Speaker-independent connected Japanese digit recognition based on phonetic approach

This paper presents a phonetic approach to speaker-independent connected Japanese digit recognition. As phonetic units, phoneme-like segments were chosen. Vowel/consonant segmentation was performed in order to identify the boundaries of these units. Knowledge was expressed as a set of word graphs, which were created manually by applying acoustic-phonetic rules to a lexicon. The optimal word sequence was obtained using word spotting and the best first search technique, where the acoustic match was performed using "attributes" which were robust descriptions of acoustic characteristics of the units. A digit recognition error rate of 15.1% was obtained for 24 six-digit strings, each spoken by ten talker. When multiple answers were allowed, the digit error rate was 10.9% (r= 1.13) and 6.2% (r= 1.75), where each digit had r answers on the average.