This study conducted an experiment focusing on the evaluation of generalization performance of the water leakage detection model proposed in the previous studies. The target model was a convolutional neural network-based model with the image data obtained through recurrence plot as the input information. This experiment used the actual leak sounds obtained at ten different observation points. Multiple combinations of training data were created for the leak detection models. The models were then evaluated by assessing their generalization performance against the unlearned actual water leakage sound. As a result of the experiment, the "two-point model", which used the observation data of two points for learning, achieved high accuracy of about 90%; however, some cases with accuracy less than 50% were also confirmed. On the other hand, it was found that the "8-point models" had high generalization performance and were able to obtain an accuracy of at least 80% for 7 of the 10 observation points.