Exploiting transitions and focussing on linguistic properties for ASR

This paper describes three cross-language ASR experiments which use hidden Markov mode lling. The first one shows that consonant identification improves when vowel transitions are used. In particular, the consonants’ place of articulation is identified better, because the vowel transitions contain formant trajectories which depend on the consonant’s place of articulation. The second experiment compares consonant identification results when acoustic parameters belonging to the consonant itself (no vowel transitions are used in the second experiment) are used as input to hidden Markov modelling directly with identification rates when acousticphonetic mapping is performed before applying hidden Markov modelling. It is shown that acoustic-phonetic mapping greatly improves consonant identification rates. In the third experiment, the acoustic parameters from the vowel transitions are also mapped onto consonantal ( not vocalic) features, as are the acoustic parameters belonging to the consonants. The additional use of vowel transitions does not lead to further improvements in the consonant identification, however. This is probably due to undertraining of the vowel transitions in the Kohonen network.