Feature-based recognition of nonsonorant consonants in Chinese speech

The recognition of nonsonorant consonants is a key problem in unlimited-vocabulary Chinese speech recognition. A feature-based nonsonorant consonant classification scheme is assumed. The unknown single-syllabic consonant-vowel utterance is first divided into consonant and vowel segments; then the features of the consonant segment are extracted. Next, the unknown consonant is categorized into subclusters at each node of a decision tree according to the feature values. This procedure is repeated until an end node is reached. The average recognition rate is 84.2%, and many of the features used are not speaker-sensitive