Abstract: In this paper, it proposes a novel speech enhancement method based on real-value discrete Gabor transform, noisy speech is transformed to the joint time-frequency domain by fast real-value discrete Gabor transform in which the Gaussian is used as the transform kernel function. The noise spectral is estimated by the improved algorithm of minimum statics and optimal smoothing, then the speech spectral gain is got by spectral subtraction in the joint time-frequency domain, the clean speech is got by inverse transform by real-value discrete Gabor transform. Experimental results show that this method enhances the segSNR（Segmental Source-to-Noise Ratio） and improves the speech quality compared to other algorithms based on spectral subtraction.

Key words: real-value discrete Gabor transform, noise estimate, spectral subtraction

摘要： 提出了一个基于实值离散Gabor变换的新的谱减法语音增强，采用高斯窗作为综合窗，利用已有的快速实值离散Gabor变换将语音变换到时频域，噪声估计采用改进的最优滤波和最小统计的martin算法，在联合时频域进行谱减得到纯净语音增益，在得到语音增强信号后，利用实值离散Gabor逆变换将其还原输出。实验结果表明，在分段信噪比和语音质量方面均与目前主流谱减法相比均有提高。

关键词: 实值离散Gabor变换, 噪声估计, 谱减法