Research On Voice Activity Detection Algorithm In Low SNR

Voice Activity Detection(VAD),as the front-end operation of various speech signal processing systems,plays an important role in the field of speech signal processing.VAD under low SNR conditions is a technical problem in the field of speech processing,which hinders the expansion of speech processing technology into applications such as noisy environments and shortwave communications.In order to search for an ideal VAD algorithm,it can not only ensure a good accuracy rate under the environment of low SNR,but also meet the requirements of less prior knowledge and computation.In this paper,the VAD algorithm under low SNR conditions is studied and the corresponding solutions are proposed.This paper summarizes the progress and achievements of VAD research in recent decades.Several commonly used VAD algorithms are analyzed from the accuracy,stability,adaptability and computational complexity of voice endpoint detection.It is found that under the condition of low SNR,the detection accuracy of traditional voice endpoint detection algorithms declines sharply.In order to solve this problem,this paper carries out research from two aspects: improving the effect of noise reduction and enhancing the robustness of VAD algorithm.In the aspect of speech noise reduction,this paper proposes an improved spectral noise subtraction algorithm,which uses endpoint detection,piecewise noise estimation,dynamic parameter adjustment and other methods to solve the problems of noise-segment estimation,inaccurate noise estimation and large music noise existing in conventional spectral noise subtraction algorithm.In the aspect of VAD algorithm,this paper proposes an improved subband energy-entropy ratio VAD algorithm,which uses subband partition,optimization of energy calculation,optimization of spectral entropy calculation,two-level smoothing processing and other methods to effectively enhance the robustness of the algorithm under different noise conditions.On the basis of the above research,an improved algorithm combining spectral subtraction with sub-band energy-entropy ratio is proposed.The algorithm uses a priori signal-to-noise ratio estimation,spectral noise reduction,sub-band energy-entropy ratio endpoint detection and other methods to effectively enhance the robustness of the algorithm under low signal-to-noise ratio conditions.Finally,this paper synthesizes noisy speech with different noise types and signal-to-noise ratios using NOISEX-92 noise library and pure voice files,compares the conventional endpoint detection algorithm with the improved algorithm proposed in this paper,and verifies the effectiveness of the proposed algorithm.The research carried out in this paper and the improved algorithm proposed provide anew idea and solution for voice endpoint detection under low signal-to-noise ratio,which has a certain reference value.