Robust automatic train pass-by detection combining deep learning and sound level analysis

The increasing needs for controlling high noise levels motivate development of automatic sound event detection and classification methods. Little work deals with automatic train pass-by detection despite a high degree of annoyance. To this matter, an innovative approach is proposed in this paper. A generic classifier identifies vehicle noise on the raw audio signal. Then, combined short sound level analysis and mel-spectrogram-based classification refine this outcome to discard anything but train pass-bys. On various long-term signals, a 90% temporal overlap with reference demarcation is observed. This high detection rate allows a proper railway noise contribution estimation in different soundscapes.