HIR-SDD — Towards Robust Speech Deepfake Detection via Human-Inspired Reasoning

Abstract

The modern generative audio models can be used by an adversary in an unlawful manner, specifically, to impersonate other people to gain access to private information. To mitigate this issue, speech deepfake detection (SDD) methods started to evolve. Unfortunately, current SDD methods generally suffer from the lack of generalization to new audio domains and generators. More than that, they lack interpretability, especially human-like reasoning that would naturally explain the attribution of a given audio to the bona fide or spoof class and provide human-perceptible cues. In this paper, we propose HIR-SDD, a novel SDD framework that combines the strengths of Large Audio Language Models (LALMs) with the chain-of-thought reasoning derived from the novel proposed human-annotated dataset. Experimental evaluation demonstrates both the effectiveness of the proposed method and its ability to provide reasonable justifications for predictions.

Method

HIR-SDD fine-tunes SALMONN (Whisper + BEATs encoders → Q-Former → Vicuna-7B) so it produces a binary decision and a human-interpretable rationale:

Hard-label SFT. LoRA fine-tuning to emit Final Answer: Real/Fake.
CoT SFT. Training on human-derived reasoning traces to produce structured output.
Audio grounding. Deterministic perturbations (Gaussian noise, time masking, gain) push the model to anchor its reasoning to perceptible acoustic evidence.
GRPO. Reinforcement learning with rewards for format, class correctness, and an LLM judge scoring coverage / relevance / logic / helpfulness.

<think> free-form reasoning grounded in acoustic cues </think>
<reasons>[ detected cue tags from the annotation taxonomy ]</reasons>
<answer> Real / Fake </answer>

Dataset

A human-annotated reasoning corpus for SDD, released on Hugging Face — both a processed version (CoT traces + hard labels) and the raw per-annotator annotations.

41,414

audio samples

124,410

annotations

annotators

EN / RU

languages

Curated from open SDD datasets plus newly synthesized samples, with splits for binary classification or chain-of-thought training and evaluation.

Citation

@inproceedings{hirsdd2026,
  title     = {Towards Robust Speech Deepfake Detection via Human-Inspired Reasoning},
  author    = {Dvirniak, Artem and Kushnir, Evgeny and Tarasov, Dmitrii and
               Iudin, Artem and Kiriukhin, Oleg and Pautov, Mikhail and
               Korzh, Dmitrii and Rogov, Oleg Y.},
  booktitle = {Proc. INTERSPEECH 2026},
  year      = {2026},
  eprint        = {2603.10725},
  archivePrefix = {arXiv},
  url           = {https://arxiv.org/abs/2603.10725}
}