I’ve got some trouble getting my head round error recovery too, because I don’t think G.711 has any error recovery. I think other schemes have some sort of error recovery, although I think this is limited to replaying previous samples or guessing the next sample rather than actually negotiating the retransmission of a sample.
But, I don’t think G.711 has anything like that. Individual voice samples for G.711 only represent 125 micro seconds, so the loss of several samples will have no effect. But, if you pack 20ms or so of samples into an IP datagram (which is realistic), and then lose one or two datagrams, then the loss would be noticed.