CAJ | 학술논문

针对正弦激励线性预测(SELP)声码器在解码端使用整型基音周期会引入额外舍入误差的问题,提出了一种码本整型化修正算法,该算法通过对码本中的各个码字在线性域进行取整,达到与解码端反量化过程一致的效果.测试结果表明,该算法能够有效地降低基音周期对数域矢量量化的量化误差,并且能够将合成语音的平均意见得分(MOS)提高0.05左右,明显改善合成语音质量.
침대정현격려선성예측(SELP)성마기재해마단사용정형기음주기회인입액외사입오차적문제,제출료일충마본정형화수정산법,해산법통과대마본중적각개마자재선성역진행취정,체도여해마단반양화과정일치적효과.측시결과표명,해산법능구유효지강저기음주기대수역시량양화적양화오차,병차능구장합성어음적평균의견득분(MOS)제고0.05좌우,명현개선합성어음질량.
To resolve the problem of the extra rounding error in sinusoidal excitation linear prediction (SELP) vocoders caused by turning the quantized pitch parameters into integers in the decoding end, this paper proposes an integer codebook modified method. It turns each codeword into integers in the linear-field, making the quantization consist with the decoding process. The simulation results show that the proposed algorithm can reduce the vector quantization error of the pitch parameters in the logarithm field, and increase the objective mean opinion scores (MOS) of the synthesized speech by 0.05, which obviously improves the quality of the speech.