基于广义回归神经网络的面罩语音矫正研究

Research on mask speech correction based on generalized regression neural network

为了提高面罩语音的清晰度和可懂度,提出一种基于广义回归神经网络(GRNN)对线谱对(LSP)参数进行非线性建模的面罩语音矫正方法。分别提取正常语音和面罩语音的LSP参数,其次利用LSP参数对GRNN进行训练,得到矫正模型,将面罩语音的LSP参数通过矫正模型进行修正,并将结果作为参数用来合成新的语音。实验结果表明,利用GRNN训练出的矫正模型能够有效地调整面罩语音的LSP参数,在一定程度上能够恢复其频谱分布。

In order to improve the clarity and intelligibility of mask speech,a mask speech correction methodbased on generalized regression neural network(GRNN)for nonlinear modeling of line spectrum pair(LSP)parameters is proposed.The LSP parameters of normal speech and mask speech are extracted respectively,and then used to train GRNN to obtainthe correction model.The LSP parameters of mask speech are modified based on the correction model,and its results are used asparameters for new speech synthesis.The experimental results show that the correction model trained by GRNN can adjust theLSP parameters of the mask speech effectively,and recover the spectral distribution of the mask speech to a certain extent.