Speech Recognition with Word Fragment Detection Using Prosody Features for Spontaneous Speech

This investment proposed a novel approach for word fragment detection with prosody features for spontaneous speech recognition. Incomplete pronunciation of word result in ill-form fragment in word-building that causes the performance of language model in speech recognition is dramatically decreased. Instead of lexical word, prosody word is used to be building block for spontaneous speech processing recently. Prosody features are further extracted from prosody word and fed into the decision tree to judge the prosody word is complete word or word fragment. There are three categories feature sets are employed here: pitch related, intensity related, and duration related features are included. For evaluating the proposed method, the Hidden Markov models (HMMs) based speech recognition core was developed to be the baseline. The proposed method is integrated into the baseline to provide the word fragment detection capability and enrich the performance of spontaneous speech recognition. According to the experimental results, the performance of proposed method outperforms traditional speech recognition especially in insertion and deletion error. This shows that the word fragment detection can obtain the improvement for spontaneous speech recognition.