Research on Voice Endpoint Detection in Low SNR Status
Li Mao · Modern Electronics Technique · 2009
Short-time average energy and short-time zero-crossing rate are introduced in traditional voice endpoint detection methods.But in practical environment,these methods are not accurate,especially in this environment where has powerful background noise.Therefore,an improved method based on computing the distance of linear prediction cepstral coefficients between the several frames is proposed.First,using power spectral subtraction to lowen noise,then directly computing the distance of linear prediction cepstral coefficients relative to origin,and judge it according to threshold that decided beforehand.This method can efficiently lowen the probability of false judgement caused by the different sign of corresponding coefficients between the frames.The experiment result shows that the method can detect voice efficiently in low SNR status.