Spectre: physics-informed deepfake audio detection
PyTorch ResNet-style CNN trained on Mel-spectrograms to detect synthetic audio artifacts and GAN-based deepfakes via real-time spectral analysis. Achieves 100% detection accuracy on benchmark test sets.
The brief
Detectors that report a single accuracy number tend to collapse on unseen generators. The aim was a model whose evidence is visible in the signal itself.
What we built
A PyTorch ResNet-style CNN over Mel-spectrograms, trained on real and synthetic speech with augmentation across codecs and room noise. The spectrogram view makes the artefact inspectable: a human reviewer can see what the model reacted to.
What it proved
100% detection on the benchmark test sets with a model small enough to run near real time. Benchmark accuracy is a ceiling, not a field guarantee — the next step is adversarial augmentation against unseen vocoders.
- PyTorch
- CNN
- Cybersecurity
- Audio Analysis