In this paper, we investigate the performance of the WiFi-based human presence detection (HPD) method using the two-dimensional convolutional neural network (2DCNN) and long-short term memory (LSTM) under generalized-room model and room-specific model learning environments with optimal 2DCNN and LSTM parameters. The HPD method is designed to classify the WIFi-data segment into a room with person (RWP) or a room without person (RWoP) using magnitude channel state information (CSI). The preprocessing is used to perform filtering and to split the recorded CSI data into blocks, and data preparation. We created the CSI database which consists of 33 hostel room datasets, 3 laboratory room datasets and 3 home room datasets for room specific models with support of 33 volunteers, by using the commercial WiFi device with single antenna. Based on the evaluation results under unseen datasets, cross-validation and room-specific models, we observe that the LSTM based HPD method had an average sensitivity, SE =91.70%, specificity, SP =89.60% and accuracy, ACC= 90.57% that outperforms the 2DCNN based HPD method (SE =89.93%, SP= 88.23% and ACC =88.96%). Further, the LSTM based HPD method had a processing time of 1.8 ms and model size of 934 KiloBytes whereas the 2DCNN based HPD method had a processing time of 2.9 ms and model size of 10470 KiloBytes. © 2021 IEEE.