Hierarchical Speech Signal Model for Expert Knowledge based ASR/ASU Systems
Goran S. Jovanovic · International Symposium on Information Theory and its Applications · 1994
The paper presents an original hierarchical speech signal (HSS) model which is suitable for expert based development of automatic speech recognition systems (ASR) by the use of the focusing structural speech signal analysis. A general block scheme for development and building-up of these systems is formulated. It corresponds to HSS model and provides convergence of an ASR system to the automatic speech understanding (ASU) system. The suitability of ASS model for the development of ASR systems is illustrated by an original (simple version) isolated word recognition system which belongs to the class of linguistic ASR systems. Performances of that system have been verified on a large speech data base, which included data obtained by digital acquisition in the real world application environment. Beside its remarkable numerical efficiency, important feature of that system is its speaker independence and possibility to be extended for continuous speech recognition tasks.