Abstract: This paper focuses on the design of high quality speech coding with variable rate at 0.75～5.4kb/s. The new algorithm classifies input speech signals as noise, unvoiced speech, transitional speech and voiced speech, and it uses different codebooks as excited impulses according to different types of speech frames. Especially, an embedded splitting vector quantization method based on pitch synchronization to synthesize voiced speech is proposed. The speech can be recovered very well at low bit rate by making use of the itch periodicity for voiced speech frames. This new algorithm can overcome the disadvantage of CELP whose recovered speech quality will degrade quickly when the bit rate is below 4kb/s because the codebook size is too small. the informal listeing test results show that the subject quality of the algorithm exceeds that of QCELP while the average bit rate of the algorithm is only about 2kb/s, which is much lower than that of QCELP whose average bit rate is about 5kb/s. This new speech coding algorithm is therefore very apt to CDMA mobile communication systems.