Pellav3: audio deepfake detector

Pellav3 looks at every clip through two views at once: the raw waveform and the mel spectrogram. Each view is encoded into a frame sequence, the two sequences are fused frame by frame into one stream, and attentive statistics pooling turns that stream into a single score.

1.54 billion parameters, 16 kHz input.

Files

  • pellav3_detector.pt: the model weights (6.2 GB, fp32).
  • pellav3_infer.py: ready to run inference script (model class included).
  • config.json: model identifier.

Usage

pip install torch torchaudio transformers==4.48.3 numpy scipy pyloudnorm huggingface_hub
python pellav3_infer.py your_clip.wav

The weights are downloaded from the Hub on first use (ffmpeg must be on the PATH). The script prints p_fake per clip and labels a clip FAKE at p_fake of 0.5 or more.

Previous version

Pellav2 (August 2026): https://huggingface.co/Sadanie/pellav2-audio-deepfake-detector

Contact

For questions, feedback or API access, email info@pellaresearch.com.

Downloads last month
3,695
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support