Add meta.yaml (HF download-stats query file) to enable download tracking

meta.yaml

@@ -0,0 +1,34 @@
+system:
+  name: "AASIST-L"
+  slug: "aasist-l"
+  description: >
+    AASIST-L: the lightweight variant of AASIST (audio anti-spoofing using
+    integrated spectro-temporal graph attention networks). Same architecture as
+    AASIST — sinc-convolution front-end, RawNet2-style residual encoder, and
+    heterogeneous stacking graph attention over spectral and temporal sub-graphs
+    with a learnable readout — but with a narrower residual stack and graph
+    dimensions (~85k params vs AASIST's ~298k). Official clovaai/aasist
+    ASVspoof2019 LA pretrained checkpoint, FP32, deterministic
+    first-64600-sample window (no random crop).
+  code: "https://github.com/clovaai/aasist"
+  checkpoint: "https://huggingface.co/SpeechAntiSpoofingBenchmarks/AASIST-L/blob/e4185b270ec20077c918e06a45093717a1bd5e30/AASIST-L.pth"
+  params_millions: 0.085306
+  paper:
+    arxiv_id: "2110.01200"
+    url: "https://arxiv.org/abs/2110.01200"
+    bibtex: |
+      @inproceedings{jung2022aasist,
+        title={{AASIST}: Audio Anti-Spoofing Using Integrated Spectro-Temporal Graph Attention Networks},
+        author={Jung, Jee-weon and Heo, Hee-Soo and Tak, Hemlata and Shim, Hye-jin and Chung, Joon Son and Lee, Bong-Jin and Yu, Ha-Jin and Evans, Nicholas},
+        booktitle={ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
+        pages={6367--6371},
+        year={2022},
+        organization={IEEE}
+      }
+notes: >
+  Lightweight AASIST-L variant (not the full AASIST). Same _net.py Model class as
+  AASIST; differs only in model_config (filts [..,[32,24],[24,24]], gat_dims
+  [24,32], pool_ratios [0.4,0.5,0.7,0.5]) and checkpoint. Deterministic
+  first-64600-sample window (no random crop), matching clovaai/aasist
+  data_utils.pad() used at eval. Checkpoint mirrored to
+  SpeechAntiSpoofingBenchmarks/AASIST-L (pinned at publish).