Lecture speech recognition considering the speaking rate variation

摘要

In a lecture speech recognition, performance of speech recognition system degrades when a speaking rate is increased. The reason of this degradation is a change of acoustic characteristics not only in frequency domain but also in time domain. Because of these changes, normalization or compensation of the speaking rate is important. In this paper, we propose a speaking rate compensation method which selects an optimal frame period and frame length using a likelihood criterion. This method changes the frame period and length to compensate the speaking rate. However, the optimal frame period and length are different in each utterance. Therefore, our proposed method conducts speech recognition with various frame periods and lengths and determines the optimal frame period and length for the target speech using the acoustic likelihood normalized by the frame period and language likelihood. In a recognition experiment using CSJ corpus, this method improves the performance for high speaking rate speech.