AGR: A Rehearsal-to-Performance Workflow for Programming Audio Gestures in Max
summary
AGR is a Max/FluCoMa workflow that turns rehearsal examples into live-controllable actions by keeping vocabulary design, capture, feature extraction, model training and routing inside one patch. It implements fixed-length padding without time warping and a local frequency-deviation representation from sinusoidal peaks, outputs OSC/MIDI or gesture-bound audio, and runs per-instrument models (clarinet, flute, trumpet, trombone) including a 12-model ensemble deployment. Post-trigger latency sits around 20–30 ms; larger feature vectors raise runtime, and accuracy remains piece-specific so each work needs its own captured vocabulary.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 1:20What is Max and audio gesture recognition
- 3:04George Lewis and Voyager inspiration
- 5:12Live demo of audio gesture recognition
- 7:24Five-step workflow for AGR
- 9:32AGR patch for non-programmers
- 13:26Feature extraction and classifier
- 15:55Three AGR use cases
- 17:43Voyager Plus at Washington festival
- 20:17Live performance excerpt
- 21:00AGR limitations and future work
- 23:51Q&A
related talks
39mSonic Earth - An Algorithmic Performance Approach to Real-Time Environmental Sonification
43mA Progressive-Adaptive Music Generator (PAMG): An Approach to Interactive Procedural(…)
18mPlaying Fast and Loose with Music Recognition
33mContext-Aware Masking for Controlled Music Generation in MIDI
26mWhy computer science matters to musicians
39mAdaptive Music for the Web
11mMachine Learning of Personal Gesture Variation in Music Conducting
20mAccelerometers as an Instrument