The present invention is a technique for provisioning and assessment of voice quality in Internet Protocol (IP) voice applications. A predetermined quality of service (QoS) is received from a user in the form of an acceptable packet loss. The technique measures current packet loss and delay in the network and in the terminal, and determines an optimum play-out delay for use by a play-out adaptation routine. An actual QoS is reported back to the user. To enhance packet delay and loss measurement, a QoS engine component is placed on the terminal side of the packetizer, therefore including packetizer losses in any measurement of the network.