Sexual Content Audio Classifier

This repository provides the means to train a neural network that can detect whether certain sounds appear in the audio track of a video. The primary application is to detect sexual content in videos by listening for moans.

The framework can be extended to detect additional or different sounds to optimize sexual content detection.

It was created in the course of the term paper "Supporting the Detection of Sexual Content in Videos by Listening for Characteristic Sounds", which can also be found in this repository.

Requirements

Detecting Sounds

• Python 3.7

Training the Model

• Python 3.7
• cuda & cudnn libraries

Creating datasets

• Python 3.7
• Praat as a command line program.

Setup

• Create a virtual environment and install the required packages from requirements.txt.

Detecting sounds

• Execute src/predict.py with -f path/to/audio_file -m path/to/model.
• Pre-trained models are included with this repository.

Training the model

• Change parameters in src/config.py if necessary.
• Execute src/train.py.
• Execute src/validate.py.
• The trained models can be found in the dataset directory under models.

Creating datasets

python src/create_dataset_script.py -m train -f urls_train.txt -d dataset

The mode can be train or validation to create the training and validation datasets. The URL file must contain one URL per line. For each URL the audio track is downloaded and used to create the data. The resulting data is saved to the dataset directory.

The annotation of the spectrogram has to be done by hand with Praat. A tutorial for annotating with Praat can be found here. The default tiers should be replaced by a single tier named sound. The tier should only contain the label 'Moan', wherever a moan is present on the audio track. The annotated TextGrid file has to be saved to the asset directory.

Pre-trained dataset

The data set on which the accompanying paper was trained is included in the repository as a release. The data is split into two subsets, train and validation. Both sets have an images and an annotations folder. The models folder contains all pre-trained models created from the data set. Finally, in log.csv, the mapping of UUIDs to URLs the data was extracted from can be found.