Guidelines for Speech Quality

To achieve good speech quality on an IP network certain rules have to be followed. The main rules are that:

• The total end-to-end one-way delay should be kept below 150 ms. One-way delay = Packetization delay + Propagation delay + Transport delay + De-Jitter Buffer delay.
  • Packetization delay is the fixed time needed by the transmitter to gather audio samples, put them through a codec and placing the right amount of codec frames into an IP packet.
  • Propagation delay is the time it takes to travel the physical distance from one point to another. When traffic has to cover long distances, be sure that the network path is optimized and as short as possible.
  • Transport delay is the total time the packet has to spend in the different devices making up the network, like switches, routers, traffic shapers, gateways, bridges, and so on. Some devices add more latency than others. Look for the number of hops needed for a packet to go from point A to point B. Try to focus on the worst offenders and try to avoid or replace them.
  • De-jitter Buffer delay is used to take care of the variations of packet arrival caused by the network. The delay in the buffer can be long, if jitter is high and network delay is low, but this is the point where trade-off between delay and packet loss has to be considered.
  • The transmitters typically have a delay of 20 - 30 ms, jitter buffers typically delays by 40 - 60 ms and PSTN has a typical propagation delay ranging from 10 - 30 (maximum 50 - 70 ms). This gives some 30 - 80 ms delay left to the IP network, when the total delay budget should be lower than 150 ms. The recommendation is to design the IP network with less than 50 ms of delay.

• Packet loss should be kept below 2% as this has been shown to be a level after which users start to complain of poor voice quality. For Fax-over-IP 2% is not acceptable. Also try to avoid bursts of consecutive packet loss. Increasing bandwidth and good tuning usually solves the problem. It is recommended to keep the target to less than 1% loss when designing the network.
• The bandwidth required for the designed number of voice calls should be allocated so that it is available at least 90% of the time. (This means that all other traffic should be limited by traffic policies.)
  • Avoid slow speed links, extensively. Plan to upgrade bandwidth on paths where different media types should co-exist.
  • Use data packet fragmentation for low speed links.
  • Use RTP header compression for low speed links.

• When Traffic Shaping is applied, apply at least 10% over-provisioning to voice calculation. Some situations though might require higher over-provisioning ratio to work properly. It is recommended that Low Latency Queuing (LLQ) or Weighted Fair Queuing (WFQ) is the preferred choice of queuing used for the voice channels. Other methods like Random Early Discarding (RED) will further help up to avoid continuous packet loss. Be aware of that the typical behavior of shapers is that they usually throw away UDP packets first, as they are not resent. Will the traffic shaper affect RTP packets?

For more information see the ITU-T Standards G.107, G.113 and H.460.9, the description for QUALITY OF SERVICE and (if H.323 end-points are used) also the description for CALL INFORMATION LOGGING, QUALITY OF SERVICE LOGGING.