Architecture: AASIST (graph-attention anti-spoofing network), from-scratch trained
(not XLS-R backed). Base weights: official ASVspoof2019 LA pretrained checkpoint
from https://github.com/clovaai/aasist, [fine-tuned on: DESCRIBE YOUR DATA HERE].
Known limitations:
Fine-tuned primarily on English ASVspoof-style data; cross-lingual/accented performance not validated.
Sensitive to codec compression (Opus/AMR/MP3) unless explicitly augmented for it.
Expects 16kHz mono input, fixed 64600-sample (~4s) window via repeat-padding.
Label convention: 0 = spoof, 1 = bonafide (verify against a known sample before trusting in production).
