Abstract: Vector sum excited linear predictive speech coding was studied and enhanced.Presented were the algorithm design and implementation of the VSELP coding at 8kbit/s,and those enhanced at 6.4kbit/s,4.8kbit/s and 2.4kbit/s via vector quantization of LPC parameter,reduction of excitation codebooks and sub block splitting etc.The informal listening showed,that the reconstructed quality of speech (QOS) at 4.8kbit/s and above were not obviously worse than that at 8kbit/s,and of good naturalness intelligibility,but the QOS at 2.4kbit/s was perceivably worse than that at 8kbit/s.