Ranknet based English stressed syllable detection

A lot of time or frequency domain speech features had been applied to address the problem of English stressed syllable detection. Researchers had proved that the combination of multiple features is necessary to get better performance. But up to now, the tasks of seeking new feasible speech features and innovative feature fusion approaches are still open. This paper proposes a detection-by-ranking approach to address the stressed syllable detection problem based on the RankNet technique. The approach is able to find out the stressed syllable through one by one comparison of feature vectors corresponding to vowels of syllables in a multi-syllable word. This paper also introduces the fractal dimensions of each vowel as one type of the stress features. Experiments conducted on the corpus TIMIT show that the proposed feature fusion method reaches high performance and the introducing of fractal dimension is helpful for improving the detection correct rate.