Recent years have seen a rise in the number of labelled human activity datasets to support supervised learning of activity recognition. However, synchronisation and manual annotation of various multi-channel time-series data are cumbersome. Prior research focused on creating datasets that are convenient to annotate, leading to a scarcity of natural activity data that includes various subjects and activities. Therefore, an offline manual annotation tool for efficient labelling activities is desired. This work presents a semi-automatic annotation technique for multi-channel time-series human activity data, utilising a retrieval-based approach to reduce annotation effort. We present an annotation tool that accepts a variety of input data types and supports both manual and semi-automatic annotation. We benchmark the different approaches.