Prestressed concrete (PSC) girder bridges are widely used owing to their economic efficiency, durability, and effective maintenance. However, since voids in ducts may cause sudden structural collapse, it is very important to detect them early. To solve this problem, voids are detected by analyzing the impact-echo (IE) signal measured by IE equipment containing a sensor, but it is difficult to accurately detect voids in a short time even by experts. In this study, we aim to detect voids in ducts on the basis of various types of neural networks and IE signals. For more effective learning, the raw IE signal is filtered and then used in its specific range, and it is also converted into a frequency spectrum by the Fourier transform. The filtered IE signal is trained with long short-term memory (LSTM) to reflect the characteristics of its time series. The frequency spectrum is trained with a feed-forward neural network because it is not a time series. After that, a multiplication operation is performed on the outputs of each network, and a model capable of detecting the internal voids of ducts is created by training these integrated features. In the experimental results, our proposed model showed an accuracy of 97.474%.