基于统计模型实现语音信号有声/无声检测的研究

基于语音信号短时平稳性以及语音信号和噪声的统计模型，提出了一种语音信号有声/无声的检测方法。该方法可对所有语音短时帧更新噪声参数的估值，因而提高了检测的准确性，根据噪声参数以及语音和噪声的统计特性定义了似然比，确定了有声/无声的判别准则，研究了判别阈值、平滑因子等参数以及缓变非平稳背景噪声对检测结果的影响。实验结果表明，即使背景噪声具有缓变特性，利用该方法仍可得到良好的检测结果。对不同的应用，在检测过程中判别阈值的选取应结合平滑因子进行必要的优化。

Statistical Model-Based Speech/Silence Detector

An improved speech/silence detector is developed by combining a noise parameter estimation method based on short time stationarity of speech signal with a statistical model based speech/silence detection method. The effect of decision threshold value and flatness factor on the detection performance has been studied. The proposed method can estimate noise parameter in all the frames, so the detection can be more exact. Test shows that the proposed method is robust and the detection performance holds good results even if the background noise is a slowly time varying non stationary process. For diverse applications, the decision threshold value should be optimized with the flatness factor.