Subjects' eye movements were measured whilst they read and performed lines of music consisting of rhythmic information only, in conventional musical notation. The relationship between the spatial pattern of the notes displayed and of the fixations made in reading them is stochastic, and similar to that in ordinary reading, but with a tendency to fixate salient details of the notation such as notes and barlines rather than the spaces in between. Shorter notes are less likely to be fixated than longer ones, and this is determined by their performance length rather than their visual appearance. Despite the timing constraints imposed by the music, the time of execution of individual saccades appears to be entirely unrelated to the time of the execution of elements of the performance itself. However, as the tempo of performance of a given piece of music is increased, the average time between saccades decreases but their mean amplitude increases. These observations suggest a new model of the oculomotor and perceptual processes involved, in which an central, iconic representation of the fixated image is internally scanned and interpreted to a given criterion of accuracy, the scan ending when this criterion cannot be reached, and this end-point determining the position of the next fixation. It is proposed that the fullness of the buffer between the perceptual and motor processes determines the strictness of the criterion which is adopted, and hence the amplitude and timing of the eye movements.