Voids behind tunnel linings are critical factors affecting tunnels’ safety and durability. For automatic, rapid, and accurate detection of void defects behind tunnel linings, this paper proposes an intelligent recognition method of void detection based on deep learning (DL) and percussion method. Extensive indoor percussion experiments were first conducted to obtain a total of 77,925 percussion signals. Afterward, the mel-frequency cepstrum coefficients (MFCCs) are utilized for signal feature extraction, based on which a convolutional neural network (CNN) is developed for automatic void defect diagnosis. The void automated diagnosis tests are subsequently performed, and the impact of three key factors on the recognition results is investigated. The results show that the proposed CNN can accurately identify voids ranging from 0.10 to 0.30 m, with an average accuracy of 94.96% and an F1 score of 72.29%. The exploration of the slab thickness indicates that the proposed method is capable of detecting voids with an average accuracy of 94.37% and an F1 score of 74.55%, with the slab thicknesses ranging from 0.10 to 0.30 m. Furthermore, the boundary effects of concrete slabs are analyzed. Finally, an on-site validation is carried out, and the good agreements between the developed method and ultrasonic detection method indicate that the CNN-aided percussion method is feasible in practical tunnel lining void inspection tasks.