Recognition of isolated words in Bulgarian, by means of HMM

The problem of the recognition of Bulgarian words by means of HMM (hidden Markov models) is discussed. The speech signal was low-pass filtered up to 4 kHz, sampled at 10 kHz, and pushed directly into the computer's memory (IBM PC/XT). Unvoiced segments were separated, and the pitch period was evaluated. For every voiced and unvoiced segment 12 LPC (linear predictive coding) coefficients were computed. These segments were used as states q/sub i/ in HMM and their LPC coefficients-an acoustic vector y/sub t/. On the basis of the training set a HMM for every word was generated. A modified Bayesian decision rule is proposed. As a result, if the decision rule is satisfied, the classification is simple; otherwise, the classification is given in the form of ordered couples. The proposed approach shows higher accuracy and is appropriate for word, command and expression recognition.<
>