Research And Implementation Of Voice Activity Detection Algorithm Based On GMM And SVM Under Complex Environment

Voice activity detection is one of the key techniques in speech communication. The voice activity of input signal is detected using certain signal. processing technique. With the rapid development of mobile communication and VoIP, voice activity detection has been widely used in speech coding, speech enhancement, speech recognition system and many other systems to reduce encoding speed and improve the efficiency of communication systems. Traditional voice activity detection methods are usually based on threshold comparing or model matching. However, a fixed threshold cannot adapt to different noise environments, while the assumed background noise model may be greatly different from real noise environments.Therefore, the performance of the above two algorithms cannot be guaranteed in complex noise environments, Increasing the robustness and accuracy of voice activity detection in complex noise environments and application scenarios has been a key point of research, and is of great significance for the development of voice communication systems. Support vector machine is a classification method based on machine learning which performs well in audio classification and speech recognition. Therefore, to solve the problems of existing algorithms, this paper proposed novel voice activity detection algorithms by studying GMM and SVM algorithms and combining multiple speech features. The main work is as follows:Firstly, a voice activity detection algorithm based on GMM and SVM is proposed, which has high accuracy and robustness under different noise environments and SNRs. The main steps of the algorithms are as follows. In the first step, multiclass SVM is adopted to determine the noise environment,and suitable feature parameters are chosen accordingly. GMM algorithm follows to transform the feature parameters, and the GMM supervectors are used as the input of SVM. Finally, K-L kernel is formed using the distance between the GMM supervectors, and optimal SVM parameters are chosen to construct the SVM classifier, which distinguishes between speech and noise accurately. Simulation results show that the proposed algorithm performs better than the traditional voice activity detection based on GMM under babble noise, machine noise and white noise environments, especially under a low SNR circumstance.Secondly, a multifeature voice activity algorithm is proposed, aiming at improving the accuracy of far-end voice activity detection in the echo cancellation module in VoIP system with a relatively low computing complexity. In the algorithm, the energy ratio and cross correlation of far-end signal and near-end signal are calculated to exclude the circumstance that far-end signal only contains residual echo from near end, thus increasing the accuracy of far-end voice activity detection, and preventing false filtering and coefficient update. Simulation and actual test results show that the accuracy of proposed far-end voice activity detection algorithm is higher than the traditional far-end voice activity detection based on energy, and the performance of echo cancellation is significantly improved therefore.