Abstract: With the development of modern technology，car has more and more electrical devices.But it is dangerous to driver to operate these devices with hands leaving steering wheel.So speech recognition is solution to this problem，but general speech recognition system is sensitive to background noise，the reason is that the contaminated speech is difficult to find the endpoint accurately，So some technology must be developed.Speech enhancement is a popular technology.This paper describes a new band partitioning mel frequency spectral entropy endpoint technology，with these technology the start point and end point can be found accurately.After getting endpoint，spectral subtract can be used to enhance speech signal，then speech recognition can work well.

Key words: speech enchantment, endpoint detection, band partitioning mel frequency spectral entropy

摘要： 随着现代科技的发展，车载电子设备越来越多，但是如果在驾驶中司机的手离开方向盘去操作这些电子设备却是很危险的。对于这个问题的一个解决方法就是这些设备都采用语音识别作为它们的输入接口。通常的语音识别系统在噪声环境下的识别率是很低的，造成识别率下降的一个重要原因就是端点检测的不准确，因此必须发展一些技术来解决这个问题。提出了一种基于子带美尔谱熵的算法，这种算法可以准确地检测到语音的起始点和结束点，得到语音端点就可以利用谱减的方法来进行语音增强，经过增强后的语音信号就可以用普通的语音识别系统进行识别。

关键词: 语音增强, 端点检测, 子带美尔谱熵