低信噪比下基于短时谱估计的语音增强

Speech enhancement based on short-time spectral amplitude estimates in low SNR

针对在弱语音信号和低输入信噪比(SNR)情况下,基于短时谱估计的语音增强算法性能下降的问题,提出了一种结合软判决信息和人耳听觉掩蔽效应的短时谱估计算法。该算法在最小均方误差准则下引入语音存在的概率,得到软判决修正的增益函数,然后利用掩蔽门限不断地调整增益函数,进而调整噪声的抑制程度,保护微弱的语音信号,减少语音谱的失真。客观测试和主观试听表明,该算法在信噪比增益以及语音的可懂度、自然度方面都优于传统的最小均方误差估计算法。

This paper presents a short-time spectral amplitude estimation approach based on soft-decision information and the masking properties of the human auditory system to improve speech enhancement of weak speech signals in low signal to noise ratio （SNR） input environments. The approach introduces the probability of speech presence into the minimum mean square error （MMSE） estimate to get a soft-decision modified gain function. Then, a masking threshold is used to continually adjust the MMSE estimate gain function which adjusts the noise reduction. This approach significantly reduces attenuation of weak speech signals to reduce spectral distortion of the enhanced speech. Objective measurements combined with informal subjective listening tests show that the algorithm is superior to the traditional MMSE estimate method, not only in SNR gains, but also in speech intelligibility and naturalness.