Large vocabulary speech recognition in noisy environments

A ІКПѴ set of speech feature parameters based on multirate subband analysis and the Teager Energy Operator (TEO) is developed. The speech signal is first divided into nonuniform subbands in mel-scale using a multirate filter-bank, then the Teager energies of the subsignals are estimated. Finally, the feature vector is constructed by logcompression and inverse DOT computation. The new feature parameters (TEOCEP) have a robust speech recognition performance in car engine noise which has a low pass nature. In this thesis, we also present some solutions to the problem of large vocabulary speech recognition. Triphone-based Hidden Markov. Models (HMM) are used to model the vocabulary words. Although the straight forward parallel search strategy gives good recognition performance, the processing time required is found to be long and impractical. Therefore another search strategy with similar performance is described. Subvocabularies are developed during the training session to reduce the total number of words considered in the search process. The search is then performed in a tree structure by investigating one subvocabulary instead of all the words.