Using a voice range profile (VRP) based averaging system, which sorts spectra into bins by Fo and SPL, the spectrum dependency on Fo and SPL was isolated. Next, a principal components analysis (PCA) was performed. This revealed two distinct frequency regions with different dependencies on SPL: a low- and a high-frequency regime that separate at about 1 kHz.
The existence of asymmetrical spectral dependencies that the PCA analysis suggests, remains a puzzling finding. The fact that spectrum levels have a multiplicative link that is a dynamic variable could imply that the amount of vocal tract filtering varies with the sound level itself. Another explanation could be that different spectral sections link to disparate aspects of the voice source waveform, but then the fact that voice source related features can remain so detached over such a large dynamic range is unexpected.
Separate slope measures for the two frequency regions were defined and mapped over the VRP. The resulting patterns are discussed and compared to those of other recorded voice quality measures. Results of modeling experiments that could explain the mechanisms involved are also shown. Feedback on the relevance of the possible interpretations is appreciated.