence and an important criterion on evaluating the capability of the future strong AI. The performance of existing techniques for general purpose music composition has been unsatisfactory. By limiting to particular “styles” of structures and patterns, the problem becomes more tractable. In our study, we focus on the Baroque Fugue/Canons as they are polyphonic music with standard rules and structures to regulate the composing process. A 2-layer bi-directional LSTM network has been designed. Approximately 16 hour long data, which are midi files of Fugue/Canon primarily composed by Johan Sebastian Bach, have been collected from internet and translated into a text file. The length, frequency, intensity and timing were considered as training features. The checkpoint file with best validation loss was used to generate output file, and converted back into midi file. The result was evaluated by using phonologic criterions, which are Polyphony, Scale consistency, Repetitions and Tone span. The result shows that the model gained relatively strong counterpoint capability. However, it can not generate apparent melody.