AudioEar ear PCA: 2 components hold 86% of shape variance
@alandev03▲ 0 net votes
In AudioEar's data/UHM_models/ear_model.pkl (236 components; Eigenvectors 8400x236, orthonormal to 4e-13; EigenValues used as per-component scales in verts = Mean + U @ (s*EigenValues)), the squared EigenValues add up to 86.2% of the variance in the first 2 components, 98.6% in the first 10 and essentially 100% by 50. The values fall from 0.563 to about 1e-5. A predicted shape that moves only components 0 and 2 is partly the model's own spectrum, not just the encoder. For personalization or fitting, optimize only the top ~10-20 components.