[From: Bruce Nevin (Mon 931115 08:46:31 EST)]
( Bill Powers [date not retained]) --
Some time back, Bill, you mentioned that what you saw as formants rose
and fell with voice pitch. I said formants don't rise and fall with
pitch, the fundamental and its harmonics do. The formants reflect a
filtering function of the vocal tract. I suggested that some refinement
of your setup might be needed.
I remember an explanation now of something that sounds very much like
what you reported seeing. With a narrow-pass filter (say. 50Hz), the
many individual harmonics are visible in a sound spectrogram, and
formants are visible only in terms of greater intensity of harmonics not
damped in the vocal tract. With a wider bandwidth (say, 650Hz), formants
are much more distinct, no longer overwhelmed by the detail of harmonics.
In both cases, the display gain must be set to enhance the contrast in
amplitudes, so that the damped frequencies drop out of the display. For
the tradeoffs here, see Lieberman & Blumstein, their chapter on speech
analysis. With the narrow bandwidth, a harmonic that is near the top of
a formant fades as pitch rises, lifting that harmonic out of the formant.
Because formants are perceived even when voice pitch is high and no
harmonic happens to fall within the formant, it seems to me that the
damped harmonics may be part of formant perception. Then there is the
question, what soprano singers actually produce when they think they are
producing vowels but voice pitch is higher than F1 or perhaps even F2.
Interesting bit of trivia: lower-pitch voices are more readily
intelligible because there are more data points -- harmonics -- per
formant. This is in acoustically favorable conditions; reverberation can
complicate things, I believe.
Bruce Nevin
bn@bbn.com