掩蔽模型对语音增强效果影响的研究

Speech Enhancement Research Under Different Masking Models

对两种最常用的噪声估计算法——VAD噪声检测估计和基于最小值统计特性的噪声估计法,运用不同的掩蔽模型计算了增强前后的信噪比SNR。结果表明:掩蔽效应的运用对语音增强效果的影响很大,在不使用掩蔽效应的情况下,最优最小统计值法语音增强的SNR增加很多,从而可能引入更大的音乐噪声。但无论对VAD噪声检测估计法和基于最小值统计特性的噪声估计法,掩蔽阈值偏移量的少量变化对SNR的影响都不是很大,其原因主要是语音信号本身的幅度较大。

The SNRs（Signal-to-Noise Ratio） of the two most common noise estimation algorithms, VAD（Voice Activity Detection） and statistical characteristics of minimum value based estimation of noise are calculated with different masking models before and after the voice enhanced. The results show that the usage of masking effect has a big impact on speech enhancement. If masking effect is not considered, the SNR of speech enhancement with optimal statistical minimum value method increases much, which could introduce greater music noise. But a small change of the masking threshold offset has a little impact on the SNRs of the two algorithms, which may be mainly due to the large magnitude of the voice signals.