语音端点检测的仿真研究

Simulation of Speech Endpoint Detection

语音端点检测在语音处理中占有十分重要的地位。介绍了短时平均能量法、短时平均过零率法和短时能零积法三种语音端点检测法。在低信噪比环境下,利用上述方法进行语音端点检测,检测效果将大大降低。功率谱减法是一种传统的降噪方法。但是功率谱减法采用固定的无音片段作为噪声样本容易引起误差。为此,笔者综合短时能零积法和功率谱减法,提出了加权功率谱减法,并用于语音端点检测。详细地给出了加权功率谱减法的实现步骤。实验结果表明该方法具有良好的性能。

Speech endpoint detection is very important in speech processing. Short-time average energy, short-time average zero-crossing rate and short-time energy-zero-product are introduced. Detect effect will be much lower with these methods in low signal noise ratio environment. Power spectral subtraction is a traditional method for decreasing noise which easily leads to error by regard fixed silence fragment as noise sample. For the purpose, short-time energy-zero-product with power spectral subtraction was synthesized, and then a weighted power spectral subtraction was proposed for speech endpoint detection. The implementation procedure was given in detail. The experiment result shows that the method has better performance.