混合激励线性预测声码器算法的研究

【作者】 赖长庆；

【导师】 刘亚康； 朱学勇；

【作者基本信息】 电子科技大学 ， 信号与信息处理， 2003， 硕士

【摘要】 近年来，随着宽带通信技术的飞速发展，语音通信的带宽占用在有线通信领域似乎已不再成为问题了，但是在无线通信领域，带宽始终是一种宝贵的资源，尤其在军用和保密通信中，语音编码上取得的成果可以迅速带来抗干扰、保密性能和系统容量的提高。另外在语音存储领域，近年来随着各种便携数码录音装置的流行，对高合成语音质量的语音编码算法也提出了迫切的要求。这些需求正是语音编码的原动力所在。经典的线性预测(LPC)声码器具有很高的编码效率，可以极低的码率(800～2400bps)对语音信号进行编码，不幸的是它的合成语音听起来很不自然，常常夹杂着嗡嗡声，重击声或者音调噪声。混合激励(MELP)声码器是近年来提出的一种以经典LPC声码器为基础的性能优良的语音编码方案，对它的研究方兴未艾，现已取得了不少的成果，可以在1.2kbps的码率下取得MOS分为3.0左右的合成语音，并且具有比较强的抗背景噪声的性能。MELP声码器继承了经典LPC声码器编码效率高的特点，并加入了一些新的特征以模仿人的自然语音。MELP声码器采用混合脉冲和噪声激励解决了经典LPC的嗡嗡声的问题；引入了抖动浊音状态以克服音调噪声；利用参数插值、脉冲散布和自适应谱增强等措施提高合成语音的自然度和可懂度；此外还采用了多带激励，使其具有了比较强的抗背景噪声的性能。本文以美国联邦标准2.4kbps－MELP算法为基础，在MATLAB上建立起了分析MELP算法的软件平台，对其性能进行了分析并提出了一些改进的建议；另外还针对MELP算法的特点对其软硬件实现进行了探讨。本文的第二章介绍了MELP声码器模型的原理，对其特征进行了详细的阐述，重点分析了各个特征的本质及其能够对提高合成语音质量起到的作用。第三章详细介绍了MELP声码器的基本算法，对其中采用的一些先进的技术手段如多级矢量量化(MSVQ)、高分辨率基音检测方法(SRPDA)等进行了重点的讲述。另外还对MELP声码器中使用的一些技术进行了实验分析，检验其效能。第四章利用在MATLAB上搭建的分析平台上对语音信号进行了编解码的试验，分析了MELP声码器的各种特征在语音编码中起到的作用。最后针对MELP声码器的特点，对其软硬件实现提出了建议。更多还原

【Abstract】 Recently, with the development of broadband communication, it seems that the band is not a serious problem any more. But in wireless communication field , band is always a kind of rare resource. Especially in military and secret communication, any improvement in speech coding may enhance the system’s performance rapidly. Digital speech’s store is also an important field that requires high quality speech coding algorithm because now all kinds of portable digital recorder is more and more popular. The demand is the power forcing speech coding to progress. Traditionally linear prediction(LPC) vocoders are very efficient, which can encode speech from 800 to 2400bps, but unfortunately, artifacts such as buzzes, thump, and tonal noise always exist in them.Mixed excitation linear prediction(MELP) vocoder is a kind of speech coding algorithm providing superior speech quality under very low rate even 1.2kbps, as well as its capability withstanding strong background noise. MELP vocoders base on LPC vocoders. Furthermore they add some new features to mimic the natural speech. MELP vocoders utilize mixed pulse and noise as the excitation to elimate the buzzes in traditional LPC vocoders, and add a jitter voicing state to overcome the tonal noise. Parameters’ interpolation, adaptive spectrum enhancement and pulse dispersion also are adopted to improve the continuity. The synthetic speech of MELP vocoders sound much more natural and perceivable than the traditional vocoders’.Basing the American federal 2.4kbps MELP algorithm, the analysis platform was established for analyzing the and testing the performance of MELP codec. This article analyzes the capability of the MELP vocoders. Finally some advice are given to realize the vocoder in hardware or software.Chapter 2 introduces the theory of MELP vocodes, and expounds the features in detail. The essence and the function of these features is focused in this chapter. Chapter 3 introduces the basic algorithm of MELP vocoders, and some advanced skills such as muti-stage vector quantization and super resolution pitch detect algorithm. Chapter 4 utilizes the analysis platform to<WP=6>analyze the capability of MELP vocoders and the features’ performance. Finally aiming at the character of MELP algorithm, some advices about realizing it are given.更多还原