dill.fish

What I learned recording the audio from my apartment balcony for 60 days

Inspired by Teddy Warner’s Avian Visitors project, I thought it’d be fun to run a similar build on the balcony of my apartment overlooking a loud and busy Brooklyn street. I wasn’t sure if I’d capture any birds at all, so I added more generalized sound identification models (YAMNet and EfficientAT) running alongside BirdNET on my Pi to see if there was anything else of interest I could pick up. It turns out a cheap $10 Amazon mic stuck to the bottom of a patio table is actually (alarmingly) capable of audio surveillance! This is a summary of what it recorded.

First, some baseline numbers: the noise level fluctuated daily between an average floor of ~62 decibels, about the noise level of a normal conversation or office environment, and an average peak of ~67 decibels, about as loud as a vacuum. The quietest hour was typically around 4 AM, and the loudest typically between 10 AM and 1 PM, driven mostly by a month of nearby construction work. On an average day, there were 175 car horns (measured as “episodes,” not individual horn blasts), reversing beeps from 120 delivery trucks backing up, and 30 emergency vehicle sirens.

average loudness by hour · aug 17 to oct 19

The daily rhythm

The averages smooth over individual spikes, which were often much louder. Sirens were the most common source of decibel peaks, consistently measuring 100–105 decibels even from 10 floors up, about as loud as a chainsaw held at arm’s length. Other top-ranking peaks included a motorcycle revving at 4 AM, car horns, construction, and music played from car speakers at nearly the same level as a typical concert.

five of the loudest recordings · clicktap to listen

    Holidays and weather drove visible deviations from trends. During heavy rain, occurrences of music fell by 86%, bike bells fell by 76%, and car horns rose by 19%. The wettest day was also the day with the most recorded honks. On Labor Day, bike bells rose 2.8x, motorcycles 2.6x, and music 2.4x compared to a typical Monday.

    When the weather turned

    Shaded: days with 10 mm or more of rain. Dashed: Labor Day.

    Music was one of the most interesting categories. When detected by the model, music was passed to the Shazam API and identified about 25% of the time, with 9 songs logged each day on average. Side note: I learned through this project that Shazam basically only uses visual spectrographic analysis to match songs, and real-world street audio is very messy with lots of overlapping spectral artifacts. Songs which were often easily recognizable to my ears in a recording could not be identified by Shazam or other models. Seems like there’s a low-hanging opportunity here to supplement with speech detection + lyric matching and other AI-augmented methods.

    The most-identified artist was Bad Bunny. The full playlist is here, but I’ve selected a few samples in the player below to give a sense of the neighborhood vibes. The ice cream truck was identified by its song on 20 days, about one day in three, and surprisingly stayed steady through October 19th (the last day I recorded), showing no seasonal slowdown (though the sample is small). The mic also picked up a specific car that drove by on 5 separate occasions blasting the same two opera songs, always between 4 and 5:30 PM.

      What about the original intent of the project, bird sounds? I couldn’t recall ever hearing a bird call from my apartment, so when I originally set up the mic, I wasn’t expecting much. I was surprised to find there’s actually a diverse orchestra of bird songs playing outside my window every day, and I am in fact either too distracted or cold-hearted to hear and appreciate them. The birds were most active in the early morning (around 7 AM) and around 4–6 PM, and 32 species were reliably identified. Now I recognize the distinctive sounds of the Common Starling and Northern Cardinal every time I open the window.

        Even more common than bird sounds, however, were the chirps of electronic moped locks. Though I recognized the sound before starting this project, I was pretty surprised to see how frequent they actually are in the data – more common than any bird species, and it’s all the exact same double chirp pattern. After some street sleuthing, I was able to confirm that the chirps were emanating from a single brand of moped: Fly E-Bike, currently entrenched in a number of lawsuits and recalls due to fires and other safety failures. The mopeds are sold and supported by a large network of dealerships around the city which also service and (in the case of electric bikes) charge the vehicles that keep fleets of gig drivers moving. So if you live anywhere near a restaurant or cafe, you’ll probably recognize this sound too – it’s the sound of a DoorDash driver hopping off a moped, locking their bike as they head inside to pick up an order, then unlocking it and hopping back on to whisk it away to the customer. Possibly some untapped alpha here for DASH earnings or consumer discretionary spending. Once you listen to the chirp in the clips below, you’ll start to hear it absolutely everywhere.

          The project was pretty easy to set up, thanks to Teddy’s project write-up. The generalized sound models added a slight challenge, as I was surprised to find that most general sound identification models didn’t work all that well out of the box. Common urban sounds like the clinking of dishware, jangling of a chain, squeaking brakes, etc. were often comically misidentified – the models consistently recognized the venting air of a kneeling bus as a person snoring. Music with lyrics was often identified as speech, which was then filtered out for privacy. Even sounds like the moped chirps, which were quite distinctive to my ear, weren’t identified by the generalized models at all. All of these categories required fairly significant manual training on custom classifiers, e.g. by listening to clips and properly tagging their sounds. Overall, a fun project that taught me something about an under-appreciated aspect of my environment and the character of my neighborhood.