Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Interpretable All-Type Audio Deepfake Detection with Audio LLMs via Frequency-Time Reinforcement Learning

Yuankun Xie, Xiaoxuan Guo, Jiayi Zhou, Tao Wang, Jian Liu, Ruibo Fu, Xiaopeng Wang, Haonan Cheng, Long Ye

TL;DR

The paper tackles all-type audio deepfake detection (ADD) by leveraging audio large language models (ALLMs) and interpretable reasoning. It introduces a two-stage FT-GRPO training framework that cold-starts with supervised fine-tuning on frequency–time chain-of-thought data and then employs group-relative policy optimization with frequency-time constraints to improve both accuracy and interpretability. An automatic data construction pipeline produces ~340K FT-CoT demonstrations, enabling stable reasoning and robust cross-type generalization across speech, environmental sounds, singing, and music. Empirical results show state-of-the-art accuracy for a 3B ALLM on all-type ADD and substantial gains when co-training across types, with interpretable FT-grounded rationales. The work advances practical ADD by combining rationale generation, constraint-driven RL, and cross-type evaluation, offering a scalable path toward transparent, robust detectors for diverse audio content.

Abstract

Recent advances in audio large language models (ALLMs) have made high-quality synthetic audio widely accessible, increasing the risk of malicious audio deepfakes across speech, environmental sounds, singing voice, and music. Real-world audio deepfake detection (ADD) therefore requires all-type detectors that generalize across heterogeneous audio and provide interpretable decisions. Given the strong multi-task generalization ability of ALLMs, we first investigate their performance on all-type ADD under both supervised fine-tuning (SFT) and reinforcement fine-tuning (RFT). However, SFT using only binary real/fake labels tends to reduce the model to a black-box classifier, sacrificing interpretability. Meanwhile, vanilla RFT under sparse supervision is prone to reward hacking and can produce hallucinated, ungrounded rationales. To address this, we propose an automatic annotation and polishing pipeline that constructs Frequency-Time structured chain-of-thought (CoT) rationales, producing ~340K cold-start demonstrations. Building on CoT data, we propose Frequency Time-Group Relative Policy Optimization (FT-GRPO), a two-stage training paradigm that cold-starts ALLMs with SFT and then applies GRPO under rule-based frequency-time constraints. Experiments demonstrate that FT-GRPO achieves state-of-the-art performance on all-type ADD while producing interpretable, FT-grounded rationales. The data and code are available online.

Interpretable All-Type Audio Deepfake Detection with Audio LLMs via Frequency-Time Reinforcement Learning

TL;DR

The paper tackles all-type audio deepfake detection (ADD) by leveraging audio large language models (ALLMs) and interpretable reasoning. It introduces a two-stage FT-GRPO training framework that cold-starts with supervised fine-tuning on frequency–time chain-of-thought data and then employs group-relative policy optimization with frequency-time constraints to improve both accuracy and interpretability. An automatic data construction pipeline produces ~340K FT-CoT demonstrations, enabling stable reasoning and robust cross-type generalization across speech, environmental sounds, singing, and music. Empirical results show state-of-the-art accuracy for a 3B ALLM on all-type ADD and substantial gains when co-training across types, with interpretable FT-grounded rationales. The work advances practical ADD by combining rationale generation, constraint-driven RL, and cross-type evaluation, offering a scalable path toward transparent, robust detectors for diverse audio content.

Abstract

Recent advances in audio large language models (ALLMs) have made high-quality synthetic audio widely accessible, increasing the risk of malicious audio deepfakes across speech, environmental sounds, singing voice, and music. Real-world audio deepfake detection (ADD) therefore requires all-type detectors that generalize across heterogeneous audio and provide interpretable decisions. Given the strong multi-task generalization ability of ALLMs, we first investigate their performance on all-type ADD under both supervised fine-tuning (SFT) and reinforcement fine-tuning (RFT). However, SFT using only binary real/fake labels tends to reduce the model to a black-box classifier, sacrificing interpretability. Meanwhile, vanilla RFT under sparse supervision is prone to reward hacking and can produce hallucinated, ungrounded rationales. To address this, we propose an automatic annotation and polishing pipeline that constructs Frequency-Time structured chain-of-thought (CoT) rationales, producing ~340K cold-start demonstrations. Building on CoT data, we propose Frequency Time-Group Relative Policy Optimization (FT-GRPO), a two-stage training paradigm that cold-starts ALLMs with SFT and then applies GRPO under rule-based frequency-time constraints. Experiments demonstrate that FT-GRPO achieves state-of-the-art performance on all-type ADD while producing interpretable, FT-grounded rationales. The data and code are available online.
Paper Structure (28 sections, 6 equations, 10 figures, 9 tables)

This paper contains 28 sections, 6 equations, 10 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Audio Deepfake Detection
  4. ALLM for Audio Deepfake Detection
  5. Method
  6. Data Construction
  7. FT-GRPO
  8. Accuracy Reward.
  9. Format Reward.
  10. FT Reasoning-Path Reward.
  11. Total Reward.
  12. GRPO Objective.
  13. Experiments
  14. Dataset
  15. Baselines
  16. ...and 13 more sections

Figures (10)

  • Figure 1: Three ALLM-based CMs for the all-type ADD task: (a) label-only SFT, (b) RFT, and (c) our proposed FT-GRPO.
  • Figure 2: Pipeline overview. Top: data construction via Step-1 raw audio caption and Step-2 caption polish. Bottom: FT-GRPO training, consisting of Step-1 SFT cold start and Step-2 GRPO.
  • Figure 3: Four different training strategies in FT-GRPO.
  • Figure 4: Three fine-tunable modules in an ALLM.
  • Figure 5: Four training modes for speech-trained Qwen2.5-Omni-3B. $\checkmark$ denotes trainable and $\times$ denotes frozen. ACC is measured on the 19LA evaluation set.
  • ...and 5 more figures