In many speech coding systems, frames are further divided into subframes. The line spectral frequency (LSF) parameters of each subframe are used to construct the linear predictive coding (LPC) synthesis filter but only the LSF parameters of the last subframe in each frame are quantized and transmitted while those of others are obtained by interpolating in decoder. Conventional coding schemes are designed to try to minimize the spectral distortion of the quantized LSF vectors without considering those of interpolated ones. However, a considerable performance could be gained by exploiting this subframe interpolation operation. This paper presents a novel and efficient LSF coding approach. The proposed subframe-weighted LSF distance measure allows a coder to improve the LSF coding performance over all subframes of all frames. Simulation results show that both average spectral distortion (SD) and outliers of the proposed algorithm are improved. Therefore it is more excellent than conventional ones.