Word spotting plays a crucial role in document analysis, particularly for ancient palm leaf manuscripts. Khmer palm leaf manuscripts, which are written on rectangularly cut and dried palm leaf sheets, hold significant cultural value in Cambodia. These manuscripts contain valuable historical, religious, and linguistic information, making their preservation essential. However, extracting information from them is challenging due to their fragility, age, and the complexity of Khmer writing and word formation. This study focuses on word spotting and investigates the construction of a Region Proposal Network (RPN) using the You Only Look Once (YOLO) technique and Convolutional Neural Network (CNN) for the accurate and efficient identification of specific words or phrases within the documents. The proposed method is evaluated using the SleukRith dataset, which consists of 1,971 images of Khmer palm leaf manuscripts. Among these, 1,379 images are allocated to the training set, 395 to the test set, and approximately 197 to the validation set. Parameter tuning is conducted on two variables: the number of layers and the number of filters. The results demonstrate that the optimal model comprises 3 layers and 24 filters, with a threshold of 0.4. The achieved detection performance accuracy is approximately 80.86%, while the classification performance reaches 69.29% for the 33 classes of Khmer characters.