DeLTA (Deep Learning Techniques for noise Annoyance detection) Dataset

The Deep Learning Techniques for noise Annoyance detection (DeLTA) dataset comprises 2,980 15-second binaural audio recordings collected in urban public spaces across London, Venice, Granada, and Groningen (sourced from International Soundscape Database). A remote listening experiment was designed and hosted on Gorilla Experiment Builder, a professional online platform used for studying complex behaviours. The survey was then distributed via Prolific to a pool of pre-registered participants (N=1,221), and data collected between July 5th and July 23rd, 2021.

During the listening experiment, participants listened to ten 15-second-long binaural recordings of urban environments and were instructed to select all the sound sources they could identify within the recording and then to provide an annoyance rating (from 1 to 10). For the sound source recognition task, participants were provided with a list of 24 labels they could select from. To collapse these into a single set of sound sources per recording, a “consensus” approach was considered, i.e., if two or more participants identified a source as being present in a recording, this source was considered to be effectively present.  This resulted in a 2890 by 24 data frame (2890 recordings, each with up to 23 possible labels present and an average annoyance rating). On average, each recording has 3.2 identified sound sources present.

Due to the constraints of the online survey software, Mp3 files were used for the listening experiment. Higher quality 24- or 32-bit 48kHz WAV files can be made available from the authors upon request. Each binaural audio recording consists of a 2 channel Mp3 file.