Feature-based approach to speech recognition

The alternative approach to speech recognition proposed here is based on pseudo-articulatory representations (PARs), which can be described as approximations of distinctive features, and aims to establish a mapping between them and their acoustic specifications (in this case cepstral coefficients). This mapping which is used as the basis for recognition is first done for vowels. It is obtained using multiple regression analysis after all the vowels have been described in terms of phonetic features and an average cepstral vector has been calculated for each of them. Based on this vowel model, the PAR values are calculated for consonants. At this point recognition is performed using a brute search mechanism to derive PAR trajectories and subsequently dynamic programming to obtain a phone sequence. The results are not as good as when hidden Markov modelling is used, but very promising taking into account the early stage of the experiments and the novelty of the approach.