This paper investigates the packet delay variation and packet loss probabilities to be expected in typical applications of voice trunking using ATM adaptation layer type 2 (AAL2) and the consequences these may have, such as the need for sequence numbers in voice packet headers. Simulations have been run for the case of 32-kb/s ADPCM encoded voice across a physical link whose bandwidth is 1/4, 1/2, or 1 times DS1 (1.536 Mb/s). The results show the number of calls that can be admitted while meeting the ninety-ninth percentile queuing delay objective of 20 ms or less. The delay budget that can be allocated to packet queuing delay at access is about 20 ms, given a one-way end-to-end delay budget of 100 to 150 ms for conversational voice. The number of calls is also limited so that the ratio of late or lost packets does not exceed 10/sup -3/. Voice quality is known to be tolerant to packet losses of up to one in 100. When the packet delay variation is significant, as it threatens to be in this situation, sequence numbers can help a receiver to detect and to recover from anomalies-lost, early, or late packets-and make fewer reconstruction errors than if sequence numbers were absent or ignored. Three circumstances are identified where packet slips can occur with ripple effects. An analysis of these cases leads to lower bounds on the sequence number modulus. The appropriate value of the sequence number modulus depends on the desired call load and performance objectives, and can be deduced from the results presented.