Inside the Chute

I recorded the sound of trash bags falling inside the garbage chute of my apartment.

Here are some questions and objectives I had before starting the project:

  • How can I find a recorder that can be (relatively) safely attached to the walls of the chute and have a storage and battery of its own?
  • Can I get a camera inside the chute to capture the bags falling?
  • What type of microphone should I use to get as close as I can to the experience of being inside the chute?
  • How can I extract individual clips of the bags falling
  • Where do the bags go?

00:00 00:22 17:29 20:44

This final version is made using a small recorder from Amazon, ffmpeg, and Sennheiser’s AMBEO Orbit.

Around 240 clips are extracted by limiting the shorted lengths and db level, so that ffmpeg extracts every clip above -24db and longer than 4seconds. From there I manually sorted through them to find 160 that are actually garbage bags falling. Some clips are loud but not garbage bags falling, such as the rustling of the ziploc bag as I tape the recorder to the chute.

Then the clips are concatenated from lowest db level to highest. I used google text to speech AI plugin with ffmpeg to insert a robot voice saying the numbers before each clip.

AMBEO Orbit is a VST Plugin that can be used inside audition to imitate the effect of a binaural recorder. I manually edited the concatenated clip to achieve a similar effect.

All scripts were written with the help of ChatGPT and YouTube tutorials. 

 

Screenshot

I also tried versions of surround sound, binaural effect using pydub, concat from short to long, and concat without the robot voice counter. You can find them and the python scripts I used here:

(Link to google drive)

What I tried but didn’t work out:

  1. Night Vision camera

The bags falling did not trigger the camera to capture, while the recorder captured sound of bags falling during the same time period.

2. Training a model

tutorial for reference

I followed tutorials on YouTube to train a Tensorflow model with the 160 recordings of bags falling, and around 140 clips of silence and other loud but not bags falling recordings. It is meant to learn from the spectrograms like this:

👆example of my result from the preprocessing function of the training model

I got a model, but it failed to recognize the sound of garbage bags falling in a 2h long recording (which I had left out of the training data).

Due to the time limit, I don’t understand the vocabulary tensor flow uses (like these), which could be the issue:

It may also be the training clips I used that are not well chosen. This will take more time to learn and change.

3. Graphing

I was able to use Matplotlib with ffmpeg to dynamically graph the timestamps of the video at which the garbage bags were falling. This would work better if the learning model could successfully distinguish garbage bags from other sounds.

4. Anaglyph binaural plugin

This is also a VST3 plugin that has a few more features than AMBEO Orbit, but it did not show up in audition or garbage band after installing.

http://anaglyph.dalembert.upmc.fr/index.html