Speech processing algorithm for isolated words recognition

Abstract

Speech processing algorithms have been intensive developed since 70's and today are often implemented in daily-use devices (personal computers, mobile phones, smartphones etc.). Unfortunately, advanced algorithms have relatively high calculation costs thus need efficient (and expensive) implementation hardware. In this paper a simple speech-processing algorithm able to recognize a spoken word from previously created constant words set has been presented. This functionality is useful in many applications e.g. in voice controlled switching devices. The described algorithm has relatively low cost with sufficient efficiency and could be implemented in a simple and cheap hardware platform. The basic idea is based on the signal analysis in time domain, where the envelope of the signal is calculated and compared with previous created pattern stored in memory. The algorithm of pattern set creation is based on piecewise linear approximation. Moreover, user could create a collection of words which have to be recognized. The proposed algorithm was written in MATLAB software and tested ‘offline’ on recorded wave files and ‘online’ with music card. Next step of the research will be the implementation of the algorithm in the low-cost 32-bit ARM core microcontroller. Details of used algorithms, first tests, occurred problems and finally conclusions are presented in the paper.