Estimation of vocal-tract areas: The need for acoustical measurements

It is well known that the transfer function of a lossless vocal tract (much less that of a tract with losses) does not uniquely specify its area function. Further, the speech wave does not uniquely specify the transfer function. In spite of this proven nonuniqueness, several investigators during the past few years have proposed methods for estimating area functions from the speech wave. In this paper we will critically examine some of the tacit assumptions on which these proposals are based. The validity of the assumptions is not easily tested, nor is it feasible to estimate the effects of departures from them. For reliable recovery, therefore, we will argue for a renewed interest in the method based on the measured driving point impulse response at the lips. This method recovers the area function exactly under the plane wave approximation for a lossless vocal tract, as well as for tracts with certain types of distributed losses. We will briefly review this method and present evidence that departures from the basic assumptions do not drastically affect the reconstructed area functions. We will also point out the practical problems that still remain to be solved.