Optimizing the length of an environmental audio fingerprint for place classification

One of the many possible sources for identifying a place is environmental sound. Ambient sound can be used by itself or in combination with other methods, like GPS, WiFi, etc. A way of identifying a place with sound is using "fingerprinting", which tries to match features of sound in similar places with the one being registered. Nevertheless, one of the many parameters in this process relates to the length of the audio both for the patterns and for the current recording. Several authors use a given time length (e.g. 10, 15, 30 seconds; however, they fail to provide any justification about the time length of the audio fingerprint for creating their classification models. In this paper, we propose to optimize the time length for classifying an environmental audio signal and for increasing the accuracy of the our classification model that uses support vector machine (SVM) as a classifier, and we perform an experimental evaluation. Our results show that the length of environmental audio signal should be between 30 and 40 seconds to get a model with 94.28 % of accuracy.