自适应谱减法实时语音增强

Real-time Speech Enhancement by Adaptive Spectral Subtraction Method

本文针对非平稳噪声和强背景噪声下声音信号难以提取的实际问题,提出了一种自适应谱减算法。该算法设计了动态阈值,给出了噪声谱与纯净语音谱预估迭代更新机制与具体实施方案。实验仿真表明,该算法能有效地去噪滤波,显著地提高语音识别系统性能与可懂度,且在不同的噪声环境和信噪比条件下具有鲁棒性。本算法复杂度低,计算代价小,实时性强,简洁易实现,做到了有效性与实时性双满足。

This paper deals with the practical difficulty of extracting the audio signal in the case of non-stationary noise and strong background noise by proposing an adaptive spectral reduction algorithm. A dynamic threshold algorithm is devised. Iterative update mechanism and the specific implementation are contrived in the clean speech spectrum and noise spectrum estimating. Simulation experiments show that the algorithm can effectively de-noising filter, significantly improving the intelligibility of speech recognition system performance and read. The method is robust in different noise environments and SNR. The algorithm enjoys low complexity, small computational cost, strong real timeliness and easy implementation. Both effectiveness and timeliness are being met.