基于会议电话中的实时语音降噪算法研究

Speech Real Time Noise Reduction Algorithm in Conferencetelephone

在兼顾降噪性能和功耗的基础上,提出一种基于多特征估计的实时语音降噪算法。算法首先将输入信号进行快速傅里叶变化,并根据概率密度进行噪声估计;然后分别计算输入信号的3个特征频谱平坦度、频谱差异度和似然值,并将特征值映射到激活函数上求出语音概率;最后根据该概率来更新滤波器系数,从而获得去噪的信号。与改进谱减法、自适应维纳滤波法和基于调制深度的算法的实验对比显示,提出算法的信噪比改善情况明显好于其他算法,最大改善幅度达到16 d B左右。此外,主观语音质量评估的效果也高于其他算法。

A real-time speech noise reduction algorithm based on multi-feature estimation is proposed based on the performance of noise reduction and power consumption.The input signal is firstly processed by Fast Fourier transform to estimate noise based on the probability density.Then,the input signal is further to be calculated the three characteristics included the smoothness,spectrum difference and likelihood value,which are mapped to the activation function to find the speech probability.Finally,the filter coefficients are updated according to the probability to obtain the denoising signal.Compared with the improved spectral subtraction,adaptive Wiener filtering method and the modulation depth algorithm,the proposed method shows the signal-to-noise ratio is maximum to 16 dB and obviously superior to other algorithms.In addition,the method shows more effective than other algorithms in subjective speech quality assessment application.