Skip to main content

Help fight noise pollution by identifying these sound clips collected from urban sensors

Help fight noise pollution by identifying these sound clips collected from urban sensors

/

Scientists are training machine-learning algorithms to recognize different city sounds

Share this story

Cities are noisy places, and a team of scientists believe that sensors, artificial intelligence, and some generous volunteers can help solve the problem. Sounds of New York City (SONYC) is asking citizen scientists to listen to 10-second sound clips collected by sensors around the city and identify what they hear.

Users are presented with a spectrogram visualization of the audio and a menu of options (“small-sounding engine,” “dog barking,” “ice cream truck”) and have to select all the options that apply. This information will then be fed to an algorithm that will learn to better identify the sources of noise on its own. Hopefully, all this will lead to a better understanding of noise pollution and better tools for fighting it.

SONYC is a five-year collaboration between the city of New York, New York University, and Ohio State University. The Verge spoke to Mark Cartwright, an NYU postdoc who’s working on the project, to learn more about how labeling sound clips could lead to a quieter city. This interview has been lightly edited for clarity.

The ultimate goal of the SONYC project is to combat noise pollution. What are the negative effects of noise pollution, aside from being merely annoying?

It’s one of the most complicated issues in the city. It creates health effects, sleep disruption, hearing loss, learning impairment, and there are economic effects as well. We would like to make New York a quieter place.

The project as a whole tries to monitor, analyze, and mitigate the noise pollution. First, we deployed a sensor network throughout the city and then we trained our models to identify the different sources of noise in the city, along with the loudness that they’re measuring. From this, we can have a better understanding of the noise and then build tools for city agencies to help them understand the noise better and enforce the noise codes.

Volunteers are asked to identify the type of noise. Why does the type matter if you already know how loud the area is?

It makes things more actionable. If we just knew that there’s a loud noise in a particular part of the city and we didn’t know what caused it, it’s hard to mitigate the noise. But if we know that it’s a pile driver that’s causing this large peak in sound, we could potentially measure it if that was exceeding the thresholds in the noise code, for example. We could deploy inspectors to investigate that further.

Maybe some New Yorkers have some subconscious understanding of what these sounds are

We’re working closely with city agencies to help them and build tools for them. That’s more future work, so I can’t say a lot about it now, but we want to help for example, construction workers understand the noise footprint and self-regulate so they don’t get fined and citizens understand noise in the city and make more informed decisions about their life. We also want to make this data available for people to build their own tools on top of, as well.

How does machine learning come into all of this?

In theory, we could have thousands and thousands of people listening all the time to identify these different sounds and notate them, but we would need a ton of people. The idea is to scale that process with machine learning.

Instead of having people continuously listen to sensors, we can have them listen to some smaller number of recordings and label them. And we give that information to machine learning algorithms and train them to do that task, as well, quickly. We’re asking citizen scientists to volunteer their time and listen to recordings and label them, and we hope that since this is a problem a lot of people care about that they will help us work toward a solution.

How many recordings need to be annotated?

We estimate about 50,000. We have a lot of redundancy in place, too. We require that each recording is annotated by three different people and we look at agreement amongst those people.

Do you worry about mistakes and people not being good at identifying the source of noise?

It’s a hard task. We’d love to get volunteers that might have a bit more experience, like potentially some construction workers who understand the difference between a jackhammer and a hoe ram.

One reason why we like to target New Yorkers is not just that it’s relevant because they live here, but maybe some New Yorkers have some subconscious understanding of what these sounds are. You may not realize it, but you possibly can tell the difference in sound between a street cleaner and a saw. If you live in New York and you’ve been walking around and see the street cleaners and associate sound with that, when you’re listening to a recording you might be better at distinguishing those even though they’re both very noisy. And there are weird sounds that happen in New York all the time, like the sound of cars rolling over street plates from road construction, cellar doors, and so on.