Speech recognition method with context dependent hidden Markov models

摘要

The method exploits the additional tonal contexts at the ends and the beginnings of the words (M,N). State transitions (Z1 to Z4) hold good for the tonal transitions (Li-1,Li). Hidden Markov models for these tones are supplemented by special context models for the beginning and end of the word. The same consideration applies to the tones (Lj,Lj+1) and the state transitions (ZN to ZN+3) of the other word. At least one of the tonal contexts represents a context boundary which is evaluated in the speech recognition. The tonal contexts are used for the modelling of the characteristics in the co-articulation during the transition from one word to the other.