Russian Neural Network “Detective” Aims to Unmask Audio Deepfakes in Real Time

At Astrakhan State Technical University, engineers are building a neural network module capable of identifying manipulated audio files. The system analyzes sound in real time and generates a detailed report highlighting signs of falsification.

The project is led by Olga Kurmanova, a postgraduate researcher in the Department of Automated Information Processing and Control Systems, under the supervision of Nadezhda Davidyuk, head of the university’s Information Security department and a candidate of technical sciences.

A Neural Network Against Deception

The team assembled a dataset of 70,000 audio files, including original recordings and samples generated using modern speech synthesis algorithms. The model was trained on this data to detect subtle digital artifacts that are inaudible to the human ear.

The system can identify several types of falsification, including fully synthesized speech generated through TTS technology, voice cloning with altered timbre, digital editing and splicing, where fragments from different recordings are merged into a single file.

Unlike many existing services, the module works exclusively with audio and does not analyze video, making it highly specialized. It does not simply label a file as “fake” or “original.” Instead, it examines the structure of the signal and indicates which parameters triggered suspicion.

A Tool Against Fraud

Developers emphasize that fake voice messages are increasingly used in scams. Fraudsters imitate the voices of relatives or colleagues in messaging apps and phone calls. The new tool is intended to help identify such recordings more quickly and expose scammers.

The university plans to expand the project. In the future, the team aims to build a broader decision-support system based on the module, capable of analyzing audio, video, and text. The platform would help combat digital forgeries across multiple formats.