There is a lot of work done to detect fake news and fake video, but not enough to detect fake voices which is extremely dangerous for people's reputations.

What it does

A chrome extension will alert you while watching a video if it detects the audio is fake/synthesized or real.

How we built it

We use multiple 1D convolutional neural network layers followed by two fully connected layers and a third fully connected output layer. We implement the model architecture with Keras. Our approach achieves 94% binary classification accuracy on the ASVspoof 2019 real PA dataset, making it excellent at detecting most synthesized voices. We used the Chrome API on a front-end app to create the chrome extension. The chrome extension notices when you land on a youtube video, takes the ID of the video, makes a request to a Flask Server, the server downloads the youtube video and converts it to an audio file, then converts it into a Numpy set of sound waves, which is fed through the model, and the result is returned to the extension.

Challenges we ran into

Chrome Extension Devtools was awfully buggy. Deploying the application. Converting the downloaded file into Audio. Deploying the model.

Accomplishments that we're proud of

Extremely good results on our neural network when we only started researching an architecture on the bus ride here. Very pretty chrome extension. Very complex software system.


Built With

Share this project: