自适应语音分段增强算法

AN ADAPTIVE SEGMENTATION AND WAVELET-DOMAIN ENHANCEMENT ALGORITHM (ASWE) FOR SPEECH SIGNAL

近年来小波理论在信号分析、处理中得到了广泛的应用,本文提出了一种自适应分段小波域语音增强(ASWE)算法,即采用局部余弦包变换对语音信号自适应分段,然后对每一段语音采用基于小波变换的语音增强处理。该方法不需要噪声的先验知识,且适合于缓慢变化的非平稳噪声,最后的仿真实验表明,该方法比直接用小波去噪效果好,是一种有效的语音增强技术。

Speech enhancement is one of the most improtant branches in speech signal processing. The theory of wavelet transform has been successfully applied in signal analysis and signal processing in recent years. One typical application in signal de-noising is the method proposed by Donoho based on the theory of nonlinear threshold in wavelet domain. This method can also be used in speech enhancement. Speech signal has two basic classes, voiced and unvoiced. The voiced signal has the feature of periodicity while the unvoiced signal has the same properties as the white noise. So, if the nonlinear thresholding in wavelet domain is directly applied to process one continuous speech, some part of the speech may be processed as noise. A belter choice to enhance speech is to select different wavelet processing methods for voiced and unvoiced signal. An adaptive segmentation and wavelet-domain enhancement algorithm (ASWE) for speech signal is proposed in this paper. The speech signal is adaptively segmented using local cosine packet transform firstly, and then each segment of speech is enhanced based on wavelet transform. This method does not need any priori information of the noise, and it is suitable for the nonstationary noise environments. The simulation results show that this method can offer better performance than the method using wavelet transform directly can.