|Abstract:||This article shows that the usual speaker-based account of h-aspire in French can explain at most three of the four phonological processes in which it is involved, whereas a listener-oriented account can explain all of them. On a descriptive level, the behaviour of h-aspire is accounted for with a grammar model that involves a 'control loop', whose crucial ingredient is 'listener-oriented faithfulness constraints'. These constraints evaluate phonological recoverability, which is the extent to which the speaker thinks the listener will be able to recover the phonological message. On a more reductionist level, however, the pronunciation of h-aspire and its variation is accounted for with a new, very simple, grammar model that can be called 'phonology and phonetics in parallel'. This model uses a single constraint set for all four processes involved in bidirectional three-level phonology, namely perception, recognition, phonological production, and phonetic implementation. In this model, the phenomenon of phonological recoverability is not built in, as in control-loop grammars, but emerges from the interaction of four equally simple learning algorithms.