以测度作为神经网络输入的客观音质评价研究

Object speech quality assessment by neural network with distance measure as inputs

利用径向基函数网络(RBFN)进行语音质量客观评价,以避免在回归分析中选取具体函数的困难.选取3种距离测度而非语音文件本身作为神经网络的输入,使得网络输入维数大大减小,网络结构大大简化.且对径向基函数网络结构作了修正,使其更便于作音质评价.作者在网络参数和结构学习中采用平滑后的训练集,有效减少了随机因素对客观评测结果的影响,也大大减少了网络结构的复杂性.主客观评价结果的相关性实验中,相关系数达0.96以上,这表明了该方法的可靠性.

In this paper radial-basis function network （RBFN） is used for objective speech quality assessment in order to avoid the difficulty of choose regression function, Three kind of distance measure rather than the speech itself are chosen as the inputs of the neural network, so the dimensions of neural network＇s input is de- creased greatly. Thus the structure of neural network is simplified greatly, The structure of RBFN is also modified so as to convenient for speech quality assessment. Using smoothed training set for learning of parameters and structure of RBFN, the affection of random factors on the result of objective assessment can be reduced effectively, and the neural network＇ s structure can be simplified. The experimental results show that the structure of obtained RBFN is very simple, and the correlation coefficient between the subject seores and object MOS estimate is above 0.96. This shows that the method is reliable.