engleski

Environmental Sound Recognition Based on SOM Codebook

Sound is a source of information that is used for a scene recognition. One approach is to model dierent scene sounds with help of a codebook. The codebook contains a predened number of codes in it. The record of a scene is divided to a certain number of segments and for each segment the most similar code from a codebook is found. At the end the record is represented as a histogram where each value species the number of occurrences of a speci c sound (code from a codebook). Those histograms are used later within a classier for the classication. In this work a database of various environmental sounds was recorded. Sounds were divided into 9 classes and were used for creating the codebook. The codebook is created with Self Organizing Map. As a classier Support Vector Machine is used. Results showed that with a proper database design a good recognition accuracy can be achieved. With 9 classes accuracy of 64.38% and with 6 classes accuracy of 80.82% is achieved.

Environmental Sound Recognition; SOM Codebook

nije evidentirano