Inference playground
Score audio for synthetic speech — a calibrated verdict, where it looks synthetic, and how it was decided. Trained on 2019–2022-era synthesis.
- 1Load target
- Run detection
- Review verdict
Point the detector at a suspect clip — drop an audio file (or pick one via “Try a sample”), then run detection. Accuracy drops on audio unlike the training corpus.
Detection targetload a clip, then run detection
converted to 16-bit WAV locally — loads as the target, never auto-scores
Batch scoringqueue audio files, then score them one at a time through the same detector as Detect
Every clip is scored live by the service — nothing cached or estimated. Calibrated P(synthetic) is calibrated on In-the-Wild (~12.9% EER out-of-domain for the best single model); expect substantially worse accuracy on audio unlike the training corpus. Abstained and uncalibrated rows deliberately show no probability.
| Quality / reason | ||||||
|---|---|---|---|---|---|---|
No files queued Drop audio files above to build a scoring queue. | ||||||