Digital speech analysis using sequential estimation techniques

Two new digital speech analysis methods for sequentially identifying the coefficients of the linear prediction model are presented; the methods are based on the stochastic approximation and Kalman filter sequential estimation algorithms. Speech synthesized using the predictor coefficients identified by the Kalman filter algorithm is highly intelligible and comparable in quality to that obtained by the autocorrelation and covariance methods. Speech synthesized using predictor coefficients identified by the stochastic approximation algorithm is also highly intelligible but of lower quality. The analysis and synthesis procedures use hand-picked pitch and voiced/unvoiced information, and the predictor coefficients are converted to PARCOR coefficients for checking stability and transmission to the receiver. The sequential techniques are shown to be real-time feasible and closely related to the more familiar autocorrelation and covariance methods for speech analysis.