Design and analysis of approximation algorithm for high performance computing in speech recognition

Abstract

A comparative study in number of computation required for word recognition in different states of dynamic time warping algorithms is made. It is observed that the number of computations required in the pattern matching reduces with the increase in the number of state up to five state dynamic time warping. It is also observed that in a single state dynamic time warping algorithm the boundary will be of two parts, first is on start and second is on last while in the division, boundary level will be more and more, viz., for two state there will be three boundaries. It is found that the five parts are ideal since each word consists of fundamentally five phonemes (on an average) that is combination of vowels and consonants. The number of computations required reduces using the multi-state dynamic time warping algorithm (high performance computing) for speaker dependent isolated word recognition in speech systems. newlineExperiments are performed over the data available in the database for six state dynamic time warping, seven state dynamic time warping and so on. It is observed that the number of computations decreases significantly but the recognition score also decreases rapidly. So, one should avoid performing the six state dynamic time warping and on wards for speech recognition systems. newline