A codebook shaping method for perceptual quality improvement of CELP coders

At low bit-rates, the error weighting filter based noise shaping technique of the conventional CELP coders often yields audible noise in spectral valley regions due to incomplete noise shaping. In this paper, we propose a simple but effective method to solve the problem. Firstly, we conducted psychoacoustic experiments to show that deterministic noise is more tolerable to the human ear than random noise in spectral valley regions. To exploit this useful auditory property, we apply a spectral shaping filter to the codevectors before the fixed codebook search so as the vector quantizer to be overloaded in spectral valley regions, which makes the quantizer emphasize formant regions while reducing the noise in the spectral valleys. Subjective quality tests show that the proposed method applied to the 5.3 kbps ITU-T G.723.1 CELP coder significantly improves the perceptual quality of the coded speech with little increase of computational cost