摘要
利用语音短时谱变化相对较慢、其邻近LSP(LineSpectrumPair)失量存在充分相关这一特性,提出了一种新的谱编码方法,即LSP编码的一步插值预测矢量量化。本文设计了一个18bit/frame分裂矢量量化方案用于量化预测残差,当帧变化周期为30ms时,平均谱失真仅为1.178dB。
Taking advantage of the sufficident correlation of successive LSP(Line Spectrum Pair)vectors because of relatively slow variation of the short term spectrum of the speech waveform,a new spectral coding approach,one-step interpolation predictive vector quantization of LSP codingis pro-posed.18 bit/frame split vector quantization scheme is designed to quan-tize the prediction residual.The average spectral distortion of 1.178 dB can be achieved when frame period is 30 ms.
出处
《长春邮电学院学报》
1995年第4期1-7,共7页
Journal of Changchun Post and Telecommunication Institute
基金
邮电部中青年科研基金
关键词
线谱对
矢量量化
语音合成
插值
编码
Speech processing
Linear preestimating coding
Line spec-trum pair
Vector quantization