摘要
研究表明超高斯分布更加贴近语音信号的实际分布,然而语音信号很难用单一的概率密度函数准确描述,针对这一情况,提出了一种用超高斯混合模型对语音信号幅度谱建模的新方法,并推导了基于此模型的幅度谱最小均方误差估的估计式。仿真结果表明:与传统的短时谱估计算法相比,该算法不仅能够进一步提高增强语音的信噪比,而且可以有效减小增强语音的失真度,提高增强语音的主观感知质量。
Recent research indicates that the speech spectral amplitude distributions could be fairly described with super-Gaussian probability density function. However, the complexity of speech signal determines that the distribution statistics of speech signal could not be well described by single simple function. Thus a super-Gaussian mixture model for speech spectral amplitude is proposed, and with this model, a minimum mean-square error （MMSE） estimator for speech signals spectral amplitude is derived. The simulation results show that this algorithm based on Gaussian and super-Gaussian speech model could achieve better noise suppression and lower speech distortion as compared with the conventional short-time spectral amplitude estimation algorithm.
出处
《通信技术》
2013年第6期137-141,共5页
Communications Technology
关键词
语音增强
超高斯混合模型
最小均方误差
speech enhancement
super-Gaussian mixture model
minimum mean-square error
