DBN网络在语音质量评估中的应用

Application of DBN Network in speech quality evaluation

针对传统的语音质量评估中主观MOS分与失真测度拟合形式过于简单,无法表现语音质量与失真距离之间复杂的听觉关系这一缺点,提出了一种基于深度置信网络的语音质量评估方法。该方法首先对语音样本进行预处理,之后提取Mel特征参数,利用深度置信网学习特征参数失真测度与主观MOS分之间的关系,从而完成语音质量的评估。实验结果表明,基于该算法的语音质量评估效果好于传统的评估算法与基于BP网络的评估算法。

In view of the shortcoming of the complex auditory relationship between the subjective MOS and the distortion measure in the traditional speech quality evaluation, a method of speech quality evaluation based on depth confidence net- work is proposed. Firstly, the speech sample is preproeessed, the mel feature parameter is extracted, and the relationship between the characteristic parameter distortion measure and the subjective MoS is studied by using the depth confidence net, thus the evaluation of speech quality is completed. The experimental results show-that the proposed algorithm is better than the traditional evaluation algorithm and the evaluation algorithm based on BP network.