Acoustic recognition of noise-like environmental sounds by using artificial neural network

• Published in: Expert systems with applications 2021-12, Vol.184, p.115484, Article 115484
• Main Authors: Simonović, Miloš, Kovandžić, Marko, Ćirić, Ivan, Nikolić, Vlastimir
• Format: Article
• Language: English
• Subjects: Acoustic noise; Acoustic recognition; Acoustics; Algorithms; Artificial intelligence; Artificial neural network; Artificial neural networks; Background noise; Classification; Expert systems; Frequency spectrum; Heuristic methods; Insects; Internet of Things; Noise; Noise-like sounds; Perception; Preprocessing; Recognition; Sound source recognition; White noise
• ISSN: 0957-4174; 1873-6793
• DOI: 10.1016/j.eswa.2021.115484

•Recognition of noise-like sounds using frequency spectrum as feature vector.•Hybrid procedure for the recognition of environmental sounds based on heuristics.•Optimal preprocessing for improving accuracy in the presence of disturbances. In spite of establishing audio perception techniques and tremendous progress of computer capabilities, artificial perception of sound videlicet natural ability to hear is at the initial stage. The latest phenomena in the evolution of expert systems, like internet of things, put a focus to the mass applications that run on embedded platforms often with pure computational capacity. Modern sensing applications require simplicity, universality, and excellent performance. Insects, obviously, realize limited but satisfactory interaction with the environment using minimum resources. The experiment is motivated by the assumption that the similar principle of control can be applied in artificial control systems. In such a manner, theoretical and practical research was conducted in order of defining optimal procedure for the recognition of short, cognitively unpretentious noise-like sounds, in real conditions, on the basis of previous experience. The result is optimal hybrid procedure for the recognition of noise-like environmental sounds built of heuristic algorithms completely. The experiment reports the success in recognizing unfavourable sounds using frequency spectrum as feature vector. The ability of abstraction was tested by employing samples of different abstraction level and robustness with respect to white noise and confusing sounds. Optimal preprocessing was suggested for the improving accuracy, employing white noise and confusing sounds for estimating preprocessing parameters. Built of ultimate algorithms, the procedure is useful for a broad range of research. This is of exceptional importance because acoustic perception in its full complexity can be approached only if the problem is observed multidisciplinary.