基于谱特征自适应估计的激光相干语音探测信号增强方法

Enhancement of speech detected by laser coherent detection method based on spectral feature adaptation

针对激光相干语音探测引入的缓变宽带背景噪声和测振目标造成的信道作用,本文提出了基于分析-重合成框架,针对特定说话人的语音增强方法。该方法首先提取观测语音特征:基音频率、浊音概率、MCEP系数,其中,MCEP系数是能够表示谱包络形状的谱包络特征。通过观测语音的谱包络特征和预训练的对应说话人语音谱包络特征GMM,估计对应的纯净语音谱包络特征,再与观测语音的基音频率和浊音概率一起重新合成语音信号,实现语音增强。噪声和信道参数的估计通过最大化观测语音谱包络特征后验概率的自适应估计实现,然后通过MMSE估计得到纯净语音谱包络特征的估计值。合成信号实验和实际信号采集实验检验了本文提出算法在激光相干语音探测场景下的去噪和均衡能力。

To address the issue of slowly varying broadband background noise and channel effects caused by vibration of the target in laser-coherent speech detection,this paper proposes a speech enhancement method for specific speakers based on an analysis-resynthesis framework.This method first extracts the features from the observed signal:pitch,voiced speech probability,and MCEP coefficients,where MCEP coefficients represent the spectral envelope features which can capture the shape of the spectral envelope.A GMM trained by speech features of the corresponding speaker is used to help estimate the spectral envelope features of the clean speech from the spectral envelope features of the observed speech,and then the speech signal is resynthesized by combining it with pitch and voiced speech probability estimated from the observed speech to achieve speech enhancement.The estimation of noise and channel parameters is achieved by adaptation,which maximize the posterior probability of the observed speech′s spectral envelope features,and then the estimation of the clean speech spectral envelope features is obtained by MMSE estimation.Both synthesized signal experiments and actual signal acquisition experiments verify the denoising and equalization capabilities of the algorithm in laser coherent speech detection scenarios.