论文部分内容阅读
在现有格型矢量量化(LVQ)方法的基础上,通过对线谱对(LSP)参数进行精度预处理,使其数值更接近于格点从而减少量化失真。仿真实现了改进方法,并采用MELP算法中4级MSVQ方法作为对比。得到的结果表明,较之MSVQ,LVQ量化LSP参数的精度提升了约90%的同时,合成语音的谱失真(SD)降低了约2/3,客观评估得分和近似MOS分略有提高。
On the basis of the existing lattice vector quantization (LVQ) method, the preprocessing of line spectrum pair (LSP) parameters is performed to make the values closer to the lattice points to reduce the quantization distortion. The simulation method realizes the improved method, and adopts the 4-step MSVQ method in the MELP algorithm as a comparison. The obtained results show that compared with MSVQ, the accuracy of LVQ quantized LSP parameters is improved by about 90% and the synthetic speech spectral distortion (SD) is reduced by about 2/3. The objective assessment score and the approximate MOS score slightly increase.