The ability of musically trained listeners to detect changes in a learned melody was evaluated in homophonic (H) and polyphonic (P) musical passages. Subjects first learned a 12-note melody in isolation, then listened to three-voice fourth species counterpoints (P context) and three-voice harmonizations (H context) that contained the melody in either the high, middle, or low spectral positions. In a discrimination task, subjects were to respond to changes in the melody only, although changes could also occur in the other two voices. A signal detection analysis found that: (1) performance in the P context was better than the H context for the high and middle melody, but was poorer than the H context for the low melody; and (2) performance for the middle melody was the same as for the high melody location. This second result suggests that the findings of previous work [e.g., Palmer and Holleran, Percept. Psychophys. 56, 301--312 (1994); Acker and Pastore, J. Acoust. Soc. Am. 99, 2482(A) (1996)], where change detection was worst in the middle voice, could be due to this voice never containing the melody. The current results will be discussed in terms of Gestalt principles and Baroque compositional practices. Finally, the differential effects of the chord-related and unrelated changes will be discussed.