基于SELP的150b／s语音压缩编码算法  

150 b /s speech compression coding algorithm based on SELP

作　　者：常亮[1] 徐敬德[1] 崔慧娟[1] 唐昆[1] 

机构地区：[1]清华大学电子工程系清华信息科学与技术国家实验室,北京100084

出　　处：《清华大学学报（自然科学版）》2013年第7期967-971,976,共6页Journal of Tsinghua University(Science and Technology)

基　　金：国家自然科学基金资助项目(60572081)

摘　　要：针对极低速率语音压缩编码中比特资源有限，量化精度严重不足的问题，该文提出了一种新的编码策略——减少量化传输的内容，提高重要内容的量化精度。语音经过低通滤波器将最不重要的3～4kHz频谱滤掉，并相应的将采样率从8kHz降低到6kHz，同时保持每帧样点数不变。这样各个参数的联合帧数就减少为原来的3／4，在比特数不变的情况下，可以有效地提高量化精度。另外，对于线性预测系数（1inearpredictioncoefficient，LPC）而言，由于语音谱从原来的0～4kHz变为现在的0～3kHz，LPC的预测阶数可以从10降低为8，参数维数降低，量化精度可以得到进一步提高。在此框架下，结合子带清浊音（band-passvoi-cing，BPVC）解码端恢复算法，实现了高质量极低速率150b／s语音压缩编码算法。与现有的两种150b／s算法相比，客观平均意见得分（meanopinionscore，MOS）分别提高了0．051和0．067，同时LPC参数的谱失真分别降低了0．09和0．16，改进了合成语音质量，提高了可懂度。A speech coding strategy is presented to improve the low quantization accuracy resulting from limited bit resources in ultra-low bit-rate speech coding algorithms. The algorithm improves the quarttization accuracy by reducing the speech content that needs to be quantized and transmitted. First, the original speech goes through a low pass filter to filter out the least important 3 - 4 kHz speech spectrum and is then down-sampled from 8 kHz to 6 kHz, with the number of samples in each speech frame kept unchanged. The number of speech frames that can be jointly quantized can then be reduced to 3/4 of the original method, which improves the quantization accuracy for the same bit-rate condition. The speech spectral reduction from 0~4 kHz to 0~3 kHz reduces the prediction order of the linear prediction coefficients （LPC） from 10 to 8, so the total LPC parameter dimension is also reduced which further improves the quantization accuracy. Finally a high quality ultra-low hit-rate 150 b/s speech coding algorithm is developed with incorporates a band-pass voicing classification recovery algorithm. The algorithm increases the objective mean opinion score （MOS） by 0. 051 and 0. 067 compared to two 150 b/s speech coding algorithms and decreases the spectral distortion by 0.09 and 0.16, which suggests that both the quality and the intelligibility of the synthesized speech are improved.

关 键 词：极低速率语音压缩编码 低通滤波 降采样 联合编码 线性预测系数 

分 类 号：TN912.32[电子电信—通信与信息系统]