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

2026-01-06

arXiv.org

AI Summary

This work addresses the challenge of generalizing audio deepfake detection across diverse audio types—including speech, environmental sounds, singing, and music—where existing methods struggle to balance performance and interpretability. The authors propose a two-stage training framework based on Audio Large Language Models (ALLMs). First, they construct interpretable supervision signals via a frequency-time structured chain-of-thought (CoT) with automatic annotation. Subsequently, they perform reinforcement fine-tuning by integrating supervised fine-tuning (SFT) with a novel Frequency-Time Grouped Relative Policy Optimization (FT-GRPO). The resulting model achieves state-of-the-art performance across all audio forgery detection tasks while generating human-interpretable reasoning grounded in frequency-time features, effectively mitigating reward hacking and hallucination issues.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Fact-Checking / Misinformation Detection (NLP Focus)
• Computer Vision: Diffusion Models for Vision

Application Category

• Economics, Online Markets and Human Computation: LLM based quality controls for crowd work
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Agentic search

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.

Problem

• audio deepfake detection
• all-type audio
• interpretability
• audio large language models
• frequency-time analysis

Innovation

• Audio LLMs
• Frequency-Time CoT
• FT-GRPO
• All-Type Deepfake Detection
• Interpretable AI