BachBot

Automatic stylistic composition

Feynman Liang

Tsingua University, 23 October 2017

Overview

• Music theory primer
• Corpus preparation
• Neural networks primer
• Results and analysis
• Musical Turing test

Music theory primer

Piano Roll

Sheet Music

Corpus preparation

Transpose/Quantize

Handling Polyphony

• Question: How many chords can be constructed from 4 voices, each with 128 pitches?
• Answer: $O(128^4)$

Handling Polyphony

• Reduce vocabulary size by treating parts independently
• Tokenize individual notes (instead of chords) in SATB order, explicit frame delimiters "|||"

								
START
(59, True)
(56, True)
(52, True)
(47, True)
|||
(59, True)
(56, True)
(52, True)
(47, True)
|||
(.)
(57, False)
(52, False)
(48, False)
(45, False)
|||
(.)
(57, True)
(52, True)
(48, True)
(45, True)
|||
END
								
							

Corpus statistics

Neural networks primer

Neurons

Feedforward Network

Memory Cell

Unrolling / Stacking

Sequential prediction

Backprop

LSTM

Results and analysis

Dropout

Neuron activations

Model learns music theory

Sample compositions

Music Turing Test

bachbot.com

Participants

Participants

Results

Results

Conclusion

• Deep LSTM generative model for composing, completing, and generating polyphonic music
  • Integrated into Google Magenta
• Model appears to learn music theoretic without prior knowledge
• Largest music Turing test to date with over 1779 participants
  • Average performance on music Turing test only 7% better than random guessing