关键词: 对数能量, 倒谱距离, 模糊C均值聚类, 贝叶斯信息准则（BIC）, 端点检测

Abstract: Endpoint detection is one of the key technologies of speech recognition, in order to overcome the undesirable detection results of traditional cepstrum distance in speech endpoint detection algorithm under low signal to noise ratio, combined logarithmic energy feature（LE） with cepstrum features（C） for endpoint detection, proposes a new logarithmic energy cepstrum features（LEC）, uses fuzzy C-means clustering and Bayesian information criterion to estimate features threshold, achieves better endpoint judgment, conducts the SNR simulation from －5 dB to 15 dB with noisy speech under three kinds of typical noise. The results indicate that the LEC method’ detection error rates is just 20.25% and significantly lower than cepstrum and logarithmic energy method, it also can effectively determine the speech endpoint and improve voice recognition results.

Key words: logarithmic energy, cepstrum distance, Fuzzy C-means clustering, Bayesian Information Criterions（BIC）, Endpoint detection