Providing real-time voice support over multihop ad hoc wireless networks (AWNs) is a challenging task. The standard retransmission-based strategies proposed in the literature are poorly matched to voice applications because of timeliness and large overheads involved in transmitting small-sized voice packets. To make a voice application feasible over AWNs, the perceived voice quality must be improved while not significantly increasing the packet overhead. We suggest packet-level media-dependent adaptive forward error correction (FEC) at the application layer in tandem with multipath transport for improving the voice quality. Since adaptive FEC masks packet losses in the network, at the medium access control (MAC) layer, we avoid retransmissions (hence, no acknowledgments) in order to reduce the control overhead and end-to-end delay. Further, we exploit the combined strengths of layered coding and multiple description (MD) coding for supporting error-resilient voice communication in AWNs. We propose an efficient packetlzation scheme in which the important substream of the voice stream is protected adaptively with FEC depending on the loss rate present In the network and is transmitted over two maximally node-disjoint paths. The less Important substream of the voice stream is encoded into two descriptions, which are then transmitted over two maximally node-disjoint paths. The performance of our scheme (packet-level media-dependent adaptive FEC scheme) is evaluated In terms of two parameters: residual packet loss rate (RPLR, packet loss rate after FEC recovery) and average burst length (ABL, average length of consecutive packet losses after FEC recovery) of voice data after FEC recovery. The sets of equations leading to the analytical formulation of both RPLR and ABL are first given for a renewal error process. The values of both these parameters depend on FEC-Offset (r, the distance between original voice frame and piggybacked redundant voice frame) and loss rate present in the network. Then, these parameters are computed for a Gilbert-Elliott (GE) two-state Markov error model and compared with experimental data. Our scheme adaptively selects the FEC-Offset (it chooses 7' that minimizes RPLR and ABL as much as possible) based on the loss rate feedback obtained from the destination. The proposed scheme achieves significant gains in terms of reduced frame loss rate (FLR), reduced control overhead, and minimum end-to-end delay and almost doubles the perceived voice quality compared to the existing approaches. © 2008 IEEE.