Research On Multi-Band And Mixed Excitation Speech Coding At Low Bit Rates

AbstractThe current situation of speech coding is summarized in this paper. The principles of the multi-band excitation(MBE) and mixed excitation linear prediction(MELP) speech coding are analyzed and introduced in detail with emphasis on the implement method of MBE and MELP speech coding at low bit rate.In the study of MBE speech coding, a new efficient practical look-back and look-ahead pitch tracking smoothing algorithm is presented based on the dynamic programming approach. The experimental results show that it can remove the umper?of pitch between the frames, guarantee the exactness of pitch track and improve the accuracy of pitch estimation. The two stage discrete cosine transformation coding scheme is designed. Using this scheme spectral magnitude parameters changing frame-by-frame are quantized adaptively and the coded bit rate of speech is largely reduced. Finally, a 2.4/1.2kb/s MBE speech coder is designed completely and implemented with the digital signal processor TMS32OC3 1. Recently our vocoder has been applied in many communication systems. The application result and DRT test show that the 2.4/1.2kb/s MBE vocoder achieves the practical requirement already.In the research of mixed excitation linear prediction speech coding, a new real-time normalized correlation function pitch detection algorithm is proposed. The integer pitch calculation, the fractional pitch refinement, and the pitch doubling check procedure are applied in the algorithm. The experimental results show that the proposed algorithm can obtain the exact pitch value within the current frame. A new multi-stage vector quantizer(MSVQ) of line spectral frequency parameters is designed. The objective and subjective evaluations show that the proposed scheme offers transparent quantization quality with 25b/frame. Finally, a 2.4kb/s MELP speech coder is designed completely. It contains five additional features. These are: mixed excitation, aperiodic pulses, adaptive spectral enhancement, pulse dispersion, and Fourier magnitude modeling. These features allow MELP speech coder to mimic more of the characteristics of natural human speech. The computer simulation and informal listening test show that the reconstructed speech quality with this coder at 2.4kb/s is closed to CELP algorithm at 4.8kb/s, and superior to 2.4kb/s MBE algorithm proposed in this paper.