Search for unexplored effects in speech production

Speech coders invariably spend about 1/3 of their bits replicating a number of effects collectively termed "excitation". The need for so much bandwidth stems from two causes. The frame rate must be relatively high, because transient changes must be resolved. The data needed for each frame is high because unpredictable broad-band components must be reproduced. Here we discuss three projects to learn more about these elusive aspects of speech. One project models the transient behavior. Two others seek to characterize stochastic processes that accompany periodic vibration in voiced sounds. 1. Sequential effects in excitation Figure 1 (a) shows a waveform of the beginning of the word heed. Fig. 1 (b) is probable glottal behavior based on data for a similar utterance — word-initial /h&/. The instrumentation is photoglotography (PGG), also known as transillumination. A fiberoptic light source is inserted through the nose, and positioned above the glottis. In a darkened room, the outline of the glottis can be seen on the lower larynx and pharynx. A light sensor "watching" about 5 cm in this region gets a reasonably linear response with glottal area. Fig. 1 (c) is Linear Prediction residual for the waveform in (a). Height of the sharp impulses reflect abruptness of glottal closure. In Fig. 1, the glottis is open for /h/; then closes for the vowel. Glottal vibration starts shortly before glottal "rest area" gets roughly to zero. In optical evidence, glottal area seems to reach a rather stable state perhaps 25 msec after vibration begins. However, the waveform and prediction residual show substantial change for another 60 msec. Something in the larynx is still changing. But what? Here is offered a physiological explanation. A model is developed and its coefficients are tuned from available data. Although the model was suggested by a proposed explanation, it does not depend on details of that proposal. The model is anchored only in the notion that measurable glottal area and LP residual derive from a single variable, and conversely that behavior of that variable can be inferred from glottal instrumentation and LP data. Fig. 2 is a sketch of the glottal mechanism. The L-shaped structures shaded gray are the arytenoid cartilages. The vocal cords attach to the tips of these cartilages, the vocal processes. The glottis is closed by tensing several muscles — principal among them, the exterior thyro-arytenoids. As tension increases, the arytenoids rotate and the vocal cords move together. At some point, however, the opposing arytenoid tips press together, and can move no further. But there is a mechanism for the thyro-arytenoids to have effect after arytenoid motion is blocked. The cords are pushed apart in the middle by air pressure, and by stretch forces tending to restore the breathing state. The exterior thyro-arytenoids are curved and push inward through a draw-string action. Increasing tension, after the arytenoids block, pushes the cords in at the sides. This has a modest effect on glottal area. But a more profound effect is to alter the geometry of glottal closure. Fig. 3 shows the progress of glottal closure for three adjustments. The lower left is a magnified view of glottal area vs. time for conditions a, b and c; the lower middle shows glottal area derivative, and lower right, the spectral consequences. Condition a is a partially open glottis as might occur in /z/ (exaggerated). The arytenoid tips are separated a small distance, making the vocal-cord configuration somewhat triangular. During vibration, vocal-cord closure begins at the front and progresses toward the arytenoids in the back. This "zipper" action has a time course of glottal area closely resembling a decaying exponential. This contributes a 6 dB/octave spectrum down-turn at a frequency FT varying roughly inversely with measurable glottal area (see Fig. 5). (b) Figure 1 For the /hi/ in "heed"; (a) waveform; (b) approximate glottal (inferred from PGG for /h&/); (c) Linear Prediction residual. Arytenoid