Re-recorded Dataset for Speech Tampering Detection and Classification

Each sample in the dataset contains the corresponding Mel spectrogram and a ground-truth mask that marks the tampered regions. Because the time axis of the spectrogram maps directly to audio duration, the mask can indicate the temporal positions of tampering.Collecting large-scale replay data in real environments involves practical difficulties such as venue arrangement, device coordination, and high labor costs.

To address this, we use acoustic simulation based on the Image Source Method (ISM) to mimic the indoor replay process. This method computes the Room Impulse Response (RIR) to reproduce the propagation paths, reflection delays, and energy attenuation of sound waves in enclosed spaces, allowing us to generate samples with realistic replay characteristics under controllable conditions. Observing the spectrograms, the simulated replay produces blurred boundaries and horizontal artifacts highly similar to those caused by real replay, indicating that the simulation effectively restores the interference introduced by actual replay, and is even more severe in certain aspects.The dataset contains 14,000 replayed audio samples in total, with 3,500 samples each for deletion tampering, same-source splicing, cross-source splicing, and untampered audio.