Recognition of the connected words can be performed by segmenting the connected words automatically into the recognition units which are then identified. But the traditional recognition units such as phoneme, syllable, or demisyllable make the segmentation and identification problems difficult. In the traditional methods, segmentation is difficult because the unstable points in speech signals are selected as the segmenting points. Furthermore, the coarticulation and transition effects cause the identification to be difficult.
In this thesis, a new recognition unit for the Korean connected word recognition system is proposed. In the proposed recognition unit, connected words are segmented at the stable points in speech waveforms. Thus the segmentation is very easy and the identification problem can be easily solved because the coarticulation and transition parts are included in the recognition units.
Experiments are carried out with over 200 Korean connected words which are composed of 1 to 4-digit strings. Experimental result shows over 91% recognition rate for the Korean connected digits.