Simplified design of the Speech Recognition System

Disadvantages of the currently used Speech Recognition Systems (SRSs) and alternative ways of their evolution are presented and discussed. In our opinion, SRSs are rather static structures with a lot of predefined knowledge that is built into them upon their creation, and usually remaining unchanged or unadapted during the recognition process. Several possible ways of increasing the amount of the dynamic, automatically learned knowledge in the next generation of SRSs are discussed; this is of a particular importance for under-resourced languages. A group of SRSs, i.e. compact SRSs with a limited vocabulary based on the Neural Network as an acoustic model, is of a particular interest. Its structure is more compatible with the recent developments in the field of distributed and parallel processing. Two experimental systems are presented and tested on a simple phoneme recognition task. One system is a fairly complete SRS based on the Neural Network as an acoustic model and Viterbi search as a time model. The second system is much simpler using only the Neural Network as an acoustic model. It will support our further research in this field.