Sound event recognition and acoustic scenes retrieval

The signal of sound events, which defined in environmental sounds, may contain a lot of important information. In the computer systems, audio signals need to perform some processes such as the conversion into the meaningful data, classification and recovery of signals. The necessity of these processes is increasing day by day. It can be used in popular work areas like defense industry, security systems, multimedia search engines and internet of objects. It could be very difficult problem because sound events have no specific characteristic. Moreover in their background, there could be consecutive or overlapping sounds. In this study; we examine and develop performances of sound event identification and acoustic scenes classification. Since it is thought that it can contribute to success of study, various feature extraction methods have been tried and various deep neural network models have been used. To the best of our knowledge, method of the Spatial Pyramid Pooling (SPP), which was successful in imagery identification problems, was first applied to the acoustic scenes classification problem. In our experiments, it has been shown that it contributes to the success on spectrogram features. Moreover, in this study, we added to develop an effective Query-by-Example sound information retrieval system using acoustically and semantically similarities. We investigated; the result of effective acoustic similarity model could be compared against the result of Query-by-Keyword systems.