【作者】 李靓；

【导师】 鲍长春；

【作者基本信息】 北京工业大学 ， 模式识别与智能系统， 2005， 博士

【摘要】 语音编码是数字语音通信系统中最重要、最基本的核心功能之一,用以压缩语音信号的数字表示而使表达这些信号所需的比特数最小。就目前的语音编码现状而言,8kb/s 以上的技术已经标准化和产品化,4kb/s 的国际标准也正在制定中。然而,随着个人计算机和国际互联网的快速发展,在移动通信、多媒体通信以及计算机网络通信等领域还将需要具有通信质量的2kb/s 语音压缩技术。因此,研究如何在2kb/s 获得具有通信质量的重建语音是语音编码中一个十分重要的课题。本论文在特征波形内插(CWI)编码算法的基础上提出了一种高质量的2kb/s 波形内插(WI)语音编码算法,主要研究成果包括: (1) 研究了两种语音线谱频率(LSF)参数的矢量量化方法。利用LSF 参数的有序性,提出了一种基于二阶时间分解(TD)模型的高效LSF 参数量化方法。实验结果表明在平均编码速率为500b/s 时,该量化方法可以获得一个相对较低的平均谱失真和相对较高的算法平均总延时。为了保证较低的算法延时,只能对LSF 参数逐帧进行矢量量化。基于此,本文利用LSF 参数的帧内和帧间相关性,提出了一种编码速率为800b/s 的LSF 参数的预测式瞬时联合多级分裂矢量量化(PSJ_MSVQ)方案。该量化方案不仅保持了较高的量化性能,而且降低了码字搜索复杂度和码字存储容量。(2) 研究了一种用于WI 语音编码模型的基音检测算法。利用小波变换的优良性能和时域波形类似性方法,提出了一种基于二进小波变换和归一化自相关函数的基音周期检测算法(DWT_NACF_PDA),与G.729 语音编码标准中所采用的仅基于归一化自相关函数的基音检测算法相比,该算法在检测性能与计算复杂度上具有优势。接下来,以该算法为基础提出了一种基于二进小波变换和归一化互相关函数的基音检测算法(DWT_NCCF_PDA),并将其用于WI 语音编码模型的基音检测。实验分析表明,DWT_NCCF_PDA 的性能与基于NCCF 的基音检测算法(NCCF_PDA)相当,但优于DWT_NACF_PDA。主观听力测试结果表明DWT_NCCF_PDA 为确保2kb/s WI 语音编码算法产生高质量的重建语音奠定了基础。(3) 提出了特征波形(CW)及其功率的量化方案。为提高编码效率,编码时将功率归一化的CW 分解为慢渐变波形(SEW)和快渐变波形(REW)。通过利用SEW 和REW 的不同感性特点、合成分析(A-b-S)技术和感觉加权技术,提更多还原

【Abstract】 Speech coding, which generally refers to the process of reducing the number of bits required for adequately representing speech signals in digital form, is one of the most important and fundamental function in modern digital speech communication systems. At present, speech coding technologies at above 8kb/s have already been standardized and achieved toll-quality. There is an ongoing standardization effort conducted by ITU-T for a 4kb/s speech coder with toll-quality. But with the rapid advancement of personal computer and international network, it’s urgent to need a kind of speech coder at 2kb/s with communication quality in the area of mobile communications, multimedia communications and computer communications. Therefore, how to achieve communication quality at 2kb/s is a very important research issue. This dissertation presents a high quality Waveform Interpolation (WI) speech coding algorithm at 2kb/s based on the Characteristic Waveform Interpolation (CWI) algorithm. The main research results are as follows: (1) Two kinds of vector quantization (VQ) methods of line spectrum frequency (LSF) parameters are proposed. An efficient quantization method of LSF parameters is proposed based on second order temporal decomposition (TD) model by using the LSF ordering property. Experimental results show that a relative low spectral distortion and high average algorithm delay can be obtained at below 500b/s. So we quantize LSF parameters each frame to overcome the shortcoming of high delay. The Predictive Simultaneous Joint Muli-stage Split Vector Quantization (PSJ_MSVQ) scheme at 800b/s is designed by using the correlation of LSF parameters between frames and inter frames. The vector quantizer achieves an average spectral distortion of less than 1dB at 20 bits/frame with low computational complexity and storage requirements. (2) An algorithm of pitch detection for WI coding model is presented. Firstly, a pitch detection algorithm based on the Dyadic Wavelet Transform and the Normalized AutoCorrelation Function (DWT_NACF_PDA) is proposed. Compared with the normalized autocorrelation-based pitch detection algorithm in ITU-T G.729 standard, the DWT_NACF_PDA has a lower pitch-estimation error and computational complexity. After that, a pitch detection algorithm based on the Dyadic Wavelet Transform and the Normalized Cross-Correlation Function (DWT_NCCF_PDA) is更多还原