G.729D Speech Codec Implementation On TMS320C64X

Speech signal compression technique is an important technique in multimedia communication. It can reduce speech signal transformation rate and improve the utilizing efficiency of network bandwidth. In 1996, International Telecommunication Union (ITU) brought out G.729 recommendation using conjugate-structure algebraic-coder-excited linear-prediction (CS-ACELP) algorithm. This algorithm has 15ms algorithmic delay and 8kbit/s encoding rate, which lowers the encoding rate while gives high synthetic speech quality and can satisfy the requirement of multimedia communication. Several improved editions were proposed by ITU-T later. In these editions, G.729D edition have a low encoding rate by using CS-ACELP algorithm, which is improved in codebook structure, pitch delay, post process etc. and reduced encoding complexity. It has a wide prospect and has a chance to replace G.723.1 in H.324 recommendation as a new speech encode standard.In recent years, along with the fast development of digital processor, the implementation of high complex speech codec comes true. TMS320C64X DSP produced by TI is provided with VLIW instruction, two level data cache, cross through road, eight level pipe line and has high calculate capability, which can make a efficient development based on suitable development tools.This paper introduces the algorithm structure of G.729, and the improvement part of G.729D, focusing on the LP analysis, transformation of LP to LSP coefficients, pitch detection, searching of adaptive codebooks, etc. As to the complex calculation parts of program, some algorithms has been improved. Besides, the DSP structure characteristic and the development tools are introduced briefly. It is proposed to implement a scheme of G.729D speech codec based on TMS320C64X DSP platform with multiple optimization method. Finally, the performance is tested and evaluated and the optimized programs have met with the expectation in data storage and operational complexity. The distortion of the decoded speech is rather small.