evau

Experimental playground for a thesis on music composition and reconstruction with evolutionary algorithms. The project keeps the evolutionary machinery relatively stable and varies the fitness function.

Experiments

exp_01_baseline.py

Renders a small showcase for the shared synth vocabulary. It exists to showcase the renderer and timbral palette.

exp_02_audio_global_similarity.py

Whole-piece CLAP embedding similarity against a target audio file. It compares one rendered candidate against one target clip with cosine similarity in CLAP's embedding space

exp_03_audio_chunk_distance.py

Chunk-wise CLAP-based Fréchet-style distance against a target recording.

Important caveat:

• This is an adaptation of a dataset-level distribution metric to a single-piece setting.
• It treats fixed-size chunks from one target recording as a local embedding distribution.

exp_05_symbolic_stats.py

Symbolic evaluator inspired by Manaris-style Zipfian structure.

Important caveat:

• This is not a direct corpus-based reproduction of the original paper.
• The current version trains synthetic critics on Zipf-like token distributions.

exp_06_clamp.py

Text-conditioned symbolic generation using bundled CLaMP 3 similarity as the fitness signal.