高质量300 bit/s声码器算法

A vocoder to obtain high quality synthetic speech at 300 bit/s is presented based on mixed excitation linear prediction (MELP) which extracts only few parameters each frame. To obtain high quantization efficiency, vector quantization is performed on parameters of the super-frame composed by eight frames. The quantization efficiency of band pass voicing coefficients (BPVC) is improved based on estimation using code-book mapping over mode transition. Codebook sizes of pitch parameter for different unvoiced/voiced model are jointly optimized to improve the quantization efficiency. Meanwhile, multi-stage vector quantization with inter-stage prediction is performed for linear spectral frequency parameters (LSF) to reduce the spectral distortion. Simulation results show that the intelligibility of this 300 bit/s vocoder is quite good and the natural tone is fine. The diagnostic rhyme test (DRT) score is 84. 2%.