一种基于统计模型的改进谱减降噪算法

An improved statistical model-based spectral subtraction method for noisy speech

提出了基于语音和噪声的傅里叶系数服从统计模型分布的假设,将基于统计模型的信噪比更新和噪声更新的方法应用于谱减法,试图解决传统谱减法中存在的音乐噪声和语音失真的问题。将提出算法与多通道谱减法和基于对数的最小均方幅度谱估计方法进行客观评价分析。利用频率加权分段信噪比评价方法、语音质量感知评价及综合质量测量等3种指标进行去噪效果评价。结果表明,所提出的基于统计模型的降噪算法效果优于MBSS,且接近Log-MMSE。

A hypothesis that the Fourier transform coefficients of speech and noise are subject to different statistical model distributions is introduced. And the statistical model-based SNR （Signal to Noise Ratio） update and noise update method is proposed to solve the problem that traditional speech spectral subtraction method may introduce musical noise and speech distortion. Subjective evaluation is carried out to compare the performances of the proposed spectral sub- traction method with two other noise reduction methods called MBSS （Multi-band Spectral Subtraction） and Log-MMSE （Minimum Mean-Square Error Log-spectral amplitude estimator）. Three evaluation tools： Frequency Weighted Segmental SNR （FWSegSNR）, Perceptual Evaluation of Speech Quality （PESQ） and Composite Quality Measure （CQM） are uti/ized. Evaluation results show that the noise reduction performance of the proposed method is better than MBSS, and close to Log-MMSE.