Abstract
This thesis describes our research on stress detection and intonation assessment. Stress detection is to find the most prominent syllable in an utterance of a multi-syllabic English word. In an intonation assessment system, an utterance pronounced by the student is compared with the one pronounced by the teacher, and a score is produced based on similarity of pitch contours. Four methods are compared in stress detection, including three methods proposed by past research and a single-stage stress detection method proposed in this thesis. Our method trains classifiers for each n-syllabic words (n=2-5) using pitch-related features, energy-related features and duration extracted from the vowel segment of each syllable. The proposed method yields the best recognition rate among the four methods, outperforming the previous method by about 2%. On the other hand, we also compare the detection performance by using different combinations of features, including pitch-related features, energy-related features, and duration. We found that the best evaluation result is produced when all features are combined. The intonation assessment mechanism proposed in this thesis includes two parts: assessment for each word and for the whole sentence. In order to build the mechanism, a corpus containing two types of utterances was recorded, namely, utterances whose intonation patterns are similar to the teachers’ and utterances whose intonation patterns that are not so similar. This makes the classification task a two-class classification problem. Three methods are compared in classifying similar and dissimilar utterances, including one method proposed by past research, one method using Gaussian mixture model, and one method using a support vector machine. Among the three methods, the one using a support vector machine gives the best recognition rate, outperforming the previous method by about 14%. We also evaluated the performance by using different combinations of features, including correlation coefficient, root-mean-square error, and sorted error vector. The best evaluation result is obtained when only correlation coefficient is used, and adding other features reduce recognition rates. Lastly, we evaluate the performance of the system by using different pitch-tracking methods, UPDUDP and the pitch-tracking function provided by Praat, and by extracting pitch contours from different sound segments: vowel phones only and from all pitched phones. The best evaluation result is obtained when the pitch-tracking function provided by Praat is used to extract pitch contours from all pitched phones. In order to evaluate the performance of the intonation assessment system, a smaller corpus was recorded and rated by eight human-raters. The inter-rater correlation coefficient is 0.80, and the correlation coefficient of the system scores and human scores is 0.65, showing satisfactory performance of the system.