Third International Conference on Spoken Language Processing (ICSLP 94)

Yokohama, Japan
September 18-22, 1994

State Duration Constraint using Syllable Duration for Speech Recognition

Yumi Wakita, Eiichi Tsuboka

Central Research Laboratories, Matsushita Electric Industrial Co., Ltd, Kyoto, Japan

For speech recognition using HMMs, we propose an adaptive syllable duration constraint method. The method constrains syllable durations using a relation each syllable included in the same utterance [1]. The duration of t-th syllable d(t) is predicted by using d0(l)... d0(t-1) the durations of syllables which have been recognized . After a syllable is recognized, if the durations of the t-th syllable is very different from the predicted value, the result is rejected. Advantages of this method are

This paper proposes a new method of state duration constraint using SDP. At first the duration of s-th state of t-th syllable is predicted using the duration of t-th syllable which is predicted by SDP. Next the matching period of the state is constrained using the predicted state duration.

We evaluate this method using word and sentence recognition. For word recognition (100 words and 9 speakers, open test ), the error reduction is 14% and the matching speed is 25% shorter. For sentence recognition ( 50 sentences and 6 speakers, open test ), the error reduction is 46% and the matching speed is 50% shorter.

Full Paper

Bibliographic reference.  Wakita, Yumi / Tsuboka, Eiichi (1994): "State duration constraint using syllable duration for speech recognition", In ICSLP-1994, 195-198.