摘要
文中提出一种压缩感知框架采样下的语音编码方案。根据压缩感知原理,利用行阶梯矩阵投影产生的观测序列保留了部分语音信息的时域特征,利用正弦字典和匹配追踪算法对观测序列进行建模,对于每帧观测序列的模型参数,根据各自特性采用合适的编码方式进行编码。在解码端对解码后的观测序列利用基追踪算法重构合成语音,并后置低通滤波器提高合成语音的人耳听觉效果。仿真实验表明,提出的编码方案在2.8~5.7 kbps时得到的合成语音平均MOS分为2.81~3.23,在压缩感知框架下取得了较好的语音编码效果。
A novel speech coding method based on compressed sensing is proposed in this paper. Based on compressed sensing theory ,the row echelon matrix retains parts of speech time domain features in the measurements,and utilize a sinusoidal dictionary and matching pur- suit for measurements sequence modeling. The model parameters are encoded by appropriate methods respectively. At the decoder, basis pursuit algorithm employs the decoded measurements for synthesized speech reconstruction. A rear low-pass filter is adopted to improve auditory effects. Simulation results show the average MOS scores of the synthesis speech are between 2.81 - 3.23 in low bit rate （ 2.8 N 5.7 kbps）, which achieves a preferable coding effect in compressed sensing framework.
出处
《计算机技术与发展》
2015年第4期188-192,共5页
Computer Technology and Development
基金
国家自然科学基金资助项目(61271335)
国家重点基础研究发展计划(2011CB302303)
关键词
压缩感知
行阶梯观测矩阵
正弦字典
参数编码
矢量量化
compressed sensing
row echelon matrix
sinusoidal dictionary
parameter coding
vector quantization
