TABLE OF CONTENTS Whether a rule-based (and explicative) or a concatenative (and descriptive) approach is used, the time signal evolution can either be represented by a model or not. The advantages of using a model are its capacity to reduce the acoustic signal's redundancy and to define parameters that are better fitted to acoustic processing.
A rule-based system cannot avoid such a model, as it is essential to reduce the quantity of information to design algorithmic solutions based on a finite set of controllable rules. Conversely, concatenative approaches based on pre-recorded units usually preserve the signal in it's raw time-domain form, i.e. without modelling.
The advantage of a parameterc model is twofold. By reducing redundancy in the signal, it acts as a compression scheme in cases where storage capacities for the synthesis system are limited (for instance, in embedded systems). It also offers a parameterc representation that is easier to relate to supra-segmental information.
Many models describing the time-domain evolution of speech signals have been proposed over the past 50 years. Even if the signal to be modeled is considered as the simple variation of a physical quantity with time, most of them are based on an assumption which stems from the speech production mechanisms: the source-filter decomposition assumption.
Models of speech signal representation can be grouped into two categories depending on the corresponding methodological hypothesis:
models relying on a production hypothesis describe the generation of the speech signal on the basis of physiological parameters. The goal is...
