Speech Data Rate Reduction Part I: Applicability of Modern Estimation Theory

Efficient coding of continuous speech signals for digital representation has attracted much interest in recent years. The underlying aim of efficient coding methods is to reduce the channel capacity required to represent a signal to meet a specific reconstruction fidelity criterion. To achieve this objective, modern speech data compression techniques rely on two very similar procedures. One procedure uses predictive deconvolution which subtracts from the current signal value that portion which can be predicted from its past and thus removes redundancy in the speech by removing sequential correlation. The signal thus requires fewer bits for equivalent quantization error. The second procedure involves identification of a complete mathematical model of the speech producing mechanism. This involves determination of the characteristics of the source that drives this transfer function. Data reduction is again achieved since the rate of change of the parameters of the speech model is much smaller than the rate of change of the speech waveform. This paper develops these data reduction procedures in terms of modern estimation theory, specifically a Kalman filter model, and illustrates the utility of this model as an analysis tool by means of an example based on a uniform tube which provides a qualitative assessment of the potential of the technique for application to real speech signals.