This paper describes a real-time implementation of a variable rate Qualcomm Code Excited Linear Prediction(QCELP) speech coder algorithm standardized as IS-96A. A real-time implementation of the QCELP speech coder is achieved using a customized fixed point DSP chip. Fixed point simulation of high-level language is accomplished to perform trade-off between speech quality and computational complexity in advance. Implemented QCELP speech coder can also support Data Service whose transmission rate is limited to 8.55Kbps using a bypass mode. Furthermore, architectural configuration of the speech coder including input/output operations which are imbedded in the CDMA Mobile System(CMS) is described. The implemented speech coder is evaluated in the customized evaluation module(EVM) as well as in the CMS. interim standard speech coder(IS-96A) for CDMA digital cellular system in North America and Korea. The main advantage of the variable rate QCELP speech coder is that this can reduce the average date rate required to the transmission of speech data by a factor of two or more, while the speech quality of variable rate speech coder maintains the same as that of full rate one. This eventually leads to increase in the capacity of a CDMA system. In this paper, we present the real-time implementation of QCELP using a customized DSP chip. The special efforts are required to reduce the finite word-length effects in implementing QCELP speech coder because DSP chip that we were used is a 16 bit fixed point format chip. The finite-word effects are carefully examined with a fixed format simulation of the high-level language before assembly programming by the DSP assembler. Through the fixed point simulation of QCELP algorithm, the optimum precision in each variable is determined and the effect of finite word-length such as overflow, underflow, and quantization error is studied in advance. The effect of the variable scaling method is investigated in various routines with wide dynamic ranges. The performance of QCELP with fixed point implementation is evaluated in the C fixed point simulation in terms of the Signal to Noise Ratio (SNR). And to meet the continuous demands for the data service, implemented speech coder can also support Data Service which is below 8.55Kbps data transmission rate. This can be feasible with bypassing technique that bypass the packet data to the Selector and PCM I/O, and fundamentally do not degrade the speech quality. Furthermore, the hardware architecture and software structure of QCELP speech coder and implemented H/W platform for evaluation is described.