| Voice activity detection,which detects the speech is absent or present from background noise is cruial for many application of speech signal processing.The accurate detection can decrease the noisy bit rate, improve system efficiency for speech code and improve the speech recognition accuracy with less recognition time.Because of the various noise pollution and circumstance changing ,its performance will degrade severely.So voice activity detection technology at low SNR has become a major issue in speech processing.In this dissertation, firstly the feature of noise signal and speech signal and the principle of voice activity detection technology are introduced, and then the existing typical algorithms of voice activity detection are analysed, later a statistical model-based voice activity detection algorithm is studied.In most of VAD algorithms which mainly oprate in the discrete Fourier transform (DFT) domain, it is assumed that the distribution of noisy speech spectra is characterized by the Gaussian densities ,it has been reported that the DFT coefficients of noisy speech is better modeled by the Laplacian distributions. So in this dissertation, these two model-based algorithms are analysed. One of the key issues in VAD is robust estimation of the priori signal-to-noise ratio (SNR). The drawback of the existing decision-directed parameter estimation method is analysed, and a predictive (PD) estimation is applied to estimate and renew parameters. The experiments show the algorithm with this method has better performance. Secondly the smoothed form for the likelihood ratio (LR) which makes it possible to improve the performance of the LR is introduced. In the smoothed form, a more smooth estimate for the magnitude spectra is applied instead of its instantaneous value,and the smoothing factor is determined by the experiments.At the end of the dissertation, the improved algorithms are summarized. Some problems that need more improvements in the future study are suggested. |
PDF Full Text Request