| Speech coding is the fundation of digital transmission and storage of speech. Recently,there is great interests in developing toll-quality speech coders at rates of 4kb/s and below. A high quality 4kb/s speech coding algorithm was proposed in this thesis, which is based on Code Excited Linear Prediction model and Dispersed Pulse CELP algorithm. The non-uniform and part-searching-area algebraic codebook searching algorithm is adopted in the fixed codebook. The non-uniform algebraic codebook is determined by the non-uniform statistical properties of the algebraic codebook, and the part-searching-area is determined by the periodicity of the algebraic codebook excitation vector, which makes up for the insufficient numbers of signed pulses in algebraic codebook at low bit rate. In order to preserve the continuity of pitch, different pitch detection methods are employed for voiced/unvoiced frames when these two techniques are used. The fast-search algorithm based on the weighted Euclidean distance was employed in the quantization of line spectrum frequency parameters and the complexities for calculating the the first four LSF parameters and the last six one are reduced to 23% and 43% respectively. Subjective and objective test results indicate that the reconstructed speech qualities of 4kb/s DP-CELP speech coder are improved based on these techniques, especially for the female speakers. Finally, the improved 4kb/s speech coding algorithm is implemented on the TMS320VC5410 fixed-point DSP board. About 10.25K program spaces and 20.75K data spaces are consumed by this system; about 49MIPS are needed by coder while about 5.2 MIPS by decoder. Objective test indicated there is no perceptual difference for the reconstructed speech quality compared with the original C program result. This speech processing system has a high ratio of performance to price, which is expected to be widely exploited in applications such as digital storage, multimedia communications.
|
PDF Full Text Request