空间噪声下语音信号端点检测仿真研究

The Space under the Noise Speech Signal Endpoint Detection Simulation Research

在空间噪声下对语音信号端点进行准确检测,可提高语音识别效果。语音信号端点检测时,需要在清音与背景噪声分离的基础上,对语音信号的特征量进行提取,而传统方法不能对清音和背景噪声分离进行自适应分离,提取的语音信号特征中仍混有大量噪声,降低了语音信号断点检测的准确度,存在信号检测鲁棒性差的问题。提出一种改进Fisher准则的空间噪声下语音信号端点检测方法。上述方法先融合于梅尔倒谱理论MFCC将清音信号和背景噪声视为两类分类问题,利用Fisher准则获取最佳投影方向,利用投影方向对清音与背景噪声进行自适应分离,提取语音信号的特征量,作为BP神经网络输入特征向量进行训练,利用粒子群算法对神经网络输入特征向量进行优化,完成空间噪声下语音信号端点检测。仿真结果表明,改进方法再空间噪声下可以对语音信号端点的位置进行准确地检测。

In this paper, we proposed a method for endpoint detection of speech signal under spatial noise based on modified Fisher rule. Firstly, integrated with Mel Frequency Cepstrum Coefficient （MFCC） , unvoiced sound sig- nal and background noise were regarded as two kinds of classification problems. Fisher criterion was used to obtain optimum projection direction. Then, projection direction was used to carry out self-adaption separation for the un- voiced sound and background noise. Characteristic quantity of speech signal was extracted as input feature vector of BP neural network for training. Moreover, particle swarm algorithm was used to optimize input feature vector of neural network, Finally, the endpoint detection was completed. Following conclusions can be drawn from simulation results that the modified method can detect endpoint position of speech signal precisely under spatial noise.