Internet real time voice communication involves transmitting digitized voice signals which can be lost in the packet-switched network environment of TCP/IP, thereby causing intermittent voice losses and quality degradations. Various methods such as silence substitution, waveform substitution, sample interpolation, Xor mechanism, embedded speech coding, and the combined rate and control mechanism have been proposed to enhance voice delivery and to minimize the quality impact caused by these losses. This paper proposes a quality-based dynamic voice recovery mechanism that combines network transmission control and voice recovery to deliver voice signals with optimal intelligibility and quality. This is accomplished by considering the subjective rating of different codecs that are used in the coding and transmission of digital audio and network packet loss conditions. The dynamic mechanism results in voice delivery that, at minimum, satisfies voice intelligibility while tolerating moderate packet loss caused by network congestion. This mechanism has been successfully incorporated into the Internet Telephone Software System developed at the School of Applied Science, Nanyang Technological University.