Rate Distribution Between Model and Signal

Knowledge of a statistical model of the signal can be used to increase coding efficiency. A common approach is to use a fixed model structure with parameters that adapt to the signal. The model parameters and a signal representation that depends on the model are encoded. We show that, if the signal is divided into segments of a particular duration, and the model structure is fixed, then the optimal bit allocation for the model parameters does not vary with the overall rate. We discuss in detail the parameter rate for the autoregressive (AR) model. Our approach shows that the square error criterion in the signal domain is consistent with the commonly used root mean square log spectral error for the model parameters. Without using perceptual knowledge, we obtain a rate allocation for the model that is consistent with what is commonly used. This model rate is independent of overall coding rate. We provide experimental results for the application of the autoregressive model to speech that confirm the theory.