A real-time subband speech codec is implemented by the author on TMS320C25 digital signal processor based hardware for speech compression down to 0.25 bit per sample. The Binomial Quadrature Mirror Filters (B-QMFs) are employed as the basic block for the hierarchical, dynamic tree, analysis/synthesis filter bank. After the analysis, available bits are optimally allocated among subbands. Band statistics are updated to overcome the nonstationarities of speech signal. The energy compaction performance of this codec is compared with different subband and DCT in simulations for speech signals and AR(1) source models. It is shown that the speech codec implemented provides an intelligible speech quality even at 0.25 bits per sample.
|