Scamurai

Inspiration

Due to the rise of AI-generated voices, scams have become a lot more convincing and harder to detect. It is especially harder for the elderly who are targeted through phone calls and voicemails. We wanted to build a tool that would help people verify whether a voice is real or fake, before they act on it.

This is how we created Scamurai, an application that detects deepfake audio and protects users from scams.

How we built it

We built Scamurai using RawNet2. It is a deep learning model that processes raw audio directly.

This is how the code works:

Upload an audio file (voicemail or recording)
Preprocessing of the file (resampling, trimming, normalization)
Deep learning model: Audio -> Sinc Filters -> Residual CNN -> GRU -> Prediction
Outputs Real or Fake with a confidence score

We trained the model using a gender-based labeled dataset of real and spoofed audio, and integrated the system with a mobile app where users can upload audio for analysis.

What we learned

Through this project, we learned:

How to preprocess raw audio for machine learning
How frequency-based filters (SincConv) can expose deepfake artifacts
How temporal models like GRUs catch inconsistencies in speech over time
How to build an end-to-end ML pipeline from data to deployment

Challenges we ran into

Working directly with raw audio was challenging. We had to manage different lengths of audio files to make sure they were consistent for training, and debug training stability and performance. Pushing the data files onto GitHub and training the model itself took a while to do, which caused a time constraint for us.

We also faced challenges in understanding and implementing the RawNet2 architecture, as well as integrating the code into a mobile application.

Accomplishments that we're proud of

We are proud of finding an extensive, gender-balanced dataset to train the model on, which caused us to get high-confidence scores when it detects real or fake audio. Due to this, we were able to meet our project and problem expectations. Furthermore, we were able to integrate it to be live, which means that our application can directly help people falling into these scams.