Do Androids Dream of Electronic Dance Music?
summary
A pipeline converts 110 MIDI files (royalty-free deadmau5 samples) into TFRecords and trains an LSTM via TensorFlow and Magenta (mostly Python with ~5% Ruby glue). About ~1,000 training epochs on a MacBook Air take ~40 minutes and produce MIDI output that reproduces deadmau5‑like arpeggios but sometimes hits off‑key notes; model tuning, larger datasets (130k files) or longer training reduce artifacts. Code is being open‑sourced; copyright questions remain.
AI-written summary · about
chapters (12)
- 0:00Introduction and speakers
- 1:04Meet the presenters
- 3:57What is machine learning
- 7:05Neural networks explained
- 13:52Why LSTM for music
- 14:15What is MIDI
- 17:08Finding and preparing data
- 20:02Code demo in Ruby
- 21:03Generating music with TensorFlow
- 23:33Live music generation demo
- 30:54Summary and future plans
- 32:19Q&A session
related talks
46mMusic Generation with Magenta: Using Machine Learning in
41mMusic and Machine Learning
52mAI Lofi HipHop - How (Not) To Generate AI Music
36mMusic and Art with ML
45mAI Lofi HipHop - How Not To Generate AI-Music
21mMusical Deep Neural Networks in the Browser
47mExpressive AI-Generated Music With Google's Performance RNN
28mMusic and AI in the Browser with TensorFlow js and Magenta js