Analytic modeling and simulation of 802.11 and 802.11e have flagged up a number of potential performance problems with voice in an infrastructure mode network. This body of work has also predicted a number of possible solutions, such as the use of TXOP. In this paper we reproduce a number of these problems in a 802.11 testbed, and demonstrate that analytic models usefully bound performance. We study the impact of buffer size, which is difficult to fully capture in analytic models. We observe an asymmetry in buffer requirements between stations and the access point and explore how buffering and the use of TXOP impact throughput and delay of voice.