B. Y. Reis
C. H. Coker
R. A. Kubli
AT&T Bell Labs., Murray Hill, NJ 07974
Speech synthesis and coding schemes commonly incorporate a random noise component during voicing. It is hard to isolate this noise in recorded speech because of the inaccuracy of characterizing a modulated random process from a few nonidentical periods with a strong periodic component. Direct measurement on humans poses many clinical difficulties. Computational hydrodynamic modeling is difficult, and reliable modeling of turbulence is an unsolved problem. Here, direct steady-state measurements are made on a detailed, life-size mechanical model. A molded rubber model of the vocal cords is driven synchronously with the sampling clock so that precise repetitions of pitch periods are generated. The random component of the recording is isolated by subtracting away the periodic component (obtained by ensemble averaging across 1000 pitch periods). The noise power and its spectrum at each point in the pitch period are estimated by ensemble averaging the autocorrelations of the noise. The noise results are well explained in terms of the directly measured transglottal pressure and glottal area. In high-passed human speech, bursts of noise appear to overlap the closed-glottis interval. To explain this, it is necessary to include transit time for glottal air bursts to reach an obstruction 2--3 cm downstream. [sup a)]Also a student at MIT.