Adhered cladding is widely used in building façade decoration; however, interfacial voids severely compromise structural safety during service. The study focuses on adhered ceramic tiles, conducting single-point impact and vibration acquisition tests on specimens with pre-defined void defects of varying sizes and shapes. Test results indicate that while Frequency Response Function (FRF) analysis reveals correlations between void size, resonance frequency, and peak amplitude, it fails to reliably distinguish small voids or specific boundary conditions due to spectral overlap with intact regions. To address these limitations, a residual neural network (ResNet)-based void classification model is developed by fusing two-dimensional impact force images and logarithmic Mel-spectrogram features of acoustic signals. Results demonstrate that the proposed fusion strategy significantly enhances detection accuracy and noise robustness compared to single-modality approaches. The model achieved classification accuracies exceeding 99.6% for multi-class void type identification and 99.9% for binary damage detection under clean conditions. Notably, even at a low signal-to-noise ratio (SNR) of −5 dB, the fusion models maintained accuracies above 94%. These findings validate the efficacy of force-sound fusion features for reliable, automated void detection in building exteriors under complex environmental conditions.