You can not select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
52 lines
2.7 KiB
52 lines
2.7 KiB
The Klatt formant synthesizer
|
|
|
|
ample, if a formant frequency is halved, amplitudes of all higher for-
|
|
|
|
mants are decreased by 12 dB, i.e. (.5)?, as shown in part (c) of
|
|
Figure 12-11.
|
|
|
|
5. The frequencies of two adjacent formants cannot come any closer
|
|
|
|
than about 200 Hz because of coupling between the natural modes of
|
|
|
|
- the vocal tract. However, if two formants approach each other by
|
|
|
|
about this amount, both formant peaks are increased by an additional
|
|
|
|
3 to 6 dB, as shown in part (d) of Figure 12-11.
|
|
|
|
The amplitudes of the formant peaks generated by the parallel vocal tract
|
|
model have been constrained so that, if Al to A5 are all set to 60 dB, the transfer
|
|
function will approximate that found in the cascade model. This is accomplished
|
|
by: 1) adjusting the gain of the higher frequency formants to take into account
|
|
frequency changes in lower formants (since a higher formant rides on the skirts of
|
|
the transfer function of all lower formants in a cascade model (Fant, 1960)), 2) in-
|
|
corporating rules to cause formant amplitudes to increase whenever two formant
|
|
frequencies come into proximity, and 3) using a first difference calculation to
|
|
remove low-frequency energy from the higher formants; this energy would other-
|
|
wise distort the spectrum in the region of F1 during the synthesis of some vowels
|
|
(Holmes, 1973).
|
|
|
|
- 'The magnitude of the vocal tract transfer functions of the cascade and parallel
|
|
vocal tract models are compared in Figure 12-12 for several vowels. The match is
|
|
quite good in the vicinity of formant peaks, but the parallel model introduces trans-
|
|
fer function zeros (notches) in the spectrum between formant peaks. The notches
|
|
are of relatively little perceptual importance because energy in the formant peak
|
|
adjacent to the notch on the low-frequency side tends to mask the detectability of a
|
|
spectral notch (Gauffin and Sundberg, 1974).
|
|
|
|
Many early parallel synthesizers were programmed to add together formant
|
|
outputs without filtering out the energy at low frequencies from resonators other
|
|
than F1. In other cases, formant outputs were combined in alternating signs. The
|
|
deleterious effects of these choices are illustrated in Figure 12-13. Some vowel
|
|
spectra are poorly modeled in both of these parallel methods of synthesis. The per-
|
|
ceptual degradation is less in the alternating sign case because spectral notches are
|
|
less perceptible than energy-fill in a spectral valley between two formants. Com-
|
|
parison of Figure 12-12 and Figure 12-13 indicates that our parallel configuration
|
|
is better than either of those shown in Figure 12-13.
|
|
|
|
A nasal formant resonator RNP appears in the parallel branch to assist in the
|
|
approximation of nasal murmurs and vowel nasalization when the cascade branch
|
|
|
|
147
|