The AT&T LD-CELP speech coding algorithm was ratified in 1992 as CCITT standard G728. The LD-CELP algorithm meets low-delay requirements and offers very good speech quality and error robustness. Floating-point implementations of the algorithm are known on the AT&T DSP32C [3] and the Texas Instruments TMS320C30, but the algorithm has a reputation to be difficult to implement in fixed-point. We implemented the algorithm on the SGS-Thomson ST18933 general-purpose fixed-point processor. This paper describes our implementation, specifying data format at different stages of the algorithm and other algorithmic details. Several problems were encountered during this implementation and the solutions we propose here can be used in subsequent implementations and/or in a fixed-point specification of the algorithm.