基于神经网络由语音预测视位参数

Predicting Viseme Parameters from Speech Based on Neural Network

语音是由多个发音器官共同作用产生的,发音器官动作与语音之间有着内在的必然联系.研究了利用神经网络预测视位参数中的选择语音参数、确定输入语音时域范围、优化神经网络结构等因素.实验结果表明,线性预测参数加短时能量优于其他语音参数,前向协同发音较后向协同发音影响更大,反馈对前馈神经网络的性能有所改善.考虑到实验采用的是任意连续语流,均方误差约为0.0114的实验结果还是很有吸引力的.

Speech is produced by co-operation of all speech organs, and there are inherent relations between speech and movement of speech organs. To predict viseme parameters from speech using neural network, input speech parameters selection, time domain and structure of neural network were studied. Experiment results show that LPC coefficient plus short time energy are superior to other speech parameters, forward co-articulation is more server than backward co-articulation, and a delay feedback can improve the forward neural network performance. Considering experiments were based on unlimited vocabulary and continuous speech, the 0.0114 mean square error (MSE) is quite promising.