Gabor Latent Refactorization reduces structural bottlenecks in end-to-end audio models

Researchers demonstrate that state-of-the-art strided convolutional encoders impose two structural bottlenecks on access to time-frequency-localized primitives: collapsing them into alias equivalence classes and limiting frequency resolution. These bottlenecks result in collapse rates of 31-35% and filter bandwidths 10-35x above the theoretical resolution bound.

• Strided convolutional encoders collapse primitives into alias equivalence classes, establishing a bound on representational capacity.
• Encoders limit frequency resolution available to learned filters, restricting separability.
• Gabor Latent Refactorization (GLRF) is introduced as a lightweight post-hoc intervention that re-expresses encoder latents in a frequency-localized basis.
• GLRF reduces filter bandwidths from 10-35x to 1.5-3x of the theoretical resolution bound while preserving reconstruction fidelity.

This intervention recovers access to frequency-localized primitives, improving steerability and interpretability without requiring retraining.