Theory suggests that aggressive signals must be costly if they are to be reliable. Recent research in birds has shown, however, that in many species the best predictors of impending attack are low amplitude vocal signals, soft songs or soft calls, that seem cheap to produce and easy to cheat. This observation leads to two related but separate questions: 1) Why use low amplitude signals to communicate aggressiveness? and 2) What maintains the reliability of soft signals of aggression? We review potential answers to both questions and present evidence relevant to each. While some hypotheses are logically sound, others have logical flaws, and most of the hypotheses have yet to be critically tested. One exception is the hypothesis that the reliability of soft signals of aggressiveness is maintained by receiver retaliation, which has been supported by experimental evidence in multiple species. We emphasize the need for further research, particularly to answer the question of why soft song is soft, and outline future research directions.