# Codecfake-W2V2-AASIST-Codec — implementation notes

Codec-trained **W2V2-AASIST** of the Codecfake paper (Xie et al., arXiv 2405.04880), plan model #8, built
2026-11-01. Everything except the checkpoint is identical to `../Codecfake-W2V2-AASIST-CoTrain/` (same `_net.py`
copy; see that dir's notes for the port, XLS-R truncation bit-identity proof, allow-list unpickler, score
definition, TF32-off). Helper scripts (`run_sets.sh`, `trt_export_both.sh`, `trt_jspaw.sh`, `parity.py`) live in
the CoTrain dir and cover both models.

- Code: xieyuankun/Codecfake @ `72c32e2840eef95fbeade349425fa9c5391fa11e`, **no LICENSE** → linked, not re-hosted.
- Checkpoint: `pretrained_model/codec_w2v2aasist/anti-spoofing_feat_model.pt` (sha256 `5a606d1f…c25a3`), trained on
  Codecfake train only; fetched at load time from the pinned raw GitHub URL → `head_codec_w2v2aasist.safetensors`.
  Strict load 245 keys; pickled temperatures/pool ratios identical to the port.
- Params 76,554,634. Score = logits[:,0] − logits[:,1] (class 0 = bona fide). trained_on: none.

## Reproduction vs authors' `result/ITW_result.txt` (no 19LA file shipped for this model)
Full InTheWild (31 779): EER ours 9.274% vs authors 9.609%; Pearson (logit) 0.9989, Spearman 0.9990,
median |Δ logit| 0.044 — the same 16-bit-FLAC packaging residual measured for the CoTrain sibling (float
source wavs: median |Δ| 0.0035).

## Throughput / TRT (2026-11-01, shared GPUs)
PyTorch: same arch as CoTrain (288 / 104 utt/s on 4070 Ti SUPER / 5060 Ti). TRT sidecar
`trt_codecfake-w2v2-aasist-codec.json`:
| card | precision | best batch | utt/s | build Spearman | J-SPAW_LA EER PT → TRT | Spearman vs PT |
|---|---|---|---|---|---|---|
| RTX 4070 Ti SUPER | FP32 (FP16 0.9997 < 0.9999) | 32 | 395 | 1.0000 | 0.1252 → 0.1252 | 0.999999 |
| RTX 5060 Ti | FP16 (0.99995) | 16 | 784 | 0.99995 | 0.1252 → 0.1252 | 0.99996 |
Both accepted.