The ORAVYS V26 WavLM Codec model represents a significant leap in deepfake audio detection. By combining WavLM's self-supervised speech representations with codec-aware preprocessing, we achieve an F1 score of 0.9995 on our held-out test set of over 10,000 samples.
Architecture
V26 builds on the WavLM Large backbone (316M parameters) with a custom classification head trained on our proprietary multi-corpus dataset. The codec-aware layer detects artifacts introduced by neural vocoders, speech synthesis systems, and voice conversion tools, even after compression through common audio codecs like Opus and AAC.
Training Data
Our training corpus spans 1.51M+ voice vectors across multiple domains: LibriTTS-R, VCTK, LJSpeech, CommonVoice, ASVspoof2019, and WaveFake. All datasets are commercially licensed and properly attributed.
Ensemble Integration
V26 operates as the primary detector in our DAAF (Deepfake Audio Authentication Framework) ensemble with a weight of 0.35, working alongside V25.4, V23 Meta-LoRA, V17 PyTorch, V3 ONNX, V2 ONNX, and the genuine-biased V19 model. This multi-model approach ensures robustness against adversarial attacks and novel synthesis methods.
Results
On our cross-corpus evaluation benchmark, V26 achieves:
- F1 Score: 0.9995
- Equal Error Rate: 0.03%
- False Acceptance Rate: less than 0.05%
- Latency: under 2 seconds for 30-second audio clips
These results position ORAVYS as the most accurate commercially available deepfake detection system.