Voice-based authentication is increasingly used in banking, call centers, and smart assistants. Voice phishing (vishing) attacks are also getting more common, where grandparents are getting calls from scammers disguised as their grandchildren. This is because deep learning methods can now generate highly realistic synthetic voices, making it possible to impersonate a target and bypass voice authentication systems. Detecting these deepfake audio attacks is therefore a pressing security challenge.
This project will explore the effectiveness of AI models in distinguishing real human speech from synthetic speech. The study will involve building or collecting a dataset of real and generated audio samples, training machine learning and deep learning models for classification, and evaluating how well these models perform against different types of voice synthesis systems.
Project Goal:
- Build a dataset of real and fake audio samples.
- Benchmark state-of-the-art ML and transformer based architectures for detecting deepfake audio.
Research Tasks:
Conducting a literature review on voice authentication and audio deepfake detection.
Collecting a dataset of real vs synthetic voice samples (possibly by harvesting openly available voice samples and creating AI generated version).
Training and evaluating baseline models (e.g., spectrogram features + logistic regression, CNNs, RNNs).
Experimenting with modern approaches such as pretrained audio transformers (e.g., wav2vec2, HuBERT).