A Stressed Syllable Labeling Approach Using Fractal Dimensions

In this paper, a fractal analysis approach is presented to label the stressed syllable in spoken English evaluation applications. Firstly, two popular fractal dimension computing methods are introduced and are compared for the purpose of extracting stressed features in speech signals. According to the experiment results, the morphic covering algorithm wins out and the stress labeling approach based on it is further developed. The prototype is constructed based on the Sphinx-4 platform. To evaluate the contribution of proposed approach to English lexical stress detection, several classical features including energy, duration and pitch are also used to perform English stress detection. Finally, experiments for the stressed syllable labeling are conducted. The experiment results show that the accuracy rate of 85.83% is achieved for the proposed approach, which outperforms all the results based on classical features mentioned above.