基于GMM的声音活动检测方法  

A VOICE ACTIVITY DETECTION METHOD BASED ON GMM

作　　者：陈奇川[1] 蔡骏[1] 林茜[1] 

机构地区：[1]厦门大学计算机科学系,福建厦门361005

出　　处：《计算机应用与软件》2009年第2期60-62,75,共4页Computer Applications and Software

基　　金：福建省自然科学基金项目(2006J0043)

摘　　要：为了提高声音活动检测的鲁棒性,提出了一种基于GMM模型的声音活动检测方法。此方法在频谱特征空间中建立背景噪音和语音的GMM模型,然后采用模型匹配的方法对被测信号进行区分。此方法自适应更新GMM模型的参数,使之可以适应环境的变化。实验结果显示该方法在噪音环境中比传统的声音活动检测方法具有更高的准确率。To improve the robustness of the voice activity detection （ VAD）, a GMM-based approach for VAD has been proposed in this paper. With this method, two GMMs are constructed to model the noise and the speech respectively in spectrum feature space, and the signal frames to be detected are discriminated in the way of GMMs matching. This method is designed to self-adapt the GMM parameters＇ updating to accommodate environmental variation. Experimental results show that the proposed method generally performs better in accuracy than tradi tional VAD approaches in noisy environments.

关 键 词：声音活动检测 特征空间 GMM模型 参数自适应 语音识别 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] TP391[自动化与计算机技术—计算机科学与技术]