In rendering a virtual sound, an adaptive inverse filtering is needed for crosstalk cancellation. Although various adaptive filtering algorithms have been proposed for crosstalk cancellation, few have been effective. The least-mean square (LMS) algorithm is known to be accurate, but it has a slow convergence rate, especially for colored inputs. In the past, perceptual characteristics of human auditory system have not been considered. In this paper, we present a novel crosstalk cancellation for fixed speaker-based spatial audio rendering using psychoacoustic model. The performance of crosstalk cancellations is evaluated according to parts of output. The performance of crosstalk cancellation of one part is evaluated by the mean-squared error between filtered output and desired output. The performance of crosstalk cancellation of the other part is evaluated by comparing filtered output to masking threshold of other part signal. Experimental results show that the proposed crosstalk canceller converges faster than previous method and the proposed crosstalk canceller outperforms previous method in perceptual.