Repeat recognition for environmental sounds

This work focuses on recognition of repeats in continuous environmental sounds. Environmental sounds, which are very informative in our daily lives, often repeat several times. Repeat recognition for environmental sounds is essential for compact representation of these sounds and for forecasting the future. In our method, input environmental sound signal is partitioned into several units according to the shapes of the power envelopes; the auditory distance between every pair of units is computed. Repeating parts are then detected by using the approximate matching algorithm. Experimental results show 73.3% of recognition rate in counting the repeats of 30 environmental sounds, the length of each of which ranges from 20 seconds to 2 minutes.