4 kb/s improved CELP coder with efficient vector quantization

A 4 kb/s LCELP (learned code excited linear prediction) speech coding method is presented. In order to improve conventional CELP speech quality at around 4 kb/s, in LCELP, a learned codebook designed by the LBG method is used to efficiently represent an excitation signal. An effective orthogonalization method is developed. Delayed-decision search is adopted to search for a globally optimal code-vector combination over a frame. In order to allocate more bits to the excitation codebook, filter LSP parameters are efficiently quantized by using a multistage vector quantizer with multiple candidates. Experimental results show that 4 kb/s LCELP achieves an average of 10.8 dB segmental signal-to-noise ratio, which is 0.6 dB higher than that for 4.8 kb/s conventional CELP