A domestic sound dataset designed for sound event detection research while reducing the privacy risk created by intelligible speech in recordings from people’s homes.
The released dataset captures everyday domestic soundscapes using AudioMoth recorders installed in living rooms and kitchens. Recordings were made in Belgium and organised by home, providing long-form acoustic material rather than short isolated sound events.
The release workflow removes speech before publication so that non-speech acoustic information remains available for machine-listening research without distributing intelligible conversations.
The dataset supports work on sound event detection under realistic residential conditions, including daily activities, household devices and other naturally occurring acoustic events. PANNs predictions are supplied alongside the audio as machine-generated annotations.
