About classifying sounds in protected environments

Recently we have proposed a low complexity solution for classifying sounds in wildlife regions. The final goal of this classification was the design of a system that detects intruders in these regions. This paper proposes a different approach, one that uses Mel-frequency cepstral coefficients in a Support Vector Machines framework. The sounds of interest are represented by recordings from humans, cars, birds and animals. The tests are performed on 4 databases of 100 recordings each. Real environments are simulated by considering several types of noises. At the cost of a significantly increased complexity the new approach proves to be more robust. Since low complexity systems are more likely to be feasible for wildlife applications, the complexity issue is discussed and a solution is proposed.