This paper proposes a data-driven acoustic method for localizing internal defects in concrete test specimens. MVDR beamforming is applied to array measurements to obtain a time-varying sound energy. Defects are modeled as producing persistent local anomalies within specific time–frequency windows. For each frame, Top-K salient points from the residual energy field are extracted and aggregated across time, followed by spatial clustering to estimate defect regions. Tucker decomposition further suppresses background responses. Experiments on a concrete test specimen with six known voids show that the method reliably reveals defect-related spatial clusters without template correction.