Pitch Accent Prediction Using Ensemble Machine Learning

In this study, we applied ensemble machine learning to predict pitch accent. With decision tree as the baseline algorithm, we use popular ensemble method-boosting, at different experiment conditions, such as using acoustic features only, use text-based only, using both acoustic and text-based features to evaluate the performance of ensemble machine learning method. Pitch and energy related acoustic features are derived from statistic methods, and we consider context influences to pitch and energy related features. Models of pitch accent (accent and unaccented) are built at the syllable level. At the same time, we compare support vector machine (SVM) to predict pitch accent at same experiment conditions. Results showed that in all experiments ensemble machine learning achieved improved performance. The best result obtained using ensemble machine learning is 82.60% accuracy to Mandarin read speech.