Detecting Screams From Home Audio Recordings to Identify Tantrums: Exploratory Study Using Transfer Machine Learning.

Area of Science:

• Machine learning applications in behavioral science.
• Audio signal processing for behavioral analysis.

Background:

• Traditional methods for assessing children's behavioral disorders rely on qualitative reports, which can be inconvenient and prone to biases.
• A data-driven approach using machine learning can offer a more objective and efficient quantification of behavioral disorders.

Purpose of the Study:

• To evaluate the efficacy of a machine learning model in detecting screams from audio streams.
• To determine if a model trained on public audio datasets can generalize to at-home recordings.

Main Methods:

• Utilized a subset of the AudioSet dataset and audio from the TV show Supernanny for model training and validation.
• Employed a convolutional neural network for audio feature extraction and a gradient-boosted tree model for scream classification.
• Manually annotated and refined scream events in the datasets for accurate model training.

Main Results:

• The model achieved a receiver operating characteristic (ROC)-area under the curve (AUC) of 0.86 on the AudioSet data.
• Achieved an ROC-AUC of 0.95 and 42% average precision on Supernanny audio, demonstrating strong performance in detecting rare scream events.

Conclusions:

• A machine learning model trained on public data shows promise for identifying behavioral indicators like tantrums in audio recordings.
• This approach can supplement or replace the collection of costly, privacy-protected clinical data for behavioral disorder assessment.