AMR-WB到AMR转码中合成滤波器转换算法

Transformation algorithms of synthesis filter in transcoding from AMR-WB to AMR

提出AMR-WB到AMR转码中的2种合成滤波器转换算法.第1种是基于采样率转换和Prony算法的转换,首先将AMR-WB合成滤波器的单位采样响应进行采样率转换,然后根据最小二乘法,使得新的滤波器的单位采样响应和采样率转换后的响应的误差最小化.第2种是基于自相关值内插的转换算法,首先由AMR-WB语音的LPC参数倒推出自相关,然后采用三次样条内插出AMR语音的自相关,最后利用Levinson-Durbin算法计算LPC参数,即得到解码端的合成滤波器.算法复杂度分析表明,2种算法的计算复杂度都低于Tandem转码.实验结果表明,2种算法都可以得到比较小的谱失真.第2种算法的谱失真在浊音帧比第1种算法略大,在清音帧谱失真有时较大,但是由于清音激励的随机性,对合成清音质量影响不大.

Two translation algorithms of synthesis filter are presented in the transcoding from adaptive multi-rate wideband（AMR-WB） to adaptive multi-rate （AMR）.The first one is based on the conversion of sampling rate and Prony algorithm： the sampling rate of the unit sampling response of AMR-WB synthesis filter is converted,and the error between the unit sampling response of the translated synthesis filter and the synthesis filter whose sampling rate has been converted is minimized according to the least square method.The second one is based on the interpolation of autocorrelations： the autocorrelations are deduced from the LPC parameters of AMR-WB;the autocorrelations of AMR speech are obtained through cubic spline interpolation;finally,the LPC parameters of AMR are computed through Levinson-Durbin algorithm.Complexity analysis indicates that,compared to tandem transcoding,the computational complexity of these two algorithms is lower.Experimental results show that,the spectral distortion（SD） of these two algorithms is small,but for voiced frames,it is a little larger in the second algorithm than that in the first algorithm.For unvoiced frames,the SD in the second algorithm is sometimes high,but it has little effect on the synthesized speech due to the randomicity of excitation in unvoiced speech.