Connectionist architecture for all Mandarin syllables recognition

This paper presents a modular connectionist architecture for all Mandarin syllables recognition. The technique used is based on the time-delay neural networks (TDNN). The architecture developed is capable of recognizing all 35 finals, 21 initials and 4 tones of the entire vocabulary of isolated Chinese syllables. Experimental results show a recognition accuracy of 93.9% for the finals, 92.7% for the initials and 99.3% for the tones, giving rise to an overall syllable recognition rate of about 90%.