Making machines that make music
summary
Models Carnatic melody from mined audio, phonetic-matches Indic titles, and deduplicates clips into a dataset, then converts audio to frequency melographs and MIDI. Uses Clojure/Overtone code to synthesize gamaka curves (Welch envelopes) and runs Markov chains at multiple temporal granularities (beat/gatti/sub-beat). Outcome: raga identification and automatic transcription work reliably, continuous-synthesis of ornamentation improves but higher-order sequence models fail to capture long-range context, and a repetition-based generator plus electronic percussion produces usable music.
AI-written summary · about
chapters (4)
related talks
41mMaking Machines that Make Music
41mMusic and Machine Learning
01h 04mKeynote 3. Interactive generative models for music
51mThe Merger of Machine Learning, Music and Programming
18mPlaying Fast and Loose with Music Recognition
43mProgramming Music with Overtone
42mAfrican Polyphony & Polyrhythm
54mIntegral scherzo: how machines compose and sing