This competition is part of the TUT course SGN-41007 Pattern Recognition and Machine Learning.

Acoustic scene classification attempts to categorize audio signals into predetermined classes. Thus the problem setting is similar to speech recognition, except that the target classes are different and a lot more heterogeneous. Yet, there are several possible applications for acoustic event recognition; for example, your mobile phone might automatically detect that you are in a meeting and turn the device to silent mode automatically.

In this competition, the data consists of recordings of 15 different contexts (classes) such as beach, home, restaurant, etc. The recordings have been cut to 10 second long segments, of which the organizers have computed the mel-spectra as the input to your algorithm.

The organizers would like to thank the Audio Research Group of Tampere University of Technology (TUT) for kindly providing access to the competition data. Parts of the data have been used for annual DCASE competitions; see e.g., DCASE2017.

Best result was achieved with Support vector machine -classifier (Sklearn SVC) by using hyperparameters:

{
    "C": 45.0, "decision_function_shape":
    "ovr", "gamma": 1.12,
    "kernel": "rbf"
}

Cross validation score for SVC was 72.48% which had 70.4% private score in competition results