APPLICATION OF SEQUENTIAL ESTIMATION TO TIME-VARYING ENVIRONMEN'I' COMPENSATION
Dong Kook, Sang Kim, Ryong Kim · 1997
Sequential approaches are proposed to compens,ate for the effects of the nonstationary environment. Unlike the batch approach- es, the proposed methods derive a different parameter estimate for each time using the sequential expectation maximization (EM) algorithm. More- over, we also propose the forward-backward estimation scheme as an improvement of the sequential parameter estimation. Environment compensation has been considered to be essential for robust speech recognition in adverse conditions where there exist unwanted added noise or spectral tilt. A number of prevailing environment compensation tech- niques usually make an approximation to the nonlinear speech contamination procedure by a simplified linear model (l, 2, 3, 41. Based upon the simplified model, parameters concerned with the characteristics of the adverse environ- ment are estimated and used to transform the noisy speech features to the clean features. In this paper, we propose sequential approaches to environment compensa- tion. Conventional techniques, in general, make an assumption that within an utterance the characteristics of the background environment do not change, and use batch algorithms for parameter estimation. However, this assumption is far from the real situation in which the characteristics of the environmen- t are usually time-varying. Unlike the batch approaches, our methods are based on the assumption that the characteristics of the environrnent are s- lowly varying, and a different parameter estimate is obtained for each time. In order to compute the parameter estimate efficiently, we apply the sequential expectataon maxzmzzataon (EM) algorithm (5). As in many Sequential pa- rameter estimation approaches, forgetting factors are introduced to track the time-varying parameter. Moreover, the forward-backward estimation scheme is also proposed as an improvement of the sequential parameter (estimation. Both the forward and backward sequential EM algorithms are proceeded in the opposite directions to compute two kinds of parameter estimates at each