On Multi-Level Machines for Continuous Speech Recognition

In this paper we introduce the concepts of multi-level machines and of multi-level dynamic programung. These machines are well suited for the difficult continuous speech recognition problem because firstly, they permit the integration of several knowledge sources, secondly, they allow an optimal search based on dynamic programmng and thirdly, they can deal with semantic constraints. Furthermore these semantic constraints have the interesting property to be dynamic, i.e. they can be modified easily by the speech recognition system itself. The important consequence of this property is that the multi-level machines presented in this paper have a potential self-leaminy hability. In this paper we present the concepts of multi-level machines and multi-level dynamic programming. These muulti-level machines, on one hand, permit the integration of several knowledge sources such as phonology, syntax, semantic etc, and on the other-hand can take into acount local semantic constraints. As it will be shown these semantic constraints are "dynamic", in the sense that the speech recognition system can modify them easily. In section 2 we begin by showing how a 1-level machine can be built from simple finite-state automatons. These automatons are called cells because the entire machine is built from these elementary machines. The description is done in such a way that the iterative process to generate a general n-level machine can be induced easily. The semantic links are also describes and we put in evidence the fact that firstly, they can be modified easily, and secondly that such modifications can be realized by the speech recognition system itself. This interesting property confer to the system a self-learning hability. In section 3 we introduce the formalism necessary to explain how an optimal seauch can be realized in an n-level machine. From this mathematical discussion we show that the solution found is optimal for all the levels of the machine.This other intersting property confirms the power of the n-level machines. The basic of the machine is a simple finite-state machine as illustrated by figure 1 (1). FIGURE 1. The basic cell of a multi-level machine : a simple finite state automaton. The basic cell is characterized by a set of starting states ES, a set of ending states FS, and a set of terminal symbols T. For example, in the case of figure and An element of the starting state set and an element of the ending state set are particularized : they are called respectively the …