Non-intrusive load monitoring (NILM) provides homeowners with detailed feedback on their electricity usage, but an open area is appliance labeling and generalizable appliance models that can be trained in one home and deployed in another. We therefore propose a semi-supervised learning appliance annotation scheme for home appliance signatures (SARAA). SARAA utilizes time series of appliance turn on and turn off events to tune generic appliance classifiers to appliances in the target home using a mixture of labeled and unlabeled data. Achieving this goal requires the development of a stopping criterion for semi-supervised learning, and we propose and evaluate a stopping heuristic for one-nearest neighbor semi-supervised learning of appliance signature time series. Starting with only a single labeled instance in the target home, SARAA produces classifiers with median F1 scores only 14.8% lower than benchmark classifiers trained on the fully labeled ground truth data in the target home, outperforming classifiers trained only on data from other homes, which have a median F1 score that is 51.23% poorer than the benchmark. The results of this paper will help develop NILM systems, which can automatically learn appliance signatures without user input and facilitate wider adoption of NILM technologies.