基于高斯粒子滤波器和TVAR模型的语音增强技术

Speech enhancement technique based on gaussian particle filter and time-varying autoregressive model

为了提高对时变自回归(TVAR)模型参数的估计精度,改善语音信号增强效果,通过将TVAR模型的参数转换为格型滤波器的反射系数,给出了一种判断模型稳定性的简单方法。将TVAR模型的信号和反射系数矢量增广为状态矢量后,应用高斯粒子滤波器(GPF)估计TVAR的模型参数,构造了语音增强算法。通过蒙特卡洛仿真实验比较了扩展卡尔曼滤波器(EKF)、标准粒子滤波器(PF)和GPF的语音增强效果,结果表明本文提出的TVAR模型能较好地描述语音信号的变化特性;PF比传统的卡尔曼滤波具有更好的滤波能力,而GPF能够在一定程度上克服粒子的退化现象,具有比标准PF更强的估计TVAR模型参数的能力,从而获得了更好的语音增强和去噪效果。

In order to enhance estimation precision of time-varying autoregressive model （TVAR） parameters and improve speech enhancement performance, a simple method to judge stability of model can be realized through transforming parameters of TVAR model into reflection coefficients of grid filter. When TVAR model signal and reflection coefficients were extended to state vector, Gaussian Particle Filter （GPF） was applied to estimate parameters of TVAR model. Monte Carlo experiments were done to evaluate the enhancement effect of Extended Kalman Filter （EKF）, standard Particle Filter（PF） and GPF, and the simulation results indicate that TVAR model is more efficient in describing the variation characteristic of speech signal ; and PF has stronger filtering ability than Kalman filter, moreover GPF can overcome degeneracy problem of standard PF, resulting in stronger estimation ability, better enhancement and noise filtering effect.