Add codec-residual source-tracing model (MLAAD v5 + STOPA)

README.md

@@ -0,0 +1,90 @@
+---
+license: mit
+library_name: sourcetrace
+pipeline_tag: audio-classification
+tags:
+  - audio
+  - audio-deepfake-detection
+  - source-tracing
+  - open-set-recognition
+  - speech
+  - mlaad
+  - stopa
+base_model:
+  - microsoft/wavlm-large
+  - facebook/encodec_24khz
+datasets:
+  - mueller91/MLAAD
+language:
+  - en
+---
+
+# Codec-Residual Open-Set Audio Deepfake Source Tracing
+
+Fitted model checkpoint for **open-set audio deepfake source tracing** — attributing a
+synthetic utterance to the **generator** (acoustic model × vocoder) that produced it, and
+**rejecting** generators unseen in training. Evaluated on **MLAAD v5** (source tracing) and
+**STOPA** (speaker/attack tracing).
+
+- **Code / reproduction:** https://github.com/pujariaditya/ood_source_tracing
+- **Checkpoint:** `mlaad_v5.pt` — the trained factorized gated head (frozen WavLM-Large front-end
+  is loaded separately from `microsoft/wavlm-large`; not stored here).
+
+## Results (this checkpoint, vs published SOTA)
+
+| Benchmark | Metric | This model | SOTA |
+|-----------|--------|-----------|------|
+| MLAAD v5 | FPR95 | **0.75%** | 3.36% |
+| MLAAD v5 | OOD-EER | 2.81% | — |
+| STOPA | unknown-attack EER | **9.38%** | 16.43% |
+
+SOTA: MLAAD v5 FPR95 = PANDA / Neamtu-2026 (arXiv 2606.10758); STOPA unknown-attack EER =
+Firc/Chhibber (arXiv 2509.24674). Numbers are at the paper's protocol scale.
+
+## Method
+
+Frozen **WavLM-Large** (layers 4-7, mean|std pooling) + a classical **phase/modulation
+signature** channel + a novel **neural codec-reconstruction-residual** channel (EnCodec-24kHz
+at 3 bandwidths). These feed a **factorized gated embedding head** (acoustic-model / vocoder /
+codec sub-spaces with AM-driven FiLM gates). Open-set scoring fuses a per-class **conformal
+margin** with a **relative-Mahalanobis** density (with language-covariance inflation). See the
+GitHub repo's `docs/METHOD.md` for details.
+
+## Usage
+
+```bash
+git clone https://github.com/pujariaditya/ood_source_tracing
+cd ood_source_tracing
+pip install -e .
+
+# fetch this checkpoint
+hf download RootAccess4Life/ood-source-tracing mlaad_v5.pt --local-dir checkpoints
+
+# evaluate (reproduces v5-FPR 0.75)
+python scripts/evaluate.py --task mlaad_v5 --checkpoint checkpoints/mlaad_v5.pt
+```
+
+Programmatic load:
+
+```python
+from sourcetrace.method import Method
+model = Method.load("mlaad_v5.pt")
+scores = model.score_openset(features)   # higher = more likely a known generator
+```
+
+The checkpoint is a `torch.save` dict produced by `Method.save` (trained head weights + the
+conformal calibration, relative-Mahalanobis density, PCA-whiten transfer, and codec
+normalization statistics). It can also be regenerated from scratch: `python scripts/train.py
+--task mlaad_v5 --seed 0`.
+
+## Base models & data
+
+- Front-end: [`microsoft/wavlm-large`](https://huggingface.co/microsoft/wavlm-large)
+- Codec channel: [`facebook/encodec_24khz`](https://huggingface.co/facebook/encodec_24khz)
+- Training data: [MLAAD](https://huggingface.co/datasets/mueller91/MLAAD) (v5, PANDA
+  family-level split) and [STOPA](https://doi.org/10.5281/zenodo.15606628).
+
+## License
+
+MIT (see the GitHub repository). Base-model and dataset licenses are held by their respective
+providers.

mlaad_v5.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1bcddc3ab369066cc5d4fdb4befbc1c2a7972db62e4ff88375744789511ab04d
+size 78364343