Record a few seconds or upload a clip. VoiceShield flags synthetic and cloned speech in seconds, and your audio never leaves your device.
Test clips are from the ASVspoof 2019 evaluation set, never used in training: real speakers and 13 different speech-synthesis and voice-conversion systems.
| Loading… |
A graph-attention network (AASIST) reads the raw waveform: learnable band-pass filters, a residual encoder, then one graph over frequency bands and one over time frames, fused by a heterogeneous graph with a master node. It learns the tiny spectral and timing artefacts that vocoders leave behind.
Trained on ASVspoof 2019 LA (English read speech, 2019-era synthesis). Newer commercial voice-cloning tools can be harder to catch: treat this as a screening signal, not proof. Code, benchmark and limits.