声码器激励参数的描述模型  

A model for describing the excitation parameters of vocoder

作　　者：贺天宏[1] 张建伟[1] 崔慧娟[1] 唐昆[1] 

机构地区：[1]清华大学电子工程系微波与数字通信国家重点实验室,北京100084

出　　处：《电声技术》2003年第4期52-55,共4页Audio Engineering

基　　金：国家自然科学基金项目(69972020);国家重点基础研究资助项目(G1998030406).

摘　　要：在低速率声码器中,对浊音激励的描述直接影响了重建语音的质量。为提高音质,提出了一种两级浊音激励谱形状描述模型。在混合激励线性预测(MELP)声码器中浊音激励谱幅度峰值描述模型的基础上,增加了对整体谱的表述,同时给出了相应的矢量量化方法。测试表明此模型可使重建语音的音质得以明显改善,主观自然度测试结果达到了63%,对男声更为突出。此模型既保持了在激励谱低频带上描述准确的优点,又大大提高了对全频带描述的准确性,可应用于多种声码器中。In low-rate speech coding,the characterization of voiced excitation directly affects the quality of the decoded speech.In order to improve the quality,this paper presents a two-stage residual spectral shape description model.It is based on the voiced excitation spectral description model used in Mixed Excitation Linear Prediction(MELP)vocoder.Also,the corresponding two-stage vector quanti-zation technique is provided to quantize the excitation spectrum envelope.The subjective naturalness is63%in informal listening test.It shows that the subjective quality of the decoded speech using two-stage spectral shape model is rated to be obviously better than that of the MELP encoder,especially for male's speech.This model not only maintains the accurate characterization of the excitation spec-trum in low-band,but also improves the accuracy in full-band.It can be adopted freely in various kinds of vocoder with better perceptual quality.

关 键 词：声码器 激励参数 两级谱形状模型 模型误差 矢量量化 语音编码 

分 类 号：TN912.32[电子电信—通信与信息系统]