Phonetic data: speaker normalization + speaker as random factor = overaccounting for individual variation?

[azlergarcia]

Hi! I've got an off-topic question to do with how normalizing data by participant/speaker in addition to including participant/speaker as a random factor might influence model estimates. In phonetics, data like vowel formants (for instance, F1 and F2) tend to be normalized to account for the differences caused by speakers' different mouth sizes, amongst others (http://lingtools.uoregon.edu/norm/about_normalization1.php). This is particularly tied to sex as well, usually resulting in lower F1 and lower F2 values for females than males. Some normalization methods include Bark distances, log transformations, individual log means, z-scoring... (see link for more if interested: https://marissabarlaz.github.io/portfolio/vowelnormalization/#types-of-vowel-normalization).

Most if not all LMMs using such data also include speaker as a random factor, but I'm worried this might result in over-accounting for individual/speaker variability since you already got rid of much of it by normalizing your data. What would be your take on it? Do you find speaker-normalizing your data still that necessary if randomizing speaker will do a similar work anyway? Hope the question's not too field-specific.

Thanks in advance!