Abstract Listeners show strong agreement on which tone should come next in an unfamiliar melody. Cognitive models of expectancy often involve listener responses to instrumental renditions of vocal melodies, and model expectancy for the next tone on its pitch height relative to previous tones in working memory. But if sung melodies encourage an expectation of singability, then listeners should expect the next tone to fall within the singer’s modal register regardless of preceding context. Nonmusicians heard either sung or instrumental renditions of vocal melody fragments and made retrospective goodness-of-continuation ratings on the final (probe) tone. Half of the probe tones were within the modal register; half were above. Ratings were modelled on a cognitive model based on pitch relations and a singability model based on absolute pitch height alone. The cognitive model best fit ratings on instrumental melodies but, as expected, the singability model best fit ratings on sung melodies. Ratings were higher for sung melodies than instrumental in the modal register, but lower for sung than instrumental in the upper register. An alternative statistical learning account based on probe tones’ novelty rather than singability was not supported. Vocal constraints offer a sensorimotor framework for understanding melodic expectancy in naturalistic song.