Deconvolution and vocal-tract parameter estimation of speech signals by higher-order statistics based inverse filters

The authors propose a two-step method for deconvolution and vocal-tract parameter estimation of (non-Gaussian) voiced speech signals. In the first step, the driving input (a non-Gaussian pseudo-periodic positive pulse train) to the vocal-tract filter which can be nonminimum-phase is estimated from speech data by a higher-order statistics (HOS) based inverse filter. In the second step, autoregressive moving average (ARMA) parameters of the vocal-tract filter are estimated with the estimated input and speech data by a prediction error system identification method (an input-output system identification method). Finally, some experimental results with real speech data are provided.<>