Design Of300bps Speech Algorithm Based On Melp And Realization On DSP

Low-rate speech coding technology has been a hot research topic.The600bps-2.4Kbps speech coding algorithms have achieved great success and have been widely used, but300bps extreme low-rate speech coding algorithm is still a huge challenge. Existing300bps speech coding algorithms take more attention to improve the quality of synthetic speech and they don’t care the complexity of the algorithm can meet the requirements of DSP platform or not. Hence, studying a300bps speech coding algorithm whose implement is easy and synthetic speech quality is high is very meaningful.The thesis designs a new extreme low-rate speech codec algorithm at300bps on the basis of MELP model. The algorithm composes a super-frame of eight frames. The parameters of band pass voicing coefficients (BPVC), pitch and gain are quantized by8-frame-joint vector quantization. The parameter of linear spectral frequency (LSF) is quantized by4-frame-joint multi-stage vector quantization. The BPVC quantization codebook is obtained by probabilistic statistics. In order to improve quantization performance, the algorithm uses the correlation between voice/unvoice sound information and other parameters to divide codec into different modes. In order to enhance the naturalness of synthetic speech, the algorithm uses the weighted Euclidean distance to measure the distortion of pitch and LSF. The objective MOS point of synthetic speech is2.096after performance evaluation, so the quality of synthetic speech is high.To realize the proposed algorithm on DSP platform, it is ported to low-cost and low-power TMS320C5535DSP and optimized. Performance analysis shows that the real-time implement of the algorithm can be completed. The proposed algorithm is also ported to high-performance TMS320C6416DSP and optimized. The experiment result proves that the algorithm can be implemented together with video encoding or other complex transactions on high-performance DSP.