Reliable broadcast is a fundamental primitive, widely used as a building block for data replication in distributed systems. Informally, it ensures that system members deliver the same values, even in the presence of equivocating Byzantine participants. Classical broadcast protocols are based on centralized (globally known) trust assumptions defined via sets of participants (quorums) that are likely not to fail in system executions. In this paper, we consider the reliable broadcast abstraction in decentralized trust settings, where every system participant chooses its quorums locally. We introduce a class of relaxed reliable broadcast abstractions that perfectly match these settings. We then describe a broadcast protocol that achieves optimal consistency, measured as the maximal number of different values from the same source that the system members may deliver. In particular, we establish how this optimal consistency is related to parameters of a graph representation of decentralized trust assumptions.