James D. Miller
M. J. Gottfried
D. J. Meyer
E. J. Frederick
Central Inst. for the Deaf, 818 S. Euclid, St. Louis, MO 63110
Values for the fundamental frequency and F1, F2, and F3 were obtained for a corpus of 1248 vocalic nuclei from CVCs (26 phonemically different vocalic nucleix4 speakersx2 stress rate conditionsx3 consonantal contextsx2 repetitions of each token) at 25 equally spaced times within each vocalic nucleus. This corpus included monopthongs /i, (small capital eye), (eh), (ae ligature), a, (inverted vee), (open oh), u, u/, diphthongs /au, ai, ei, ou, (open oh)i/, rhotic vowels /ir, er, ar, or, (hooked backward eh)/, and vowels followed by /(ell with tilde)/---/i(ell with tilde), i(ell with tilde), (eh)(ell with tilde), a(ell with tilde), (inverted vee)(ell with tilde), (open oh)(ell with tilde), u(ell with tilde)/. These were spoken in /b_d/, /d_d/, and /g_d/ contexts. A standard back-propagation neural network with one hidden layer was trained to identify which of these 26 vocalic nuclei was spoken. Input data for the neural network was presented as (1) an auditory-perceptual space using (x'y'z') or (2) other implementations of the fundamental frequency and the first three formants. Preliminary results indicate that the neural network is able to identify these nuclei on the basis of acoustic parameters alone. [Work supported by NIDCD.]