We propose a coder structure, called SB AS ("Subband Analysis-by-Synthesis"), for use in medium- and low-rate coding of speech and audio. We use analysis-by-synthesis to quantize the parameters of a subband signal model. For speech coding applications, long-term prediction can be easily integrated with this structure. The evaluation of SBAS for the coding of telephone-band speech at 8 kb/s gave promising results. A relatively low coding delay (10...50 ms) can be achieved with this type of coder, and its computational complexity is comparable to that of a CELP coder.