Abstract: An underwater digital voice communication system aiming to be used in shallow-water acoustic channels was studied and implemented by combination of coherent underwater acoustic communication technique and MELP (Mixed Excitation Linear Prediction) digital voice encoding technique. According to different importance of voice data in recovering speech, this system used hierarchical channel encoding scheme to protect voice data in order to avoid introducing too many redundancies, trying to improve transferring efficiency of the whole system. To assure the time continuity of speech, which is of importance to the performance of whole system, a buffer-based mechanism and a robust voice data frame-synchronization scheme were used. A channel estimation based Decision Feedback Equalizer with fractionally-spaced feedforward filter (FS-DFE) and an embedded two-order Digital Phase-Locked Loops (DPLL) was used to account for multi-path effects of acoustic channels. A sea experiment was conducted and promising results were obtained: good quality (can identify different speakers) real-time underwater voice communications were successfully realized over 6km and 11 km. This result showed that this system can be applied to half-duplex real-time underwater digital voice communications.