New acoustic simulation trains replay speech detectors without real data

Researchers Michael Neri and Tuomas Virtanen have developed a new acoustic simulation framework that generates synthetic multi-channel replay speech data to train robust voice assistant security systems. Replay attacks—where an adversary records a user's voice and plays it back to trick a voice-controlled device—pose a growing threat in smart environments. Traditional single-channel detectors struggle to generalize across different acoustic conditions, and multi-channel datasets are scarce. The framework solves this by simulating realistic multi-channel replay scenarios using publicly available resources, enabling training without any real-world recordings.

Using this synthetic data, the authors trained the state-of-the-art multi-channel replay detector M-ALRAD and tested it on the real ReMASC corpus. To better exploit spatial cues, they extended M-ALRAD with inter-channel phase difference features computed for adjacent microphone pairs, adding directional awareness to the beamformed representation. The results show that the synthetic-only training generalizes effectively to real environments, marking a significant step toward scalable, dataset-free defenses for voice assistants. The framework and datasets have been released publicly, and the work is accepted at IEEE MMSP 2026.