| |
Abstract:
Three speakers of French (two males and one female) were the
subjects of an MRI analysis of the vocal tract during the
production of sustained isolated French vowels. From a 3D
reconstruction of the vocal tract, area functions were determined
for each vowel, and the corresponding formant values were
computed with a harmonic model of the vocal tract. Using the
computation of the relative formants shifts associated to
localized area function changes, the main affiliations between
formants and cavities were found. Based on these functions, as
well as on geometrical criteria, each area function was divided
into laryngeal cavity, back cavity, constriction cavity, and
front cavity. One of the male speakers was chosen as "target
speaker". Then, a resonance based approach was used to bring the
formant patterns of the two remaining speakers (called test
speakers) closer to the ones of the target speaker. For that, a
transformation coefficient was calculated, for each vocal tract
cavity (except for the larynx), as the ratio of the formants
corresponding to the lowest resonance (except Helmholtz) of the
cavity of the test speaker and of the target speaker
respectively. Then, the [a i u] MRI area functions of the test
speaker were distorted in length, cavity by cavity, according to
these coefficients. The comparison of the accuracy of these
predictions with the actual differences measured between speakers
shows that our method is quite powerful if the lowest formant
frequencies are either half-wavelength or quarter-wavelength
resonances of the main vocal tract cavities, and if the nature of
the resonance modes associated to the formants are the same for
both subjects. Differences between speakers in the coupling of
the cavities does not seem to be a major factor of the
inter-speaker variability.
|
