A speech coding apparatus which selects an optimum code from a code book (21), the optimum code giving the minimum magnitude of error signal between the input signal and the reproduced signal obtained by a filter calculation using a linear prediction parameter from a linear predictive analysis unit (10) with respect to the codes of the code data, wherein use is made, as the codes, of a code formed by thinning to 1/M (M being an integer of two or more) the plurality of sampling values constituting the codes. To compensate for the deterioration of the quality of the reproduced signal caused by thinning the sampling values in this way, an additional linear predictive analysis unit (20) is further introduced and use made of an amended linear prediction parameter instead of the linear prediction parameter.