Baby cry detection - Building the model

Recognition of baby cry from audio signals

The aim is to automatically recognize a baby crying while sleeping. In such case, a lullaby is played to calm the baby down.

This is done by implementing a machine learning algorithm on a Raspberry Pi. The idea is to train a model on a computer and to deploy it on Raspberry Pi, which is used to record a signal and use the model to predict if it is a baby cry or not. In the former case a lullaby is played, in the latter the process (recording and predicting steps) starts again.

Code organisation

The code is organised as follows.

• ./baby_cry_mlflow/pc_main and ./baby_cry_mlflow/pc_methods folders: to run on a computer, they implement the training part
• ./baby_cry_mlflow/rpi_main and ./baby_cry_mlflow/rpi_methods folders: to run on a Raspberry Pi, they implement the predicting part

TRAINING

It includes all the steps required to train a machine learning model. First, it reads the data, it performs feature engineering and it trains the model.

The model is saved to be used in the prediction step. The training step is performed on a powerful machine, such as a personal computer.

Code to run this part is included in pc_main and pc_methods.

PREDICTION

It includes all the steps needed to make a prediction on a new signal. It reads a new signal (9 second long), it cuts it into 5 overlapping signals (5 second long), it applies the pipeline saved from the training step to make a prediction.

The prediction step is performed on a Raspberry Pi 2B. Please check baby_cry_rpi for deployment on Raspberry Pi.

Code to run this part is included in rpi_main and rpi_methods.

SIMULATION

There is a script to test the prediction step on your computer before deployment on Raspberry Pi.

A script prediction_simulation.py and 2 audio signals are provided in folder ./baby_cry_mlflow/prediction_simulation.

Run

To make it run properly, clone this repo in a folder. In the same parent folder you should also create the following tree structure:

• PARENT FOLDER
  • baby_cry_mlflow this cloned repo
  • output
    • dataset
    • model
    • prediction
  • recording

From your command line go to baby_cry_mlflow folder and run the following python scripts.

TRAINING

This step allows you to train the model. Please note that the model itself is not provided.

# Create and save trainset
python baby_cry_mlflow/pc_main/train_set.py

# Train and save model
python baby_cry_mlflow/pc_main/train_model.py

Script train_set.py saves the trainset in folder dataset and, script train_model.py saves the model in folder model. Folders dataset and model are parameters with default values that fits with the organisation shown above, they can be changed as wished.

PREDICTION

This step is to be executed on Raspberry Pi. Please refer to baby_cry_rpi

SIMULATION

This step allows you to test the model on your computer. It uses scripts from rpi_methods folder.

python baby_cry_mlflow/prediction_simulation/prediction_simulation.py

Logs

Log files are created for each step, they are saved in folder baby_cry_mlflow.

Part of the data used for training comes from ESC-50: Dataset for environmental sound classification