# AI MusiMuse # 008_COMPOSITION_ARCHITECTURE.md Version 1.0 --- # Purpose Music is not a sequence of sounds. Music is a sequence of ideas. The goal of this layer is to transform low-level musical events into a semantic representation suitable for learning musical composition. --- # Hierarchy ``` Audio ↓ MusicDNA ↓ Timeline ↓ Composition ↓ Style ↓ Composer ``` Every layer increases abstraction. --- # Timeline Timeline describes "What happened." Example ``` 00:00 Pad 00:32 Bass 00:58 Lead 01:44 Break 02:20 Noise 03:10 Lead returns ``` Timeline is objective. It is directly observable. --- # Composition Composition describes "Why did it happen." Instead of events ``` Bass Lead Pad ``` Composition contains ``` Expectation Growth Release Silence Return Resolution ``` --- # Composition Units Every composition is divided into units. Example ``` Intro ↓ Build ↓ Development ↓ Break ↓ Recovery ↓ Ending ``` These units are semantic. They are independent from instruments. --- # Musical Ideas Every unit contains one or more musical ideas. Examples ``` Ambient atmosphere Slow tension Wide space Minimal rhythm Bright climax Calm ending ``` Ideas may repeat. Ideas may evolve. Ideas may disappear. --- # Musical Graph Instead of storing music as a list, the system stores relationships. Example ``` Atmosphere ↓ Main Theme ↓ Variation A ↓ Variation B ↓ Break ↓ Main Theme ↓ Ending ``` Every node represents one musical idea. --- # Graph Node Every node contains ``` id type start end duration energy density confidence ``` --- # Graph Edge Edges describe transitions. Example ``` Theme ↓ Development ``` Transition metadata ``` energy change tempo change harmonic change density change transition type ``` --- # Transition Types Examples ``` Fade Cut Build Drop Expansion Reduction Modulation Silence ``` --- # Composition Complexity Each composition receives statistics. Examples ``` graph depth branch count average transition length average section duration average energy slope repetition ratio development ratio ``` --- # Motifs The system attempts to identify recurring ideas. Example ``` Theme A ↓ Variation ↓ Theme A ↓ Variation ↓ Theme A ``` The motif itself becomes a reusable object. --- # Sections vs Ideas Sections are physical. Ideas are semantic. Example ``` Intro contains Atmosphere + Expectation ``` --- # Emotional Curve Every composition contains an emotional curve. Examples ``` Calm ↓ Curiosity ↓ Growth ↓ Peak ↓ Release ↓ Silence ``` This curve is independent from BPM. --- # Energy Curve Energy is represented separately. Example ``` 100% /\ 80% / \ 60%__/ \____ 40% 20% 0% ``` Energy and emotion are different concepts. --- # Density Curve Density describes How many musical elements are active. Example ``` Pads Bass Lead FX Percussion Noise ``` ↓ ``` 0 1 2 5 7 3 1 ``` --- # Instrument Independence Composition must never depend on ``` Piano Synth Violin Kick ``` Instead ``` Harmony Rhythm Texture Pulse Motion Space ``` The generator may later choose any instrument. --- # Composition Template Every song eventually becomes ``` Composition Graph + Timeline + MusicDNA + Style ``` This representation is sufficient for learning. --- # Learning Goal The AI should eventually understand ``` This song builds tension. ↓ Then creates silence. ↓ Then returns the original idea. ``` instead of merely learning ``` Bass Lead Pad Kick ``` --- # Future Composer The Composer generates Composition Graph ↓ Timeline ↓ Events ↓ Rendering ↓ Audio Generation therefore begins with ideas rather than notes. --- # Design Principles Composition is deterministic. Ideas are semantic. Graphs are immutable. Transitions are first-class objects. Rendering is independent. Composition never stores audio. --- # Long-Term Goal AI MusiMuse should eventually learn how musical ideas evolve, how they relate, and how they create emotional movement. Generated music should therefore resemble the author's compositional thinking, not simply imitate existing recordings.