Localization of Speaker Timbre Across Frequency Bands and SSL Layers
October 22, 2026
Probing reveals that speaker identity is distributed across speech signals, with 1-8 kHz magnitude and vocal-tract spectral envelopes dominating human perception. While shallow SSL layers contain compact speaker information, bands above 8 kHz remain primarily discriminative for machines.
HOW THIS AFFECTS YOU
●
researcherThis identifies specific frequency bands and SSL layers to target for better voice conversion and cloning models.
●
designerYou can better understand how perceived voice similarity relates to specific spectral components.