Automatic speech recognition based on diphones

A hidden Markov model-based system of automatic continuous speech recognition with diphones is proposed. The diphones statistics for Polish language and rules of creating the observation probability vectors for diphones are presented. The authors suggest using diphones as more sufficient units which carry more suprasegmental knowledge. A method of automatic finding of the diphone segments and their parametrization, detected diphones labeling and recognition test, were realized. The effectiveness for semicontinuous speech from a data base containing about 115 sentences, for a description of the 16 parameters of the short-term spectrum and the ANN/HMM algorithm exceeds a 90% correct recognition (a result that is better by about 9% in relation to the analogous experiments that used phonemes as basic units)