基于清浊音状态空间模型语音增强算法  

Speech Enhancement Based On A Unvoiced-Voiced Model

作　　者：张海云[1] 汪志鸿[1] 杜利民[1] 

机构地区：[1]中科院声学所语音交互技术研究中心

出　　处：《微计算机信息》2006年第09S期293-295,共3页Control & Automation

摘　　要：基于语音状态模型的语音增强算法是当前语音信号处理的研究热点。把通常的LPC语音模型修正后,将得到两个语音模型:时变AR模型、时变双AR模型。但是利用这些模型增强语音时,都没有考虑到语音的清音、浊音区别。为此本文引入了语音清浊音状态空间模型,这种模型在描述语音方面比时变AR模型、时变双AR模型要强,而且物理含义明显。同时在用含噪语音信号预测纯净语音信号时,引入遗忘因子和粒子滤波算法以降低计算复杂性,减小运算量。实验证明,增强后的语音信号信噪比有一定提高,且优于传统的LPC模型。Speech enhancement algorithms based on speech state model are very popular research areas in speech signal processing. If the traditional LPC mode is modified, two speech models are derived： time-varying autoregressive （TVAR） model, time-varying two-autoregressive （Two-AR） model. Utilized these models to enhance speech, the differences between unvoiced and voiced speech are not considered. To solve the problem, a unvoiced-voiced speech model factor is introduced. In the ability of describing speech, the unvoiced-voiced model is better than LPC, TAVR and Two-AR. Besides this , its physical content is more obvious than others. When predicting clean speech through noisy speech, forgetting factor and particle filter are presented to decrease the computation complexity and computation expense. The experimental results show that the SNR of enhanced speech signal is improved to some extent, and the Unvoiced-Voiced Model is more effective than LPC model.

关 键 词：语音增强 语音清浊音状态空间模型 遗忘因子 粒子滤波 

分 类 号：TN912[电子电信—通信与信息系统]