BachBot: Composing Bach Chorales using Deep Learning
summary
BachBot uses a three-layer LSTM trained on the Music21 corpus of all known Bach chorales after transposing to C major/minor, quantizing durations, and encoding soprano→alto→tenor→bass with frame delimiters; the corpus is ~423,000 tokens with a vocabulary of 108. Training uses dropout (~0.3 favored) and next-token prediction; generated harmonizations sound baroque, and a web musical Turing test with ~1,700 responses shows human listeners beat chance by only about 5–7%.
AI-written summary · about
chapters (8)
- 0:00Introduction
- 1:15Agenda
- 2:08Music Theory
- 4:18The AI
- 5:21The Question
- 9:02Recurrent Neural Network
- 13:27Results
- 22:00Questions
related talks
32mGetting started with generative AI with AWS DeepComposer
01h 04mKeynote 3. Interactive generative models for music
54mIntegral scherzo: how machines compose and sing
41mMaking Machines that Make Music
21mMusical Deep Neural Networks in the Browser
43mMaking machines that make music
51mThe Merger of Machine Learning, Music and Programming
41mMusic and Machine Learning