Estimation of wave velocity (or slowness) from array waveform data is a basic and very important process in acoustic logging and seismic processing. A predictive method is developed to process array waveform data containing multiple wave modes. These wave modes may overlap in both time and frequency and are inseparable using conventional techniques. In this new technique, the waveform at a receiver is modeled by a combination of wave data at other receivers using a time-domain prediction theory. It is assumed that the array data contain a number of propagating modes. A minimization procedure is formulated to optimize the match between the predicted and measured waveforms, yielding slowness estimates of the wave modes across the array. Most important, the optimization is performed directly in the time domain using the entire array wave data set, including all possible data combinations. This strategy effectively reduces the noise effects and enhances the robustness of the estimation. Furthermore, the estimated slowness values can be used in formulating a procedure to split the array data into individual wave modes, allowing their behavior to be analyzed. Examples are shown to demonstrate the ability of the technique to extract wave slowness from multiple wavemode data.