In SVD-factorised WavLM-Large layer-6 features, individual content and speaker dimensions carry intensity, formants, voicing, pitch and gender, and shifting those dimensions controls pitch and intensity in resynthesis

- Janse van Rensburg and Kamper factorise WavLM-Large layer-6 features into a shared content subspace and speaker-specific transformations (rank 64, SVD) and find leading content dimensions capture intensity (R^2 = 0.72, dimension 2), only weakly the formants (R^2 0.136-0.285), and voicing, with pitch in a later dimension (12, R^2 = 0.375), while the highest-variance speaker (PCA) dimension is strongly associated with pitch and gender (R^2 >= 0.58 for jitter and pitch); adding multiples of a dimension and vocoding changes the correlated characteristic (formants could not be controlled through the speaker space), and joint content-plus-speaker interventions give finer control. Correlational dimension analysis plus causal resynthesis interventions; causal-direction wording only. [janse-van-rensburg-kamper-2026-interpreting-content-and-speaker-characteristics-in-factorised-self-su]