Reproduce “DeepBach: a Steerable Model for Bach Chorales Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Gaëtan Hadjeres, François Pachet, Frank Nielsen
DeepBach is a deep learning model developed by researchers at Sony CSL and Universit Pierre et Marie Curie that generates four-part chorales in the style of J.S. Bach, emphasizing 'steerability' through user-imposed constraints. A human listening study showed that 50% of its generated music was perceived as authentic Bach compositions.
Key insights
- The use of a 'hold' symbol in data representation is crucial for efficient Gibbs sampling, allowing single variable changes to modify note durations.
- Combining forward and backward recurrent neural networks captures complex local context and musical directionality, leading to higher quality generation.
- Pseudo-Gibbs sampling enables a 'steerable' generation process, allowing users to impose arbitrary positional constraints (notes, rhythms, parts) during composition.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.