基于HMM与SVM的语音活动检测  

Voice Activity Detection Based on HMM and SVM

作　　者：肖佳林[1] 赵聿晴 王英[1] 

机构地区：[1]湖南大学信息科学与工程学院,长沙410082

出　　处：《计算机工程》2014年第1期203-208,共6页Computer Engineering

摘　　要：工程机械强噪音环境下的噪声源较多,导致电话语音通话无法进行,且强噪声造成无效数据占用带宽。为此,提出基于隐马尔科夫模型(HMM)和支持向量机(SVM)的语音活动检测算法。该算法将提取的美尔频率倒谱系数特征向量输入到HMM识别器中,并通过Viterbi算法得到N维最佳识别结果,将其转换为SVM特征向量输入到SVM分类器中进行分类判别,得到判决结果。实验结果表明,该算法在机械工作噪音的情况下,语音检测率较静态统计类算法平均提高9%,比小波支持向量机方法提高11%,在驾驶室噪音的情况下比小波SVM方法有较小幅度的提高,但其增长速度较快,且比传统的统计类算法提高9%。In the construction machinery strong noise environment, there are lots of noise source, speech often is covered by the machine＇s noise, calls often can not success and waste bandwidth. To solve this problem, a new Voice Activity Detection（VAD） algorithm based on Hidden Markov Model（HMM） and Support Vector Machine（SVM）（HMM/SVM-VAD） is proposed. This algorithm inputs the Mel Frequency Cepstrum Coefficient（MFCC） into the HMM, and gets the N-best Recognition results by using Viterbi algorithm, and transforms the N-best recognition results to SVM feature vector. It uses the SVM to get the classification results. Experimental results show that HMM/SVM-VAD is better than the traditional statistical algorithm and wavelet SVM algorithm. In the case of machine work noise, new method improves by the average of 9% than the static statistical algorithms, improves 11% than the wavelet SVM algorithm, in the case of cab noise, the new method improves small, but it has faster growth, and improves by 9% than the traditional statistical algorithm.

关 键 词：美尔频率倒谱系数 隐马尔科夫模型 支持向量机 语音活动检测 核函数 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]