This project investigated both whether a speech utterance contains synthetic audio and where the altered segments occur. In a team of three, I co-authored the study using the PartialSpoof dataset and a PyTorch pipeline with Wav2Vec2-XLSR speech features.
The work adapted a Temporal Difference Attention Module and used a multi-task loss for detection and localization. We documented the methods and experiments in a research-style report and poster.
The system achieved an utterance-level Equal Error Rate of 0.79%, while frame-level EER was 17.34%. That gap was an important limitation: detecting a manipulated clip was substantially more successful than locating the exact forged regions.