Nowadays, the “life-log,” recording our daily activities using a camera, microphone or other sensors and retrieve those recorded data, is becoming more and more realistic. One of the applications that utilize the life-log data is the automatic generation of activity summary of the user. The present work focuses on using sound data to make the activity summary. There have been several works that classified the recorded sound based on the user’s activity. The focus of those studies was how to classify the collected data into a pre-defined set of activity classes. However, there have been no considerations what kind of activity classes were appropriate for this purpose. Moreover, we need a basic investigation for optimizing parameters of sound recognition such as window size for feature calculation. Therefore, we first investigated the optimum parameters for feature extraction, and then analyzed the acoustic similarities of sound features observed by various activities. We exploited twenty-two hours of environmental sound in a test subject’s ordinal life as the training and test data. Using the data, we analyzed the acoustic similarities of the activity sound using hierarchical clustering. As a consequence, we observed that target classes could be classified into three groups (“speaking,” “silent” and “noisy”). Misrecognitions between those groups were rare, and we observed a large number of misrecognitions within the “speaking” group.