基于TVAR模型和粒子滤波的语音增强方法

Speech Enhancement Based on TVAR Model and Particle Filter

实际的语音以及语音中掺杂的噪声一般都是非平稳的。本文详细分析了TVAR(时变自回归模型)语音系统模型,把利用TVAR模型增强语音分解成卡尔曼滤波和粒子滤波两步,以减小运算量。同时在粒子滤波中,为克服粒子退化效应,引入了粒子重采样技术提高粒子滤波精度。实验证明,这种增强语音方法无需对语音分帧处理,无需要求噪声是否平稳,能很好地跟踪语音信号的非平稳性,对系统初始值设置不敏感,增强后的语音信号信噪比得到明显改善。

Practically, both clean speech signal and noise are nonstationary. In this paper, a detailed analysis of TVAR （Time- Varying Autoregressive） speech model is presented. In order to reduce the computation, the method of speech enhancement using TVAR is divided into two steps： Kalman filter, particle filter. In the particle filter step, resample algorithm is introduced to increase the accuracy of particle filter, and to overcome particle degeneracy problem. Experimental results show that the method presented in this paper has good performance in tracing nonstationary speech signal, without requirement of splitting speech into frames and the stationarity of noise. The method is also of low sensitivity to the initialized parameter of the system. The SNR of enhanced speech signal is improved obviously.