基于Laplace-Gauss模型和简化相位判别的离散余弦变换域语音增强

Speech enhancement based on Laplacian-Gaussian model and simplified phase discrimination in Discrete Cosine Transform domain

分析了理想情况下离散余弦变换域中语音信号增益,先验信噪比及后验信噪比之间的关系,用实际数据获得了各种信噪比下增益范围的统计特性。基于语音呈Laplace分布、噪声呈Gauss分布的模型,推导了具有相位特性的增益及先验信噪比的估计公式,通过合理性分析得到了简化的相位判别准则。实验结果表明,在高斯白噪声和F16飞机噪声情况下,简化的相位判别可使低信噪比下的语音增强系统的性能得到较大的改善。

The relationship among the gain, a priori SNR and a posteriori SNR of noisy speech signal is analyzed in Discrete Cosine Transform domain. The statistical properties of the gains＇ range are obtained from the true data at several SNRs. The formulae of gain and a priori SNR dependent on the phase are derived employing the Laplacian model for speech and Gaussian model for noise. The simplification of the phase discrimination is obtained by analyzing the reasonability of the gain and a priori SNR. The experiment results indicate that the simplified phase discrimination achieves much improvement on the performance of speech enhancement at low SNR in white Gaussian noise and F16 jet noise.