基于AR-HMM在线能量调整的语音增强方法

发布时间：2018-01-07 19:28

本文关键词：基于AR-HMM在线能量调整的语音增强方法　出处：《电子学报》2014年10期 　论文类型：期刊论文

【摘要】：针对单通道语音增强技术对非平稳噪声的跟踪不准确、噪声抑制效果较差的问题,本文提出一种基于在线能量调整的语音增强方法.该方法以归一化临界带能量为特征,采用高斯混合模型对背景噪声进行分类,利用对应类型噪声的自回归隐马尔可夫模型(Auto-Regressive Hidden Markov Model,AR-HMM)和纯净语音的AR-HMM,在最小均方误差准则下估计语音和噪声的功率谱.考虑到非平稳环境中训练集和测试集的差异性,需在线调整语音模型和噪声模型中的能量,语音模型的能量调整采用迭代的期望最大化算法;噪声模型的能量调整则利用的是模型训练过程中的能量重估方法,并以最小值控制的递归平均算法确定噪声能量调整的初始值.在ITU-T G.160标准下对算法进行性能测试,测试结果表明,本文方法对非平稳噪声的跟踪效果较好,对噪声衰减量较大,收敛时间较短.
[Abstract]:For single channel speech enhancement technique for tracking nonstationary noise accurately, noise suppression effect is poor, this paper proposes a speech enhancement method based on online energy adjustment. The method based on normalization of critical band energy as the feature, using the Gauss mixture model to classify the background noise, using autoregressive hidden Markov model corresponding to the type of noise (the Auto-Regressive Hidden Markov Model, AR-HMM) and pure voice AR-HMM, in the minimum mean square error criterion to estimate the power spectrum of speech and noise. Considering the differences in the training set and test in nonstationary environments, to adjust online speech model and noise model of energy, energy adjusted by expected speech model the maximum iteration algorithm; noise model is adjusted using energy is the energy estimation method in the process of training, and control the delivery to the minimum The normalized mean algorithm is used to determine the initial value of noise energy adjustment. The performance of the algorithm is tested under the ITU-T G.160 standard. The test results show that the tracking effect of the method is good for non-stationary noise, and the noise attenuation is large and the convergence time is short.

【作者单位】：北京工业大学电子信息与控制工程学院语音与音频信号处理实验室;
【基金】：国家自然科学基金(No.61072089) 北京市教育委员会科技发展计划重点项目(No.KZ201110005005)
【分类号】：TN912.3
【正文快照】： 1引言现实中,语音信号不可避免地会受到噪声的污染,因此语音增强技术一直是语音信号处理的重要内容.自上世纪70年代以来,语音增强经历40多年的发展,单通道语音增强算法已形成较为经典的几种算法.根据是否使用参数描述模型可以分为无参数模型法和有参数模型法,无参数模型法包括

【相似文献】