Design of a command interface with a dynamic grammar speech recognition engine

This paper presents the design of an open grammar speech recognition engine. The engine is built to operate in real-time and kept simple. The objective of this paper is to show how task dependent knowledge can improve speech recognition accuracy by cancelling out parts of the command set depending on the state of the interface. The system is a phone-based recognizer and therefore the vocabulary and grammar can easily be extended without retraining the system. The hypothesis of this paper is that speech recognition accuracy will improve if higher level knowledge about the state of a task is used by the recognition engine.