This repository is my playground on creating a sound event detection model for my doorbell.
During summer I reguarly don't hear the doorbell while I am in the garden, because I am too far away from the door or because the door to the garden is closed to keep the hot air outside of our flat. Especially, when expecting a delivery this a quite annoying. Additionally, I am hearing music during work from home and I just realized, that my new headphones do also surpress the sound of the doorbell. So I would like to get a notification to my smartphone.
Existing solutions are quite invasive by modifing the electrics or using a new smart doorbell. Since I don't own the flat I'm living in, I cannot do such modifications easy. So I decided to try to detect the sound of the door bell by microphone.
I got a Rapsberry PI Zero W with a ReSpeaker 2 Mics PI Hat.
I will try to leverage my ML skills to solve this problem and to enhance them along the way. Especially, data management and reproducibility are points I would like to dive deeper.
Everything starts with sufficent data. I already created a simple bash script for recording data using arecord.
This script places a .wav file on a cifs network storage as soon as it finished the recording.
With Termux and Termux:Widget I created a simple way to start short recording from my smartphone, by invoking the recording script on the Raspberry PI Zero W via SSH.
This is handy for starting, but does not scale.
I want to have some (semi-)automatic data gathering.
This is going to be at the first run some rule or DSP based algorithm, which triggers the recording of potential interesting data.
Recorded data is placed on a network drive, which will be fetched from a data labeling system. I use an internal instance of Label Studio, mainly because it was more or less the best tool I found for audio sequence labeling.