Indoor Event Detection with Sound Data

Today, interest in sound data increases rapidly because of easy access and effectiveness of usage. Especially in indoor event recognition, sound data is a great candidate. In this paper, we focus on creating a prediction system that uses sound data for indoor event detection. For this purpose, 8 different indoor sounds are selected from the ESC 50 dataset. These sounds are converted to the spectrograms that bring out different features of the sounds. Image combination(IC) technique combines the different spectrograms of the same sound into one image. On the other hand, multi input technique uses three different spectrograms. By using multi input technique in our experiments, the accuracy rate increased compared to IC technique. We also created a prototype application to determine how well the system works with the real life data. 6 samples per sound type of 8 sound types make a total of 48 sounds recorded in home environment using smartphone microphone. For some sound types we had more than 80% accuracy rates, even 100%.