SPECIAL FEATURE Automatic Classification of Impact Sounds with Rejection of Unknown Samples Joaquim Ferreira da Silva 1 • Sofia Cavaco 1 • Gabriel Pereira Lopes 1 Received: 7 April 2017 / Accepted: 24 July 2017 Ó Ohmsha, Ltd. and Springer Japan KK 2017 Abstract The discrimination of very similar sounds is a hard task both for artificial systems and humans. For the former, the main problem lies on finding appropriate features to discriminate each class of sounds, which is an especially hard task when the sounds are very similar, such as impacts on rods of different metals. This paper presents a method to automatically select the features to be used in the classification. Given an initial large set of features, the method measures their discriminative power and builds a reduced set of new features which discriminates the sound classes very accurately. This feature selection method is part of the learning phase of a supervised classification approach also proposed here. In addition, this approach contains a module that rejects unknown sounds also very accurately. This is also an important innovation since most audio classifiers assume all test sounds belong to one of the known classes. Keywords Machine learning  Impact sounds classification  Feature selection & Joaquim Ferreira da Silva jfs@fct.unl.pt Sofia Cavaco scavaco@fct.unl.pt Gabriel Pereira Lopes gpl@fct.unl.pt 1 NOVA Laboratory for Computer Science and Informatics, Faculdade de Cieˆncias e Tecnologia, Universidade Nova de Lisboa, 2829-516 Caparica, Portugal 123 New Gener. Comput. DOI 10.1007/s00354-017-0025-z