一种新的用于语音识别的噪音估计及非线性谱相减方法

提出了一种新的噪音估计及非线性谱相减方法，通常的噪音估计一般基于语音检测方法，取噪音段的谱平均作为噪音谱的估计，该方法在信噪比较低时性能下降严重。提出的基于能量聚类的噪音谱估计方法，不依赖于语音检测直接估计噪音谱，提高了噪音谱估计的精度。还在一般非线性谱相减方法的基础上提出了改进的谱相减方法，该方法根据单个mel滤波器频带内局部的信噪比，来决定该频段内非线性谱相减的多少，细化了非线性差谱的额度，在有效抑制噪声的同时减少了语音谱的失真。

A New Method of Noise Estimation and Non-linear Spectral Subtraction Used in Speech Recognition

A new noise estimation and non-linear spectral subtraction algorithm is proposed in this article. The usual noise estimation is based on voice activity detection (VAD) deteriorates quickly when the SNR is low. This article proposed a nes noise estimation which is based on energy clustering and is independent of the VAD. It also puts forward an improved non-linear spectral subtraction algorithm. The new spectral subtraction algorithm determines the over subtraction factor according to the local SNR in one mel channel. This method reduces the speech spectral distortion in the mean time of effectively restraining the noise.