Piano Genie
summary
Piano Genie maps an 88-key piano to an eight-button controller by training a recurrent neural network autoencoder with a discrete bottleneck: an encoder learns button sequences from a few hundred piano performances and a decoder reconstructs notes in real time. Regularization enforces melodic contour so higher buttons yield higher pitches. A web demo plays improvised outputs (not reproducible twice) and a user study shows Piano Genie outperforms a scale mapping and a performance RNN on enjoyment and perceived control.
AI-written summary · about
chapters (5)
- 0:00The vision for music AI
- 2:58Introducing Piano Genie
- 5:30Auto-encoding approach
- 8:01User evaluation and outlook
- 10:48Q&A and future developments
related talks
34mModelling emotional nuance of musical performance
47mExpressive AI-Generated Music With Google's Performance RNN
41mCreating Intelligent Music Instruments with Machine Learning
01h 04mKeynote 3. Interactive generative models for music
32mGetting started with generative AI with AWS DeepComposer
05mInteractive Exploration-Exploitation Balancing for Generative Melody Composition
33mContext-Aware Masking for Controlled Music Generation in MIDI
24mPlay Them Off, Keyboard App: Building a virtual piano in JavaScript