It is well known that DPCM using linear prediction has about two bits gain over PCM for speech signal. To improve further, the predictor parameters need to be adjusted according to the changing statistic of signal.
In this paper, two types of adaptation correlation method and gradient following method are designed and implemented for linear predictors. The correlation method, though it has brief expressions in analysis, is found to be very difficult in realization.
On the other hand, the gradient following method has the advantage of simple and stable implementation.
It is shown that an adoptive second order linear predictor using gradient following algorithm provides 10.7db SNR gain over PCM for speech signal sampled at the rate of 10K samples/sec.
Adaptive linear predictor 로서 correlation method 와 Gradient following algorithm 를 구성하여 실험 하였다.
Correlation method 는 그 adaptation scheme 이나, mathmatical expression이 명확하기는 하지만 실제로 구성하는데는 analog arithmatic unit 의 accuracy 관계로 많은 문제점이 발생하였다.
Gradient following algorithm 은 parameter convergence에 대한 구체적 해석은 어렵지만 실제로 구성하는데는 훨씬 간단하며 arithmatic unit 의 accuracy 에 큰 영향을 받지 않았다. 10 K samples/sec.로 sampling 된 신호에 대하여, Gradient following algorithm 으로는 10.7 (db) 의 SNR 개선을 얻었다.
