• Delay : Two problems that result from high end-to-end delay in a voice network are echo and talker overlap. Echo becomes a problem when the round-trip delay is more than 50 milliseconds. Since echo is perceived as a significant quality problem, VoIP systems must address the need for echo control and implement some means of echo cancellation. Talker overlap (the problem of one caller stepping on the other talker's speech) becomes significant if the one-way delay becomes greater than 250 milliseconds. The end-to-end delay budget is therefore the major constraint and driving requirement for reducing delay through a packet network.
• Jitter (Delay Variability): Jitter is the variation in inter-packet arrival time as introduced by the variable transmission delay over the network. Removing jitter requires collecting packets and holding them long enough to allow the slowest packets to arrive in time to be played in the correct sequence, which causes additional delay. The jitter buffers add delay, which is used to remove the packet delay variation that each packet is subjected to as it transits the packet network.
• Packet Loss : IP networks cannot provide a guarantee that packets will be delivered at all, much less in order. Packets will be dropped under peak loads and during periods of congestion (caused, for example, by link failures or inadequate capacity). Due to the time sensitivity of voice transmissions, however, the normal TCP-based re-transmission schemes are not suitable. Approaches used to compensate for packet loss include interpolation of speech by re-playing the last packet, and sending of redundant information. Packet losses greater than 10% are generally not tolerable.

Maintenance of acceptable voice quality levels despite inevitable variations in network performance (such as congestion or link failures) is achieved using such techniques as compression, silence suppression, and QoS-enabled transport networks. Several developments in the 1990s, most notably advances in digital signal processor technology, high-powered network switches, and QoS-based protocols, have combined to enable and encourage the implementation of voice over data networks. Low-cost, high-performance DSPs can process the compression and echo cancellation algorithms efficiently.

Software pre-processing of voice conversations can also be used to further optimize voice quality. One technique, called silence suppression, detects whenever there is a gap in the speech and suppresses the transfer of things like pauses, breaths, and other periods of silence. This can amount to 50-60% of the time of a call, resulting in considerable bandwidth conservation. Since the lack of packets is interpreted as complete silence at the output, another function is needed at the receiving end to add "comfort noise" to the output.

Another software function that improves speech quality is echo cancellation. As was noted earlier, echo becomes a problem whenever the end-to-end delay for a call is greater than 50 milliseconds. Sources of delay in a packet voice call include the collection of voice samples (called accumulation delay), encoding/decoding and packetizing time, jitter buffer delays, and network transit delay. The ITU recommendation G.168 defines the performance requirements that are currently required for echo cancellers.

Engineering a VoIP network (and the equipment used to build it) involves trade-offs among the quality of the delivered speech, the reliability of the system, and the delays inherent in the system. Minimizing the end-to-end delay budget is one of the key challenges in VoIP systems. Ensuring reliability in a "best effort" environment is another. Equipment producers that offer the flexibility to configure their systems to fit the environment and thereby optimize the quality of the voice produced will have a competitive advantage.