# AI MusiMuse # 007_MUSIC_REPRESENTATION.md Version 1.0 --- # Purpose This document defines the internal representation of music used by AI MusiMuse. The system does not learn directly from raw audio. Instead, every composition is transformed into multiple abstraction layers. Each layer answers a different musical question. --- # Representation Layers Every track is represented as seven independent layers. ``` Track │ ├── Metadata ├── Audio ├── MusicDNA ├── Structure ├── Timeline ├── Composition └── Style ``` Every higher layer depends on lower layers but never modifies them. --- # Layer 1 — Metadata Metadata describes the recording itself. Examples - title - artist - album - duration - sample rate - channels - filename Metadata is never used directly for generation. --- # Layer 2 — Audio Audio is the original waveform. Usually ``` WAV or FLAC ``` The waveform is immutable. It is used only by analyzers. --- # Layer 3 — MusicDNA MusicDNA describes "What does this music sound like?" Examples ``` tempo key spectral centroid dynamic range rhythm density tonal stability ``` MusicDNA contains global characteristics. MusicDNA does not know where musical events happen. --- # Layer 4 — Structure Structure answers "What parts does this composition contain?" Typical sections ``` Intro Theme Development Bridge Break Climax Outro ``` Each section has - start - end - duration - confidence --- # Layer 5 — Timeline Timeline answers "What happens over time?" Timeline consists of ordered musical events. Example ``` 00:00 Pad starts 00:18 Noise layer appears 00:42 Bass enters 01:05 Melody begins 02:30 Harmony changes 03:40 Break 04:05 Theme returns 05:10 Outro ``` Timeline is deterministic. The same song always produces the same timeline. --- # Musical Event Every timeline entry is represented as ``` Event time duration type confidence parameters ``` Example ``` BassEnter time = 42.3 confidence = 0.96 ``` --- # Event Categories Events belong to different categories. Examples ``` Harmony Rhythm Texture Dynamics Energy Transition Instrument Effect ``` The list may grow over time. --- # Layer 6 — Composition Composition answers "How does the musical idea evolve?" Unlike Timeline, Composition is semantic. Example ``` Atmosphere ↓ Expectation ↓ Development ↓ Release ↓ Resolution ``` This layer describes musical intention rather than acoustic events. --- # Composition Graph The composition is represented as a directed graph. Example ``` Intro ↓ Atmosphere ↓ Main Theme ↓ Development ↓ Break ↓ Return ↓ Ending ``` Nodes represent ideas. Edges represent transitions. This graph becomes the primary learning object for future generators. --- # Layer 7 — Style Style is extracted from the complete library. It is never computed from a single track. Examples ``` preferred BPM preferred dynamics preferred harmonic movement preferred structures preferred energy evolution preferred transitions ``` Style describes the author's musical language. --- # Why Multiple Layers? Different AI components require different information. MusicDNA ↓ describes sound Structure ↓ describes organization Timeline ↓ describes events Composition ↓ describes ideas Style ↓ describes habits --- # Generator Input The future Composer never receives raw audio. Instead it receives ``` Style + Composition Graph + Timeline Templates + MusicDNA Statistics + Random Seed ``` --- # Generator Output The generator produces ``` Composition Graph ↓ Timeline ↓ Musical Events ↓ Rendering ↓ Audio ``` Generation is therefore hierarchical. --- # Learning Strategy The system learns progressively. Stage 1 Understand sound. Stage 2 Understand structure. Stage 3 Understand musical ideas. Stage 4 Understand author style. Stage 5 Compose new music. --- # Design Principles Raw audio is immutable. MusicDNA is immutable. Timeline is deterministic. Composition is semantic. Style is extracted from the complete library. Generation consumes abstractions. Rendering is independent from composition. --- # Long-Term Goal AI MusiMuse should eventually understand not only "What sounds are present?" but also "Why does this composition evolve the way it does?" The generated music should therefore resemble the author's musical thinking rather than copying existing tracks.