Steve Waterhouse, Anthony Robinson
In this paper we consider speech coding as a problem of speech modelling. In particular, prediction of parameterised speech over short time segments is performed using the Hierarchical Mixture of Experts (HME) (Jordan & Jacobs 1994). The HME gives two ad(cid:173) vantages over traditional non-linear function approximators such as the Multi-Layer Percept ron (MLP); a statistical understand(cid:173) ing of the operation of the predictor and provision of information about the performance of the predictor in the form of likelihood information and local error bars. These two issues are examined on both toy and real world problems of regression and time series prediction. In the speech coding context, we extend the principle of combining local predictions via the HME to a Vector Quantiza(cid:173) tion scheme in which fixed local codebooks are combined on-line for each observation.