基于能量谱熵的英语摩擦音检测方法

An English Fricative Detection Method Based on Energy Spectrum Entropy

根据摩擦音发声时的频谱特点,提出一种基于能量谱熵的摩擦音检测方法.该方法首先利用不同音素的语谱能量特点检测出音素边界.然后计算每个语音段的能量谱熵,并将超过阈值的语音段作为候选.最后根据语音段的长度、开始结束时的能量突变等对特征候选语音段后处理,去除错误候选.实验表明,在干净环境中并且容错误差为20 ms时,摩擦音的检测率达到96.9%.

According to the spectrum characteristics of fricatives, a fricative detection method based on the energy spectrum entropy is proposed. Firstly, phone boundaries are detected based on spectrum of different phonemes. Then, each spectrum entropy of speech segments is computed and the segments whose entropy exceeds the threshold are selected as candidates. Finally, post processing is conducted to remove the insertion errors according to parameters of segment length and the sudden changing of energy at segment starts and ends. The experimental results show that the accuracy of the proposed method is up to 96.9% in clean circumstance when the tolerance is 20 ms.