论文部分内容阅读
4kb/s有限状态代数码激励线性预测语音编码算法FSACELP是一种具有延时较短,合成语音质量高、算法复杂度较低的语音编码算法.在线性预测(LP)参数量化上,利用了语音帧内和帧间的相关性,对线谱对(LSP)参数使用预测式分裂式矢量量化,获得很高的量化效率.在自适应码本搜索上,采用了有限状态控制分数延时搜索的算法.在保证合成语音质量的同时,有效地降低了运算量.对于随机码本,采用了具有多模结构的代数码本,提高语音合成质量.对于激励码序列的增益,采用了预测式矢量量化,有效地提高了量化精度.经非正式听音测试,4kb/sFSACELP的合成语音质量超过了北美8kb/sVSELP,接近G7298kb/sCSACELP,MOS分约为39.
4kb / s Finite State Algebraic Code Excited Linear Predictive Speech Coding Algorithm FSACELP is a speech coding algorithm with short delay, high synthetic speech quality and low complexity. In the linear prediction (LP) parameter quantization, intra-frame and inter-frame correlation are utilized, and predictive split vector quantization is used for line spectrum pair (LSP) parameters to obtain high quantization efficiency. In the adaptive codebook search, a finite state control algorithm of fractional delay search is adopted. In ensuring the quality of synthetic speech at the same time, effectively reducing the amount of computation. For random codebook, using algebraic codebook with multi-mode structure, improve the quality of speech synthesis. For the gain of the excitation code sequence, a predictive vector quantization is adopted, which effectively improves the quantization accuracy. After an informal listening test, the synthesized voice quality of 4kb / sFS-ACELP surpassed that of North America at 8kb / sVSELP, which was close to G7298kb / sCS-ACELP. The MOS score was about 39.