| With the improvement of Information Technology, people's demand for voice communication has increased thus low-rate speech coding plays an important role in the communication system. The most basic low-rate speech coding used is linear predictive coding (LPC). LPC-10 and MELP vocoder are typical successful use of LPC. LPC-10 vocoder was the U.S. third generation of secure telephone voice communication standard since 1986; MELP vocoder replaced LPC-10 vocoder and became the new federal standard in 1997, for its clear, natural synthetic speech quality.To measure the performance of speech coding algorithms, the most important is the synthesized speech quality and the algorithm computational complexity. Therefore, how to get a reasonable compromise between the synthesized speech quality and the computational complexity became key issues. In addition, LPC-10 and MELP speech coding algorithm are not compatible with each other, as a result, communication systems which use LPC-10 and MELP vocoder can not communicate directly. This has made necessary an encoding method to convert these two algorithms to achieve their mutual communication.Based on the theory of speech coding, this paper researched the theoretical basis of low bit rate speech coding. An in-depth study of the realization of MELP speech coding algorithm theory was done with the aid of the PC, using C language in Microsoft Visual Studio 2003 environment to simulate the MELP codec. Assessment of the quality of synthesized speech was made through P.862 (PESQ) which is recommended by ITU. In order to reduce the computational complexity of MELP speech coding, the Federal standard MELP coder is modified by replacing the original LSF quantized with an MTVQ using the proposed fast search procedure instead of multistage vector quantization and at the same time, we divided the frame into three states and to found the exact pitch to improve the synthesized speech quality. The improved MELP codec can get a compromise between the synthesis speech quality and the computational complexity. To match the requirements of practical engineering, this paper used CCS to optimize the improved MELP algorithm and chose TMS320C6418 DSP platform to achieve an improved MELP vocoder and Testing the performance of the vocoder by using informal subjective test and PESQ test, the results show that the improved MELP vocoder can obtain ideal voice quality, good intelligibility and clarity at 2.4k/s. To reduce the spectrum distortion caused by Tandem transcoding, this paper introduces parameters for direct conversion between MELP and LPC-10 vocoder. Detailed analysis of LPC coefficients, pitch period, voicing decision and gain conversion by directly transcoding bit streams without secondary encoding. The results of testing by PESQ to give an objective evaluation shows that, the parameters of directly transcoding compared to Tandem conversion, not only ensure the quality of synthesized speech, but also effectively reduce the computational complexity. |
