Research On Noisy Voice Activity Detection Method

Endpoint detection is the first crucial technology, whose accuracy determines the success of the whole speech recognition system to some extent. The detection of the presence of speech embedded in various types of non-speech events and background noise is called endpoint detection, speech detection or voice activity detection (VAD).This article first briefly explains the mechanism of the speech signal, the signal acquisition and digital voice, Window Frame pretreatment theoretical knowledge. Then the last few years the number of domestic and foreign scholars endpoint detection algorithm, and a detailed analysis of the two endpoint detection algorithm and the simulation results. Later, the three of the noise endpoint detection algorithm:(1)Information entropy function in the absence of noise, can distinguish between good and the voice of the silent, but the noise, its detection capabilities of the poor. This paper put forward energy-weighted for the information entropy algorithm, to distinguish between noise and speech of the above. (2)Voice and the noise characteristic of the spectrum is a big difference. In the spectrum of noise, the frequency of changes is slow in between the gentle; Bandwidth between voice and the more radical changes. Because of this, the frequency of variance, but the variance in the frequency bands used alone results of tests under low signal-to-noise ratio is still not very satisfactory. Therefore, this article comes up with a method based on time-frequency- variance-summation method.(3)Spectral subtraction is a traditional method. However, the use of fixed without reducing the sound spectrum was prone to error as a noise sample. Spectral entropy method is effective endpoint detection, but the environment in low signal-to-noise ratio detection results will be greatly reduced. In this paper,a method that noise reduction and endpoint detection is simultaneous is proposed.