基于序贯蒙特卡罗统计的语音增强算法
文献类型：学位论文
| 作者 | 张海云 |
| 学位类别 | 博士 |
| 答辩日期 | 2005 |
| 授予单位 | 中国科学院声学研究所 |
| 授予地点 | 中国科学院声学研究所 |
| 关键词 | 语音增强 序贯蒙特卡罗统计 卡尔曼滤波器 TVAR模型 时变双AR模型 语音清音浊音状态空间模型 |
| 其他题名 | Speech Enhancement Algorithms Based On Sequential Monte Carlo Method |
| 中文摘要 | 实际环境中的语音信号经常受到其它信号或者噪声的污染，降低人或机器对语音的感知和识别的效能。因此，语音增强作为语音信号处理的一种特殊功能，成为国际上研究的热点。本文以信号处理的状态空间模型理论为基础，结合序贯蒙特卡罗方法，开展单通道语音增强算法的研究，主要创新贡献如下：（l）在TVAR（Time-varyingAutoregresive）模型的基础上，分别采用了KPF（KalmanParticleFilter）、UKF（UllscentedKalmanFilter）、UPF（UnscentedParticleFilter）三种序贯蒙特卡罗滤波器，从含噪语音信号中估计语音分量，实验表明，KPF、UKF、UPF三种滤波器在增强语音方面有着一定的作用，优于卡尔曼滤波，并且UPF滤波器的滤波性能高于KPF和UKF滤波器；（2）在非平稳噪声环境下，提出基于时变双AR模型的序贯蒙特一卜罗语音增强算法，将后验概率（纯净语音信号｝含噪语音信号）的计算分解成卡尔曼滤波和序贯蒙特卡罗统计两个处理过程，大大降低计算复杂度。测试表明，本文方法不仅有效降低运算复杂度，而且处理之后的语音质量也得到明显改善；（3）语音清浊音状态空间模型的TVAR模型，对语音清浊音进行区别处理，同时在对后验概率（纯净语音信号｝含噪语音信号）简化，计算中引入了遗忘因子，更加有效进行序贯蒙特卡罗语音增强。数值模拟表明，基于语音清浊音状态空间模型的增强算法处理后的信噪比，比基于LPC模型的普通卡尔曼滤波算法提高2-3dB。 |
| 英文摘要 | In practice, Speech signal is often contaminated by other signals or noises. So the perception ability of human or machine is degraded seriously. For the reasons, as a special ability of speech signal processing, speech enhancement becomes a very hot research area in international speech research projects. Based on the state space theory in signal processing, the thesis, which combines sequential Monte Carlo method, develops research work in speech enhancement algorithms of single channel. The main innovative characteristics of this dissertation is listed below: Based on TVAR (Time-varying Autoregresive) speech model, the thesis ultilzes there sequential Monte Carlo filters respectively, which are Kalman Particle Filter, Unscented Kalman Filter, and Unscented Particle Filter, to extract pure speech signal. The experiments show KPF, UKF and UPF have good effects on enhancing speech signal and precede Kalman filter. Moreover, the performance of UPF is better than KPF and UKF. In non-stationary noise environment, the thesis develops a speech enhancement algorithm based on time-varying AR model and sequential Monte Carlo method. And when computing poster probability, p{pure speech \ noisy speech), it is divided into two steps: Kalman filter, particle filter. After testing, it is found that the division can reduce the computation cost efficiently, and the quality of enhanced speech signal is improved to some extent. Based on speech state space model of surd and sonant, the thesis presents an algorithm to process surd and sonant differently. In order to simplify computation of p(pure speech \ noisy speech), forgetting factor is introduced to utilize sequential Monte Carlo method efficiently to enhance speech signal efficiently. Compared with Kalman filter based on LPC model, numerical simulation proves the sequential Monte Carlo algorithm based on the model improves 2~3dB. |
| 语种 | 中文 |
| 公开日期 | 2011-05-07 |
| 页码 | 89 |
| 源URL | [http://159.226.59.140/handle/311008/1036]  |
| 专题 | 声学研究所_声学所博硕士学位论文_1981-2009博硕士学位论文 |
| 推荐引用方式 GB/T 7714 | 张海云. 基于序贯蒙特卡罗统计的语音增强算法[D]. 中国科学院声学研究所. 中国科学院声学研究所. 2005. |
入库方式： OAI收割
来源：声学研究所
浏览0
下载0
收藏0
其他版本
除非特别说明，本系统中所有内容都受版权保护，并保留所有权利。