Expressive AI-Generated Music With Google's Performance RNN
summary
Magenta releases a working Performance RNN — an LSTM trained on ~20–30k solo piano MIDI performances that predicts note on/off, velocity and real-time clock moves in milliseconds, producing noticeably human expressive timing versus score-only models. The project runs in TensorFlow and ships samples and code. Related work trains a QuickDraw sketch RNN encoder‑decoder into a latent embedding that lets researchers interpolate, detect outliers and decode many believable variants; demos and datasets are open source.
AI-written summary · about
chapters (13)
- 0:00Intro
- 1:37Welcome Doug
- 3:11Dougs background
- 8:22Capturing sensory motor control
- 13:01Music style transfer
- 15:07Performance RNN
- 17:42Training
- 19:41Network Architecture
- 21:18Google Draw
- 24:28Decoder
- 34:06Creating a new art form
- 38:16Magenta
- 42:17Nerd Alert
related talks
31mDoug Eck, Research Scientist, Google
28mMusic and AI in the Browser with TensorFlow js and Magenta js
21mMusical Deep Neural Networks in the Browser
46mMusic Generation with Magenta: Using Machine Learning in
41mCreating Intelligent Music Instruments with Machine Learning
34mModelling emotional nuance of musical performance
41mMusic and Machine Learning
24mCreating an AI Musician with JavaScript