spoof-pad-wavlm-nes2net-tflite

Spoof/PAD WavLM + Nes2Net-X TFLite

This repository contains a spoof / presentation-attack-detection model exported to TFLite/LiteRT for local CPU testing.

Files

Source lineage

This is a converted derivative of a WavLM + Nes2Net-X anti-spoofing stack:

• Active checkpoint repo: nielsr/nes2net-checkpoints
• Active checkpoint file: https://huggingface.co/nielsr/nes2net-checkpoints/resolve/main/WavLM_Nes2Net_X_e75_seed420_valid0.03192785031473534.pt?download=true
• WavLM base model: microsoft/wavlm-large
• Original implementation repo / branch: https://github.com/Liu-Tianchi/Nes2Net_ASVspoof_ITW/tree/asvspoof5

GitHub implementation repos cannot be represented as Hugging Face base_model relationships, so they are listed explicitly here and in artifact_manifest.json.

Runtime contract

• Sample rate: 16000 Hz
• Input preprocessing: normalize to mono 16 kHz, then take/tile exactly 64600 samples
• Input dtype: float32
• Output: single logit interpreted as the bona-fide score
• Score direction: higher score is more bona-fide
• Thresholds are application-specific. Useful decision values include max_spoof_probability, min_bonafide_probability, and min_bonafide_logit.

Limitations

• This is an experimental converted artifact.
• License and usage terms must be checked against the upstream WavLM, Nes2Net, ASVspoof, and checkpoint sources before redistribution or production use.
• PAD thresholds are application-specific; validate on your target microphones, speakers, replay conditions, and attack types.