Analysis and recognition of similar environmental sounds

Dissertação apresentada na Faculdade de Ciências e Tecnologia da Universidade Nova de Lisboa para obtenção do grau de Mestre em Engenharia Informática

Humans have the ability to identify sound sources just by hearing a sound. Adapting the same problem to computers is called (automatic) sound recognition. Several sound recognizers have been developed throughout the years. The accuracy provided by these recognizers is influenced by the features they use and the classification method implemented. While there are many approaches in sound feature extraction and in sound classification, most have been used to classify sounds with very different characteristics. Here, we implemented a similar sound recognizer. This recognizer uses sounds with very similar properties making the recognition process harder. Therefore, we will use both temporal and spectral properties of the sound. These properties will be extracted using the Intrinsic Structures Analysis (ISA) method, which uses Independent Component Analysis and Principal Component Analysis. We will implement the classification method based on k-Nearest Neighbor algorithm. Here we prove that the features extracted in this way are powerful in sound recognition. We tested our recognizer with several sets of features the ISA method retrieves, and achieved great results. We, finally, did a user study to compare human performance distinguishing similar sounds against our recognizer. The study allowed us to conclude the sounds are in fact really similar and difficult to distinguish and that our recognizer has much more ability than humans to identify them.