A Local-Segment-Based Approach for Spotting Large Vocabulary Chinese Keywords from Mandarin Speech

Abstract

This paper presents an efficient local-segment-based approach to spotting large vocabulary Chinese keywords from a spontaneous Mandarin speech utterance. In stead of searching through the whole utterance with the aid of filler models, the approach simply searches through the local segments within the utterance to find out the most possible keywords. The monosyllable-based technique for very large vocabulary Mandarin speech recognition [1][2] was modified and applied to a three-phase framework with a special scoring method for keyword spotting. There are three main advantages in this method. First, the difficulties in training the fuller models can be avoided. Second, it is unnecessary to retrain the keyword models when the vocabulary is changed. And third, this approach not only works for small vocabulary problems, but it works for large keyword vocabulary problems as well. To improve the performance of the approach, several additional techniques are also developed to further enhance its speed and accuracy. Two tasks with completely different characteristics were tested here. The first one has 9 keywords, and each keyword includes 2 syllables. A spotting rate as high as 93.73% is obtained for the test based on this task. The second task has 2611 keywords, and each keyword includes 2 to 20 syllables. A 84.52% spotting rate for the top 10 candidates is attained with a speed requiring only 1.6 times of attained with a speed requiring only 1.6 times of utterance length when the test was performed on a Sparc 20 workstation.