SONYC Urban Sound Tagging (SONYC-UST)
Multi-class dataset of 2,794 ten-second audio recordings from the SONYC acoustic sensor network in New York, annotated by Zooniverse volunteers for 23 urban noise classes.
- Size
- 1.9 GB
- Format
- CSV, CSV.gz, MD, YAML
- License
- CC-BY-4.0
- Access
- Open Access
- Source
- Zenodo (record 2590742)
Your edits to the notebook only affect your own copy in the Hub. Changed it and want the original back? Restore it here.
Challenge
Which urban-noise classes dominate New York's soundscape, which ones co-occur, and can the citizen-science crowd be trusted against expert ground truth?
- Where to start
- Load annotations.csv reading only the metadata columns (split, sensor_id, audio_filename, annotator_id) and the eight coarse '_presence' columns (1_engine_presence … 8_dog_presence); each row is one annotator's tags for a 10-second clip. Coerce presence to numeric and treat -1 as 0. Keep volunteers (annotator_id>0) and reduce each clip to a majority vote, then count class frequencies and build a Jaccard co-occurrence matrix. On the 'validate' split, compare the volunteer majority vote against the expert ground truth (annotator_id==0).
- What to share
- A per-class prevalence bar chart, a co-occurrence heatmap of the 8 coarse classes, and a crowd-vs-expert comparison (rates plus per-class recall) with a short note on where volunteer tags are reliable and where they under-report.
💬 Discuss this dataset, ask questions and share your results in its discussion thread. Challenges are open-ended — there's no single right answer.