What Sound Sensors Say About New York City’s Noise Pollution
From late 2016 to early 2020, a team of researchers and volunteers at Sounds of New York City (SONYC) used smart sound sensors to collect 30-second clips from across the Big Apple.
The team used these clips to assess noise levels in New York City’s outdoor environments, marking the presence of several major noise sources in each clip, including machinery impact, non-machinery impact, construction equipment, voices, dogs, engines, and car alarms. With most sensors around the NYU area in downtown Manhattan and some in the outer boroughs, the data reflect outdoor sounds New Yorkers are most likely to hear in their daily lives.
Motivation
In 2026 alone, the city has received over 145000 311 complaints about noise on the sidewalk or street. The SONYC project can help understand when certain sounds can be encountered.
By exploring this dataset, one can understand, among many other things,
- whether construction crews obey the legal timeframes in which they are allowed to work and produce noise,
- the amounts of vehicles and people on New York streets at different times of day,
- patterns of construction, driving, and going out across seasons, thanks to data spanning four years
Temporal distribution
The following shart shows distribution of different sounds across a 24 hour cycle. Hover over a segment to see the share of recordings at this hour a particular sound comprised.
Data composition
Researchers and volunteers tagged the presence of 23 sounds, chosen in consultation with the New York City Department of Environmental Protection (DEP). These 23 fine-grained sound categories were then grouped into eight coarse-grained classes.
The chart below shows how often each coarse-grained category of sounds has been recorded (in other words, it counts the “presences” of each sound across the database of all available 30-second clip).
Over the course of four years, the sensors mostly picked up sounds of car engines, alerts, and human voices. Car horns comprised over half of all alert signals, and large engines made half of all engine noises. In addition, non-machinery impacts (loud bursts, kocks or slaps that are hard to place under a single category) were identified in nearly 20% of clips.
Methodology
This explorer uses D3’s native functions, namely d3.rollup() and d3.hierarchy() to create a lightweight data analysis and visualization pipeline. The original SONYC data is loaded as a .csv file from the project folder.
The visualizations are produced by a render layer in the chart components: packChart.js creates the packed category frequency view, and stripChart.js renders the 24-hour distribution.
References
Read more about the motivation and creation of this dataset see the DCASE 2020 Urban Sound Tagging with Spatiotemporal Context Task website.