Nov 22
======

Iskarous/etal:2013
------------------

"MI in the German stop place of articulation dataset is about half the
magnitude in the German manner dataset. There are several reasons for
this discrepancy: ... (3) each MI is based on 74 tokens in the first
set, but 30 in the second set"
-> I'm confused about the role that token count plays in MI. Is there
a specific reason to think that the number of tokens would affect MI
in a particular direction?

"[DAC] quantification works by first measuring, in physical units, for
each segment, the degree to which the segment influences and is
influenced by its surrounding segments. Numbers are then assigned to
each segment by the investigator that reflect the relative resistance
of the segments." / "DAC subjectively assigns each segment a number"
-> I'm curious about what physical measurements are being taken and
how this results in a subjective number. Is DAC essentially an ordinal
scale of the sounds in a particular experiment?

FIG 5 (Catalan data) -> The vertical (a) component of Bld y is a
perfect example of how much interspeaker variation there can be. The
circle speaker shows essentially no difference in MI between the two
sounds whereas the others show quite a large difference. It makes me
wonder how much we can really generalize from such a small sample.

The sequences in German were produced in the phrase "Ich habe geCVCe
gesagt." How do we know that the coarticulatory effects from the rest
of the word "geCVCe" (especially, the /g/ sound) didn't affect the
articulation of the tested consonants in a way that could impact the
study results?

The MI approach discussed in the paper examines coarticulatory overlap
by assessing the predictability of a specific articulator's position
during the production of a consonant based on a following vowel. While
this offers insights concerning typical patterns regarding
coarticulation, the conclusions drawn seem to heavily generalize
across speakers. As already highlighted in the paper's conclusion, it
might thus be reasonable to investigate the occurrence of individual
differences in coarticulation patterns, which may not necessarily
adhere to the predicted general tendencies.

While the MI scale can be used to analyze speech productions including
coarticulatory patterns, it does not directly address speech
perception. How might coarticulatory overlap measured by MI relate to
listeners' perception of invariance and coarticulation in speech
sounds?

Are there any possible demerits in prioritizing the tongue over other
articulators in degree of articulatory constraint (DAC) models?

Relating to terminology: what is the exact distinction between
invariance and coarticulation resistance?

How can these concepts of coarticulatory resistance and invariance be
applied to coarticulation in sign languages?  For instance, rather
than constraints on vocal tract articulators for each speech segment,
there would be constraints on the sign articulators (fingers, arms,
available space, etc.)  Would the "spatiotemporal measurements" used
in this approach also apply to the articulators in sign languages?
Would the 3 parameters of sign languages (hand configuration, place of
articulation, and movement) also affect coarticulation resistance in
the same way that POA and MOA affect coarticulation resistance in
spoken languages?  Would locus equations still work by replacing F2
values with movement or a different sign parameter?

Why are F2 slopes commonly used to measure coarticulation resistance
in locus equations? Why are F2 slopes preferred over F1 and F3?

One of the German corpora was recorded at both comfortable and loud
speech volumes. Could the loud speech have skewed the results? Does
intensity affect coarticulation?

MI is shown as an effective way to measure coarticulation. Could MI be
useful in clinical research to track disease progression in conditions
where coarticulation deteriorates over time?