Inverse filtering and simultaneous model matching

A new method for automatic inverse filtering is presented that is derived from the “discrete all-pole modeling” (DAP) algorithm. In the new method, the parameters of a glottal flow model (the LF model) are iteratively optimized to match the inverse filtered acoustic speech wave. At the same time, the particular configuration of the LF model enters into the inverse filtering process in each iteration step to obtain an improved estimate of the glottal flow. By applying standard multidimensional optimization procedures, the model parameters are modified to finally yield the optimal description of the glottal flow. The initializing is performed on the basis of an exhaustive search. Measures describing the voice source are calculated from the matched LF model. The complete procedure is fully automatic. The performance of the method is illustrated for one natural utterance where phonation was varied between hypofunctional and pressed.