Sinusoidal model is efficient for low bit-rate audio coding. A novel sinusoidal parameters extraction algorithm based on narrow band spectrum is described in this paper to improve sinusoidal modelling. Both the frequency and amplitude are resolved with the spectral lines inside a narrow band in signal's DFT spectrum. Frequency is extracted based on signal's autocorrelation function calculated in DFT spectrum, which improves analysis accuracy. Amplitude is obtained by minimizing the mean-square error between original signal and estimated signal in frequency domain. Experiments show that the proposed algorithm outperforms some classic existing algorithms. The analysis accuracy is improved by at least 50%. Evaluating on natural speech demonstrated that it can be applied in speech sinusoidal model.