A priori digital speech analysis

Abstract: This paper describes a priori digital speech analysis. This method takes advantage of the quasi-periodic nature of voice speech sounds to reduce the amount of computation needed to obtain the analysis parameters. The vocal tract is modeled as an all-pole digital filter (predictive coding) and pitch is detected using a modified autocorrelation method. Since voiced speech is quasi-periodic, its waveform is characterized by a high degree of redundancy from pitch period to pitch period. For these regions, the parameters which describe the mathematical voice model are relatively constant. Specifically, the predictor coefficients and the pitch period vary slowly enough in these regions to allow their preceding values to be used as additional a priori information when calculating their present values. Estimates of the predictor coefficients are obtained using a priori least squares. By incorporating previous parameter values as a priori information, less deata (a smaller window size) are required to estimate the current parameters. Also, in addition to the savings obtained from reduced arithmetic operations, it is shown that a priori least squares generates a smoother set of predictor coefficients than those generated when a priori methods are ingnord. Estimates of the pitch period are obtained using a priori pitch detection. It is shown that pitch can be estimated more efficiently (fewer real multiplies) when the previous pitch period, the predictor coefficients, and the interperiod waveform correlation are incorporated in the estimation.