HoliAntiSpoof: Speech Anti-Spoofing Arena wrapper

Packaged for the Speech Anti-Spoofing Arena from the authors' code and paper.

HoliAntiSpoof (Xu et al., 2026): an audio LLM for holistic speech anti-spoofing. The Qwen2.5-Omni-7B thinker (Whisper-style audio encoder + Qwen2 7B LLM) is fine-tuned to emit a structured JSON analysis of a clip (real/fake, spoofing method, spoofed regions, semantic influence): DoRA r=64 on the LLM self-attention q/k/v/o projections plus a fully fine-tuned audio encoder. Arena wrapper: one forward pass over the authors' prompt with the fixed answer prefix {"real_or_fake": " and score = logit("real") - logit("fake") of the next token (higher = bona fide; monotone in the authors' p(real)). DoRA merged into the base weights at load; BF16, layers split over two GPUs; the two answer logits are computed in FP32. Input: first 4 s (64600 samples) of the clip, no repeat-tiling, clips under 1 s zero-padded to 1 s (as the authors do).

• Weights: not re-hosted. The wrapper loads the authors' DoRA adapter (wsntxxn/HoliAntiSpoof, commit-pinned) on top of Qwen/Qwen2.5-Omni-7B (thinker only) and merges it at load.
• Params: 8255.263744 M (BF16, ~15 GB; split over two GPUs via HOLI_DEVICES / HOLI_SPLIT).
• Trained on (paper Table 3): ASVspoof 2019 LA train, PartialSpoof, SINE, WaveFake, CodecFake, PartialEdit, VCTK, LJSpeech, EmoFake, SpeechFake-en, DailyTalkEdit, FakeOrReal and all of In-the-Wild. The Arena marks InTheWild and EmoFake_test as n/a* (training overlap; left out of the Mean).
• Score: higher = bona fide; logit("real") - logit("fake") of the first answer token, in FP32. Input: first 4 s of each clip (Arena maintainer's choice; the authors allow up to 60 s). PyTorch only (no TensorRT engine: a 7B two-GPU model).

Arena results

EER% (lower is better); EmoSpoofTTS and LRLspoof are spoof-only and use 1-SRR% at the model's DeepVoice EER threshold. Per-dataset scores are under .eval_results/.

Citation

@article{xu2026holiantispoof,
  title={HoliAntiSpoof: Audio LLM for Holistic Speech Anti-Spoofing},
  author={Xu, Xuenan and Ren, Yiming and Liu, Liwei and Wu, Wen and Li, Baoxiang and Lu, Chaochao and Wang, Shuai and Zhang, Chao},
  journal={arXiv preprint arXiv:2602.04535},
  year={2026}
}