Simulated smart phone recordings for audio identification

Abstract

This paper studies the use of simulated recordings to perform audio identification experiments. In contrast to use actual recordings in the experiments, we use the measured room impulse response to generate simulated recordings. Doing so greatly reduces the burden of manually recording audio items for experiments. By comparing the correlations between actual and simulated recordings, we conclude that this approach is highly possible. The audio identification experiments are conducted based on the moving picture expert group audio signature descriptors to represent the simulated recordings. We also add environmental noises, provided by European Telecommunications Standards Institute, to the simulated recordings to study the performance degradation. Finally, we study if performing filtering in the descriptor domain can improve the accuracy. The experimental results show that filtering in the frequency direction yields higher accuracy for signal to noise ratio of 10 dB items.