Joint estimation of excitation and vocal tract response

Many speech analysis techniques attempt to deconvolve the speech waveform into an excitation component and a response function. The standard approach is to estimate the vocal tract model parameters first and then the excitation signal from the residual errors (or directly from a bandlimited version of the original speech signal). This paper presents an overview of a new approach for simultaneously identifying the system model parameters and detecting the unobserved random pulse-type inputs. A key component of this procedure is the application of a new decision-directed algorithm to estimate the period of the pitch pulse process. This decision-directed algorithm incorporates an exact, recursive estimator to compute the rate of a discrete-time point process used to characterize the arrival times of the pitch pulse process.