Classification of Acoustic Events in a Kitchen Environment using Multiband Spectral Entropy

When the context of a scenario is studied with help of audio, distinct problems appear, which entail a challenge for all the acoustic event recognition systems, such as, noise, mixture of different types of sound sources, among others. The methods used for attending these problems are generally focused on two processes; the feature extraction process and the classification process. In this paper we propose to use Multiband Spectral Entropy Signatures (MSES) for extracting features from acoustic events with a background of mixture of sounds occurring in a kitchen environment. MSES takes into account the randomness of the signal, making it more robust to noise, loudness and spectral flatness. To test our proposal, we created a database of a mix-up of triples from a collection of sixteen real world kitchen sounds using 3dB of signal-to-noise rate. Our benchmark in this work is MFCC feature, since it is often used for this issue. With respect to the classification process, we use two similarity distances, cosine distance and Hamming distance. Experimental results indicate that MSES outper forms the MFCC feature in robustness and effectiveness, improving the performance of the classification process.