论文部分内容阅读
提出了基于神经网络的语音谱失真测度概念。利用前向神经网络，包括多层感知器和径向基函数网络，对多维非线性函数的逼近原理，使得谱失真测度函数具备了表现人耳听觉系统的主观感知行为的能力。结合语音质量客观评价应用，我们以在大量的失真条件下得到的主观评价结果作为期望值对该网络进行训练。统计相关分析表明，基于神经网络谱失真测度的客观评价方法的主客观评价的相关性，较之传统欧氏距离以及加权欧氏距离都有了显著的提高，并具有更高的鲁棒性．该方法还具有技术独立性．
The concept of speech spectral distortion measurement based on neural network is proposed. Using the forward neural network, including multi-layer perceptrons and radial basis function networks, the approximation principle of multidimensional nonlinear functions enables the spectral distortion measure function to possess the subjective perception of the human auditory system. In combination with the objective evaluation of voice quality, we train this network with subjective evaluation results obtained under a great deal of distortion as the expectation. Statistical correlation analysis shows that the subjective and objective evaluation based on neural network spectral distortion measure has a significantly higher subjective and objective evaluation than the traditional Euclidean distance and weighted Euclidean distance, and has a higher robustness. This method is also technically independent.