一种自适应时移与阈值的DCT语音增强算法

发布时间：2018-06-11 14:45

本文选题：语音增强 + 离散余弦变换　；参考：《西安电子科技大学学报》2014年06期

【摘要】：针对现有语音增强方法在低信噪比下性能降低的问题,提出了一种自适应时移与阈值的离散余弦变换语音增强算法.首先,对软阈值函数进行改进,既能消除噪声主导帧中的噪声,也能消除语音主导帧中的噪声,并依据信噪比自适应地选择阈值,较大程度地保留了语音的原始特征.其次,依据基音周期自适应地选择分析窗时移,降低了固定分析窗时移产生的白噪声,并且引入短时自相关函数和短时平均幅度差函数相结合的加权自相关函数,来进行基音周期的检测,提高了基音周期检测的准确性和对噪声的鲁棒性.理论分析与仿真结果表明,该算法在信噪比低至-5dB噪声环境下,相比现有的经验模态分解算法和子空间算法,输出信噪比有较大提高,鲁棒性更好.
[Abstract]:In order to reduce the performance of existing speech enhancement methods under low SNR, a discrete cosine transform speech enhancement algorithm based on adaptive time-shift and threshold is proposed. Firstly, the soft threshold function is improved to eliminate the noise in the noise dominant frame as well as the noise in the speech dominant frame, and the threshold is selected adaptively according to the SNR, so that the original features of the speech can be retained to a large extent. Secondly, according to the pitch period, the time shift of the analysis window is adaptively selected, which reduces the white noise caused by the time shift of the fixed analysis window, and introduces the weighted autocorrelation function which combines the short time autocorrelation function and the short time average amplitude difference function. The accuracy of pitch period detection and the robustness to noise are improved. Theoretical analysis and simulation results show that compared with the existing empirical mode decomposition algorithm and subspace algorithm, the output SNR of the proposed algorithm is much higher and the robustness is better than that of the existing empirical mode decomposition algorithm and subspace algorithm under low SNR to -5dB noise.
【作者单位】：西北工业大学电子信息学院;
【基金】：国家自然科学基金资助项目(60872159)
【分类号】：TN912.35

【相似文献】