Fixed-point Optimization Of AMR-WB Algorithm

Voice communication is one of the most basic and important way in human communication field, and speech coding is an important technique in digital voice communication. With the rapid development of Internet, military communications and mobile communications, and the demand of higher voice quality,3GPP proposed the Adaptive Multi-Rate WideBand (ARM-WB) coding standard in March2001.Then AMR-WB was selected as wideband speech coding standard around16kbit/s named G.722.2by ITU-T in2002.The input speech is sampled,and then encoded by AMR-WB algorithm. The input speech is divided into two bands in the encoding:low frequency band50Hz-6400Hz and high frequency band640OHz-7000Hz. Compared with other speech coding algorithms, AMR-WB provides high speech quality, low average bit rates and higher intelligence. So it has broad application prospects in fields of IP telephone, the third generation mobile communication, ISDN broadband phone and video conferencing.Due to the high complexity of AMR-WB algorithm, this paper focuses on the optimization and improvements of the algorithm on the DSP platform. For AMR-WB algorithm DSP implementation, the paper researches on hardware and storage structure of the DSP chip, and DSP software development environment; for the optimization of the algorithm, the paper uses assembly optimization, loop fusion optimization, DSP internal function optimization and so on; Implement the fixed-point algorithm about multi-stage split vector quantization using the codebook classification rearrangement vector quantizer at last. The main work of the paper is as follows:(1)The paper researches on TMS320C55x hardware architecture and software development environment. It focuses on the program and data space of the CPU instruction. In the process of assembly optimization instructions are read from the program space, however parameters are passed between the CPU registers. Write the CMD file including allocation of C5509A of memory and section configuration, such as the.text section, the.data section.(2)The paper mainly researches the optimization of the algorithm. These functions that consume the instruction cycles more and are called frequently in the encoding and decoding process are optimized assembly, which are obtained through the analysis by the Profile tool in CCSv3.3. The C code of AMR-WB algorithm is optimized by C55x compiler internal functions, such as the function__sadd() and the function_min(). Merge multiple cycles that have the same number of executions and don’t affect each other, so it reduces the judgment of the cycle and the numbers of loading data.(3)Fixed-point implementation of the codebook classification rearrangement vector quantization. The input16-dimensional ISF vector is compare to the characteristic values of each sub-codebook in the rearranged codebook, then the searching range is determined. So the improved method can reduce searching times, and it not only significantly reduces the computational complexity of algorithm AMR-WB but also basicly keeps the quality of synthetic speech.(4)Test the performance of the optimized algorithm on the DSPC5509A platform, and objectively evaluate the synthesized voice.