融合几何相位信息的噪声抑制算法研究

Research on Noise Suppression Algorithm Combining Geometric Phase Information

摘要: 本文针对语音信号处理中因估计的纯净语音幅度误差会引入噪声，提出了一种通过融合几何相位信息，使幅度失真与先验信噪比和后验信噪比建立数学关系的噪声抑制算法，具有较高的可行性，提高了语音的可懂度。现有数据表明，当被处理的语音只包含衰减失真和小于6分贝的放大失真时，对语音质量影响较小，需要对增强之后的幅度谱进行简单限制。同时，需要对于放大失真超过6分贝的语音区域采取不同的约束条件，会对提高语音质量有较大的帮助。但目前的研究大多采用估计的纯净语音与原是纯净语音谱幅度进行直接对比，现实中可行性较差。本文主要提出一种融合几何相位信息，把幅度产生的失真与先验信噪比和后验信噪比建立数学关系，然后根据此先验信噪比和后验信噪比的判定条件对幅度放大失真超过6分贝以及衰减失真和小于6分贝的放大失真同时进行不同的限制约束。实验结果表明，在对数谱距离(LSD)，分段信噪比(SegSNR)，短时清晰度客观测度(STOI)等评测条件下，语音增强效果有明显的提高。总而言之，本文提出的算法简单易行，能有效地提升语音质量，提高语音的可懂度。

Abstract: In this paper, a noise suppression algorithm is proposed to establish a mathematical relationship between amplitude distortion and prior signal-to-noise ratio and post signal-to-noise ratio (SNR) in speech signal processing, due to the introduction of noise from the estimated pure speech am-plitude error. Speech enhancement process can be expressed by gain function, but the gain function will introduce two types of speech distortion: amplifying distortion and attenuating distortion of speech signal amplitude. In fact, when the processed speech contains only attenuation distortion and amplification distortion less than 6 dB, the influence on speech quality is relatively small. It is necessary to limit the amplitude spectrum after enhancement. At the same time, different constraints for speech regions with amplification distortion exceeding 6 dB will help to improve speech quality. In this paper, we propose a fusion of geometric phase information to establish the mathematical relationship between amplitude distortion and prior signal-to-noise ratio and pos-teriori signal-to-noise ratio. According to the criteria of priori SNR and post signal-to-noise ratio, the restriction of amplify distortion exceeding 6 decibels, attenuation distortion and less than 6 decibels is given at the same time. Experimental results show that the algorithm proposed in this paper is simple and feasible, and it can effectively improve speech quality and improve speech in-telligibility.