Most sub-band coders (SBC) have been realized by quadrature-mirror filter (QMF) with frequency shift and low-pass filter (LPF). However, the QMF method has the following problems: (i) Band separation is limited to the power of 2; (ii) arithmetic errors are accumulated because of the cascade connection of QMF; and (iii) it is necessary to adjust the time delay if the number of QMF connections is different. This paper presents a new architecture for SBC with flexible band assignment, less arithmetic calculations, and small storage memory. In this SBC architecture, a finite impulse response filter (FIRF) is used instead of the QMF. The center frequency of each band is selected according to the equation Fs/Fk = D *n/m, where Fs is the sampling frequency of the input signal, Fk is the center frequency of the k th band, D is the decimation factor, and m, n are integers. According to this relationship, frequency shift and the LPF can be performed at the same time. In this case, data memory for the FIRF is drastically reduced because the input of the FIRF is the same for every band. With the new SBC architecture, the 16- to 12-kbps sub-band coder/decoder was implemented on a single DSP. Band selection and suitable bit allocation are discussed.