Infant Crying Detection

This repository contains code to run an infant crying detection model for continuous audio in real-world environments, as described in this paper.

Citation Information

1. X. Yao, M. Micheletti, M. Johnson, E. Thomaz, and K. de Barbaro, "Infant Crying Detection in Real-World Environments," in ICASSP 2022

2. M. Micheletti, X. Yao, M. Johnson, and K. de Barbaro, "Validating a Model to Detect Infant Crying from Naturalistic Audio," in Behavior Research Methods 2022

Models and Main Package Versions

Trained deep spectrum model can be found at: https://utexas.box.com/s/poe7i9p9lzs6rg7yer85y2m9wh247lz0. Trained SVM model is in this repository: svm.joblib

Versions

python3/3.6.3
tensorflow-gpu==1.15.0
scikit-learn==0.23.0
pyAudioAnalysis==0.3.7
librosa==0.8.1

Code

There are two scripts: preprocessing.py and predict.py.

preprocessing.py aims to get rid of the seconds that are definitely not infant crying using frequency information. It reads an audio file and outputs a csv file containing the start_time (seconds) and end_time (seconds) of audio where there is some energy for signals higher than 350Hz. Change input/output filennames here:

audio_filename = 'P34_2.wav'
output_file = "preprocessed.csv"

predict.py gives predictions of crying/not crying at every second. It reads an andio file and the preprocessed csv file and outputs a csv file containing the predictions at each second with its timestamp. Change input/output filenames here:

preprocessed_file = "preprocessed.csv"
audio_filename = "P34_2.wav"
output_file = "predictions.csv"

Other resources

HomeBank English deBarbaro Cry Corpus (https://homebank.talkbank.org/access/Password/deBarbaroCry.html)
It contains part of the RW-Filt dataset that was used to create the model as 2 out of 24 participants did not give us permission to share their data.
To protect the privacy of participants, all crying episodes were cut into five second segments (with four-second overlap between neighboring segments). An equal length and number of five second segments of non-cry data was randomly selected from the same recording. The complete dataset totals 61.3h of labelled data with over seven hours of unique annotated crying data.