The top principal direction of utterance-averaged self-supervised speech features carries pitch and gender and other directions carry intensity, noise and formant structure, and shifting WavLM-Large layer-6 directions changes the correlated characteristic with little leakage

- Janse van Rensburg, van Niekerk and Kamper run PCA on utterance-averaged Wav2Vec2, HuBERT and WavLM features and report that the highest-variance principal direction encodes pitch and associated characteristics such as gender and jitter, with other directions correlating with intensity, noise level, the second formant and high-frequency content; in WavLM-Large layer 6 (the model analysed in depth) shifting a principal direction and vocoding changes the correlated characteristic for pitch (which plateaus at large shifts), intensity, F2, F3, spectral rolloff and ZCR but not for F1, jitter, shimmer or HNR, and the changes are largely isolated (varying the pitch direction to move pitch between 100 and 300 Hz changes intensity by only about 5 dB). Only WavLM-Large layer 6 is tagged because the layer-wise comparison covers further models and layers without exact checkpoints named. [janse-van-rensburg-etal-2026-interpreting-speaker-characteristics-in-the-dimensions-of-self-supervi]