It is the never ending battle of real time voice and video traffic against the following evil foursome. Here is a quick reference of who/what they are and the metrics required to keep them in check:
According to Cisco and IEEE recommendations the end to end one-way delay should be kept between 150 ms to 200 ms for voice traffic. In reality any greater than 250ms of consistent one-way delay will degrade voice quality and is not acceptable for production deployments.
Symptom: Unnatural delay in conversation
This is the measure of the variation of delay. In other words the inconsistency of voice packet arrival times at the receiving end. As an example, a few voice packets may arrive every 100 ms then the next (few) packets may arrive 120 to 130 ms apart. To smooth out this inconsistency so that the listener receives a consistent flow of voice a jitter buffer is commonly implemented at the receiving end to collect/buffer a large enough number of packets before playing the stream to the listener.
The recommended value to constrain one-way jitter should be less than 30 ms. Jitter buffers are typically designed to handle this amount.
Symptom: A robotic sounding voice from the other end.
A symptom of a number of possible causes which could include faulty hardware, over subscription of WAN links, corrupted packets etc.
Packet loss needs to be kept less than 1% in networks handling voice and video traffic.
Symptom: Clipping of voice and choppiness.
Lack of Bandwidth:
The bandwidth normally requires engineering based on the CODECs being implemented and the calculated expectation of the VoIP and/or video traffic flow.
The above four are the most common and major causes of disruption to the real-time nature of VoIP and Video traffic however there are other issues which can and should be considered when engineering the network. For example, serialization delay should be considered if the WAN link speed is below 768 Kbps (Cisco considers link speeds <= 768 Kbps a ‘slow speed WAN link’).
QoS mechanisms, adequate bandwidth provisioning along with mechanisms such as LFI (fragmentation) for slower WAN links can go a long way in curtailing these villains of real-time traffic.
Overall as long as the voice engineer pays attention to these major areas of contention when pre-assessing the network there is high chance for a successful VoIP deployment.