一种甚低码率声码器的设计

Design of an ultra-low bit rate vocoder

在混合激励线性预测(mixed excitation linear prediction,MELP)模型的基础上,以超帧为单位,采用多帧联合编码技术,分模式对子帧的语音特征参数进行联合量化,实现了一种码率为600 bit/s的声码器。为了进一步减小量化误差,设计出了一种基于高斯混合模型的预测分类分裂矢量量化器(predictive switched split vector quantization based on Gauss mixture model,GMM-PSSVQ),该量化器对超帧中某些子帧的线谱频率进行量化,并利用帧间预测和线性插值等方法提高编码效率。采用谱失真对设计的矢量量化器进行性能评估,并分别与多级矢量量化和预测分裂矢量量化算法进行性能比较;通过客观感知语音质量评估和主观判断韵字测试对实现的声码器进行性能测试。测试结果表明,设计的矢量量化器平均谱失真最低,实现的声码器合成语音具有较高的清晰度和可懂度。

Based on the mixed excitation linear prediction(MELP)model,this paper designs a vocoder with a bit rate of600bit/s.It adopts a multi-frame joint coding technique with the super frame,and then through the divided model to realize joint quantification for the speech feature parameters of sub frames in the super frame.To deal with the problem that the performance of the existing vector quantization is non-optimal,a predictive switched split vector quantization based on Gauss mixture model(GMM-PSSVQ)is adopted.It quantizes the line spectrum frequency of some sub frames and uses the inter prediction and linear interpolation method to improve the coding efficiency.The performance of the designed vector quantization is evaluated by spectral distortion and it is compared with the multistage vector quantization and predictive splitting vector quantization.The performance of the vocoder is tested by the perceptual evaluation of speech quality and Diagnostic Rhymer Test.Experimental results show that the proposed algorithm has the lowest average spectral distortion,and the speech synthesized by the vocoder proposed in this thesis has high clarity and intelligibility.