Spectral and tonality characterization of unwanted environmental sounds for automatic noise assessment

In environmental acoustics, especially during long-term outdoor measurements, unwanted sounds can compromise the reliability of noise assessments. Traditionally, their identification and removal require extensive manual work. Recent advances in machine learning for Sound Event Classification and Detection offer new perspectives for automating such tasks. This study, based on the publicly available DataSEC dataset, presents a spectral and tonal characterization of 40 classes of environmental sounds commonly encountered in urban and rural contexts. Third-octave spectra, spectrograms, and probabilistic tonality maps are analyzed to describe intra- and inter-class variability and to identify recurring spectral structures. The results provide a reference framework for understanding how different sources contribute to tonal and broadband features in environmental recordings. The analysis also discusses how these descriptors can support future automated workflows, including frequency-domain post-processing or spectral smoothing strategies, without claiming experimental validation of such methods. By offering comprehensive, normalized spectral and tonal features, this work provides an open reference for both acoustic professionals and researchers aiming to develop or train machine learning models for environmental sound assessment and event recognition.