多尺度卡尔曼滤波语音增强算法研究  

Research on multi-scale Kalman filtering speech enhancement algorithm

作　　者：郑彦虎 唐云[1] 张澎[1] 闵宇航 ZHENG Yan-hu;TANG Yun;ZHANG Peng;MIN Yu-hang(College of Information Sciences and Technology,Chengdu University of Technology,Chengdu 610059,China)

机构地区：[1]成都理工大学信息科学与技术学院,成都610059

出　　处：《信息技术》2021年第7期20-25,30,共7页Information Technology

基　　金：国家自然科学基金项目(61972324)。

摘　　要：传统卡尔曼滤波在追求去噪的同时,往往因量测方差不准确的估计,伴随着较为严重的语音失真和畸变,为此提出一种多尺度卡尔曼滤波语音增强方法。利用小波包多尺度频带划分能力对含噪语音进行分解,通过熵求取最佳树,使得其可以更好地拟合含噪语音,对小波包分解后以语音信号为主的低频系数进行噪声量测方差重估计,以此作为卡尔曼滤波先验知识进行迭代,实现了语音和噪声量测方差求解的分离,克服了传统卡尔曼滤波单一噪声估计的局限性而导致的对信号过衰减问题。实验表明,该算法能更好地消除噪声,减少语音失真,且具有一定的鲁棒性。It is often accompanied by serious speech distortion and distortion due to inaccurate estimation of measurement variance in the process of the traditional Kalman filter pursuing denoising.Therefore,a speech enhancement method of multi-scale Kalman filtering is proposed.The multi-scale wavelet packet frequency band partition ability is used to decompose,voice signals with noise by entropy to calculate the best tree,and make it a better fitting with noise speech.After the wavelet packet decomposition,low frequency coefficient with voice signal measurement noise variance estimation is carried out,and iterate as prior knowledge of kalman filtering and realize the seperation of speech and noise measurement variance,which overcomes the limitations of traditional single kalman filtering noise estimation which leads to signal attenuation problem.Experiment results show that this algorithm can eliminate noise and reduce speech distortion,and has certain robustness.

关 键 词：语音去噪 小波包 多尺度分解 卡尔曼滤波 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]