This paper describes the use of syllabic duration control in vocabulary-free speaker-independent speech recognition (i.e. phonetic typewriter) to significantly reduce the frequency of insertion errors. The duration control algorithm used in this paper is applicable to arbitrary unit length and is particularly effective when applied to mora-timed languages such as Japanese speech. Experimental results show that syllabic duration control can reduce the frequency of insertion errors, thus improving syllable accuracy from 47.7% to 75.6% and word accuracy from 19.5% to 41.7%. It has been also found that syllabic duration control is more effective than phonemic duration control. A combination of syllabic duration control and syllable bigram probabilities further improved syllable accuracy to 78.5% and word accuracy to 48.4%.
Bibliographic reference. Jitsuhiro, Takatoshi / Yamada, Tomokazu / Sagayama, Shigeki (1995): "Syllabic duration control for vocabulary-free speech recognition", In EUROSPEECH-1995, 15-18.