Automated Robocall Campaign Detection using Audio Embeddings

This repository contains a Proof of Concept (PoC) of uncovering robocall campaigns from raw robocall recordings based on audio similarity. The example code demonstrate the following:

• How to compute audio embeddings using two pre-trained (and fine-tuned) models using Wav2Vec2 and WavLM (on CPU and GPU)
• How to aggregate the embeddings into robocall campaigns

Dataset Details

To demonstrate this code, the dataset from Robocall Audio from the FTC’s Project Point of No Entry (GitHub link) is used.

How to run this example?

• Extract the raw audio recordings in FTC-raw-audio-ppone-normalized.zip (Google Drive link) or download audio files from robocall-audio-dataset.
• Install the relevant dependencies
• Run the example code Robocall_Campaign_Detection_GPU_and_CPU.py

Questions?

The example code is part of the paper titled "Characterizing Robocalls with Multiple Vantage Points". The paper was published at the IEEE Security & Privacy 2025 conference.

Please refer to the paper for additional details (evaluation, scaling, etc). If you found this artifact useful, please cite the paper!