A quantitative model of cross-speaker pitch
range equivalence
D. R. Ladd, Edinburgh University
This paper is a preliminary presentation of
a quantitative model that
predicts the relative F0 level of tone
phonemes produced by different
speakers.
F0 values are modeled in terms of three parameters, "level"
and
"span" (speaker-specific
parameters referring to the overall level of the
voice and the span or range of frequencies
used by the speaker) and "tone"
(phonologically specified levels within the
speaker's overall range, for
e.g. H, M and L tone). The basic form of the model is F0 = Fr +
S^T,
where Fr (reference frequency) is a
reference level at the bottom of the
speaker's range, T can take one of a
restricted range of values
corresponding to the tones of the language
(e.g. L tone has T = 1 and H
tone has T = approximately 1.45), and S
(span) is a variable that models
the F0 difference between H and L
tone. This model is shown to give a
very close fit to mean tone values for 5
speakers of a 4-level-tone
language (Mambila) and 4 speakers of a
3-level-tone language (Yoruba). For
Yoruba, the model also clearly shows that
overall raising of pitch range
in questions is a change in overall level
(Fr) rather than span (S), and
that raising of H tones before L tones is
clearly a change in span (S)
rather than in the linguistic tonal specification (T).