48 | CryBaby - Infant Interpreter

Our tech stack.
User Interface.
A visual representation of a baby cry - Mel Spectrogram

Inspiration

Many adults, especially first-time parents, aren’t properly trained to take care of babies needs. For example: 85% of accidental baby deaths are caused by suffocation and strangulation. Babies use a variety of cries to communicate their needs, such as a different cry for hunger versus pain.

What it does

CryBaby is an app designed to help parents identify the reason why their baby is crying. By analyzing the sound of the baby's cries, the app uses AI to identify potential causes such as hunger, tiredness, diaper change, or pain. It then provides advice on how to soothe the baby, such as feeding or rocking techniques. The app also tracks the baby's crying patterns over time, giving parents insight into their baby's needs and helping them adjust their routine accordingly. With CryBaby, parents can feel more confident and supported in their parenting journey.

How we built it

The app has 3 components: the model, the webserver, and the app itself. The app records audio of the baby crying, and then sends it to the webserver, which processes through a machine learning model.

To create our model, a Convolutional Neural Network, we acquired an unprocessed repository of baby audio data and cleaned, processed, and labeled our data. To get more data points, we augmented our data, e.g. changing the speed, pitch, or volume. We then trained our model on the dataset, reaching 82% accuracy, which is remarkable given our limited data, as well as lack of computing power.

Challenges we ran into

We struggled to find many datasets on which to train our AI. The first one we found was a study in 2007, but we felt it was lacking in the amount of sounds. Thus, we searched for and merged together other datasets, then normalized them to ensure we were training our AI with audio of similar quality. We spent a considerable amount of time augmenting and collecting data.

Accomplishments that we're proud of

While trained nurses can correctly identify the reason for a baby's cry only 30% of the time, our AI model boasts a nearly 80% accuracy rate.

What we learned

AI is a burgeoning technology, however, among the biggest things holding it back seems to be the lack of datasets to train these models.

What's next for CryBaby - Infant Interpreter

We seek to improve the user experience of the frontend application. As of now, it's fairly plain. However, we wish the app give off a more welcoming feel. Additionally, we believe the model can only get stronger and more accurate as more time and resources are put into it.