This work presents a semi-automated labeling system for Human Activity Recognition (HAR) datasets. A 1D-CNN trained on scripted activities is used to generate predicted labels for free-form walking segments. Labels below a confidence threshold are marked for review using closed captioned video reducing the time required to label the unscripted sequences. The publicly available TWristAR dataset collected for this work includes Empatica e4 wristband data recorded during six HAR activities and free-form walks. A complete time-stamped video record is provided which we believe makes this dataset unique and useful. Our 1D-CNN model achieves a subject-dependent accuracy of 96.4% and a subject-independent accuracy of 77.4% for six scripted activities. The source code for this work is available at https://github.com/imics-lab/Semi-Supervised-HAR-e4-Wristband.