An Adaptive Predictor for Speech Encoding

Abstract

<p>In order to improve the performance of differential encoding systems, the encoding and decoding models have to change according to the speech waveform. The speech signal can be treated as quasi-stationary processes, which over a short period of time can be modelled by a certain set of parameters. Adaptive algorithms should be viewed as means of adjusting the system parameters.</p> <p>In this thesis, a 2.048 sec. long sentence has been studied by the Box-Jenkins time series procedure to determine the order of the linear prediction model and to investigate the need for adding moving-average terms. The algorithm suggested by Box-Jenkins for parameter estimation has been employed to update the parameters of the predictor of a prediction error coder each specific period of time.</p> <p>Since it is difficult to implement this algorithm on-line an alternative scheme has been studied. It is based on using the Box-Jenkins procedure to determine a suitable ARMA model and then updating the parameters of this model using a good on-line estimation algorithm. The applicability of the recursive least-squares and the stochastic approximation algorithms has been investigated. Stochastic approximation appears more promising as it takes less time for computation with an acceptable performance.</p> <p>As a result of this study, the addition of moving average terms to the predictor's model are shown to be necessary. But when Box-Jenkins' algorithm was tested with an ARMA model with adaptive and fixed initial parameters, it did not outperform the pure autoregressive model used with the same algorithm.</p> <p>The application or the three adaptive algorithms, the Box-Jenkins' approach, the recursive least-squares and the stochastic approximation, has been studied for the PEC configuration and the performance of the predictor was evaluated in each case. The results of this study indicate that combining stochastic approximation with the time series, and including an adaptive quantizer is applicable to differential encoder configurations, mainly the DPCM, with slight modificiations, and would yield better signal-to-noise ratio.</p>