A Ranknet Based Hierarchical Sentence Stress Detection Approach For Spoken English

Computer Assisted Language Learning based on speech technology is gaining increasing attention in recent years. It provides an interactive platform for students to learn foreign languages. It is an objective of our work to develop such a system to help students to learn English sentence stress. This thesis presents a novel hierarchical method based on RankNet to detect English sentence stress.The hierarchical framework this paper proposes consists of three levels. The first level is to detect the stressed vowel nuclei of independent words in each sentence utterance disregard of the sentence rhythmic stress and the relationship of the stressed and unstressed words or the stressed and stressed words. It is the traditionally lexical stressed syllable detection. The second level detects the sentence stress upon the whole stressed nucleus of each word within each sentence utterance. The third level is rejusting the results of the second level by searching the threshold, which could separate the stressed and unstressed word mostly, from the rank values got by the sentence stress detection model. The rhythmic stress detection is important and difficult in speech recognition. The basic content of this paper is as follows:First, we employ the local feature normalization into the already existed RankNet based stressed syllable detection method, to improve its accuracy rate. Second, we train the RankNet sentence stress detection model to detect the sentence rhythmic stress. According to the RankNet theory, we fetch the speaker independent prosodic features and vowel quality features as terminals to classify each vowel segment as stressed or unstressed, and then improve the accuracy incorporating with the threshold.Finally, to evaluate the two-level stress detection approach objectively, we employ part of ISLE (Interactive Spoken Language Education) data corpus, and the approach is examined and compared to linear discrimination analysis. The ability of each dimension of an individual feature to differentiate stressed vowels in-word and in-sentence is evaluated.The experiment results show that the proposed hierarchical sentence stress detection method based on RankNet achieved a minimum error rate of 22.9% for all the stresses of all the sentences and 43.5% for the first stress in all sentences. The result suggests that the RankNet based approach developed in this paper outperforms linear discrimination based approach for sentence stress detection.