论文部分内容阅读
基于声信道特点和回声状态神经网络建模,提出了一种通过抑制环境回声而相应增强目标语音的信号处理方法.仿真实验表明,对应于模型最好的泛化能力,其储备池规模(N)及其稀疏连接度(p)的N×p取值(为储备池中互相连接的神经元数量)是极值;其训练数据量(即足够的训练时间)存在一个下限值.训练建模后,该模型不仅达到通过抑制环境回声而相应增强输出目标语音信号的目的,而且麦克风接收信道改变时,也保持有效的处理效果.
Based on the characteristics of the acoustic channel and the neural network model of echo state, a signal processing method is proposed to enhance the target speech by suppressing the environmental echo.The simulation results show that, corresponding to the best generalization ability of the model, the reserve pool size (N ) And its sparse degree of connectivity (p) (the number of interconnected neurons in the pool) is extreme; there is a lower limit for the amount of training data (ie enough training time) The model not only achieves the purpose of enhancing the output target voice signal by suppressing the environmental echo, but also maintains the effective processing effect when the receiving channel of the microphone is changed.