Re: Stop consonant identification based on initial spectra?

[Abeer Alwan <abeera@xxxxxxxxxxxxx>]

Greetings,

If I remember correctly, the Blumestein and Stevens papers showed that initial
spectra bursts account for about 85% of (correct) place responses.

Diane Kewley-Port (1983) modified the fixed time window of
Stevens and Blumstein with a running spectral display
to include the initial 40 ms of the vowel onset as well. She was able to
categorize place of articulation for stops at 88% correct (automatically not
perceptually).


Even though the initial burst carries important perceptual cues for place
of articulation, the burst is quite vulnerable to background noise. In noise,
formant transitions play a larger role in identifying place although the
importance of the transitions vs. the burst varies with the masker
shape (in white noise, the burst is masked at a higher SNR than
in speech-shaped noise; in both cases, it is masked at a higher SNR than
the formants).


W. Chen and A. Alwan, "Perception of the Place of Articulation Feature for
Plosives and Fricatives in Noise," in Proc. ICPhS, Barcelona, August, 2003.

James J. Hant and Abeer Alwan, "A Psychoacoustic-Masking Model to Predict the
Perception of Speech-Like Stimuli in Noise," Speech Communication, Vol. 40, May
2003, pp. 291-313.

Regards,
Abeer


Abeer Alwan, Prof.
Dept. of Electrical Engineering, UCLA
email: alwan@xxxxxxxxxxx        http://www.icsl.ucla.edu/~spapl