Uses DL techniques to detect whether a baby is crying (to notify young parents)
$ pip install -r requirements.txt
$ python server.py run
- URL: https://baby-cry-detector.now.sh/
- branch: release_docker (only has files needed for Inference)
- Data Collection
- We used https://freesound.org/ to collect audio files, augmented with https://urbansounddataset.weebly.com/urbansound8k.html
- Idea above was used to generate a dataset of different sounds which can occur around the place where a infant/baby is sleeping
- We tried to make a robust dataset, like
- having different animal cries which could be similar to a human baby cry
- having silence as some of the training samples
- No effort was put into trying to differentiate between a baby laugh v/s a baby cry – as the need was to detect whether a sleeping infant is up (babies normally cry once up), and even a laughing (or any sound) baby should trigger the alarm for parents (if needed, this could be attempted, and would be the real test for the model)
- Data Preprocessing
- The audio files are split into smaller chunks of audio to capture different patterns that can occur
- The audio is then converted into spectrograms - on which we then tried different Image Classification models
- Results
- CNN with Transfer Learning from a pre-trained resnet34 seems to be performing extremely good (with a CV split of 0.2)
- Add continuous audio support - where the user presses the START button and leaves the app (on a device) - and the app can detect whenever the baby starts crying (we can just run the current process every
'x'seconds on repeat until stopped by the user) - Integrate with Twilio API for an option to send a message to the user when a baby cry is detected
- Similar/Alternate to above, have a simple mobile app which could send out a notification when a baby cry is detected
- Improve the dataset with more different kinds of audio sounds that can confuse the model