The real time communications are based on UDP protocols. This protocol is connectionless and if a packet is lost it is not send again. In addition the packages loss also takes place by discarding packets that do not arrive on time at the receiver.
Nevertheless the voice is quite predictive and if the packet loss is isolated the voice can be heard in a quite optimal way. The problem is greater when packet loss occurs in burst.
The highest rate of packet loss so the voice can be heard with enough quality must be 1%. But it depends on the codec used. When the codec compression is higher this effect will be more dangerous. A 1% packet loss degrades more the voice if the communication is using G.729 codec instead of G.711 codec.
In order to avoid the packet loss problems the most effective technique is to not send silences (especially in low speed networks or with congestion). Conversations have a lot of silence moments. If we only transmit voice information we can release the bandwidth enough to avoid this problem.
This problem is also related with the jitter and the jitter buffer.