Pitch discrimination at peaks of frequency modulation is better than at troughs (Demany and McAnally, 1994). A similar asymmetry emerges within a time-domain pitch perception model based on autocorrelation. The model requires the following assumptions: (a) The neural discharge patterns must be sharpened to a single narrow pulse per period (possibly by neural convergence within the cochlear nucleus). (b) Autocorrelation must be implemented as a cross-correlation between the pulse train and a delayed pulse train convolved with a short window function. This function must be asymmetric in time. (c) Pitch discrimination must rely on higher-order modes of the autocorrelation function. Based these assumptions, the asymmetry in pitch discrimination accuracy arises as a result of a convolution of two asymmetric functions: the interpulse interval distribution near the peak or valley of frequency modulation, and the asymmetric window function. The result of this convolution is sharper for peaks than for troughs. The model thus accounts for the pitch discrimination asymmetry observed experimentally. As a by-product it can account for "hyperacute" discrimination observed at peaks of triangular modulation.