Speech Endpoint Detection Based On Gaussian Colored Noise Environment

Speech startpoint detection is the base in speech procession.Its efficiency, correction will affect the later experiment's result.In the distilling,transmission and disposal's procession,the signal will be polluted by noise,so we must denoise firstly.In this paper,Gauss-Colored-Noise is used to the original noise.Make use of a algorithm to create the Gauss-Colored-Noise's whose mean is zero and variance is one.Compare with other common noise,and conclude its quality.Analyse the Gauss-Colored- Noise's singularity, add it as addition-noise into signal and denoise according to its sigularity. The main algorithm is:decompose the signal with wavelet,calculate the modulus maxium with every waveletlevel's coefficient,find these positions and the relations between the wavelet's coefficient,judge whether signal or noise,reserve the signal,change the value of noise and then use the changed signal's value to rebuild the original signal.There are some traditional algorithms of speech startpoint detection such as:short-time energy,short-time average across-zeros rates,short-time self-correlation and so on.In this thesis the improved algorithm is:wavelet multi-level decomposition's coefficient and speech signal characteristic combine to imply speech signal startpoint detection.First, we decompose the signal with wavelet transform,distill the low-frequence's coefficient.Secondly,enframe the low-frequence's coefficient,calculate the short-time energy ,short-time average across-zeros rates of every frame.Finally,compare the calculated data with the tentative threshold.The result of experiment indicate:the improvement increases the correction and decrease the programme running time.