As an adaptive algorithm of signal decomposition, matching pursuits provides a new framework for sinusoidal modeling of speech and audio signal. In this paper, the procedure of sinusoidal modeling using matching pursuits is analyzed as well as the sinusoidal modeling algorithm using perceptually weighted matching pursuits. And a method of sinusoidal modeling with perceptual gradient is proposed. The proposed method, which adopts the adaptive feature of matching pursuits, computes dynamically a masking threshold from the currently synthesized signal using the psychoacoustic model. With the threshold, it extracts the most perceptually significant component from the residual signal. Therefore, the perceptual information contained in the synthesized signal increases as quickly as possible. The quality of the synthesized speech by this approach is rather high even if the model precision is low. Experiments prove that the method in this paper uses the features of hearing system in a better way, and the modeling is reasonable and efficient. Both the objective compare of SNR and the subjective listening test show the rationality and superiority of the new method.