利用函数逼近原理，提出了一种语音信号时间动态规正的新方法，并由此导出了可用于神经网络语音识别的新特征．重点讨论了时间动态规正的原理、新特征的提取方法和性质．实验及理论分析表明，新特征对于神经网络语音识别及语音压缩具有很好的效果 Using the principle of function approximation, a new method of time dynamic correction of speech signal is proposed, and the new features of neural network speech recognition are derived. It focuses on the principles of temporal dynamics and the extraction methods and properties of new features. Experiments and theoretical analysis show that the new features have good effect on speech recognition and speech compression of neural networks