Source-linked AI summary
ALLM4ADD: Unlocking the Capabilities of Audio Large Language Models for Audio Deepfake Detection
Hao Gu, Jiangyan Yi, Chenglong Wang, Jianhua Tao, Zheng Lian, Jiayi He, Yong Ren, Yujie Chen, Zhengqi Wen
TL;DR
High-fidelity generated speech can be misused, making audio deepfake detection important, but ALLMs' ability to detect it was uncertain. The paper evaluates zero-shot models and proposes ALLM4ADD, which frames detection as audio question answering and uses supervised fine-tuning. It reports superior performance to existing methods, including with limited training data.
Problem
Although ALLMs have advanced in audio processing, their effectiveness for audio deepfake detection was unexplored, despite the growing misuse risks of high-fidelity generated speech.
Source: §1 · pp. 1–2Method
ALLM4ADD casts audio deepfake detection as audio question answering and uses supervised fine-tuning to train ALLMs to judge query audio as fake or real.
Source: §1 · p. 2, §4.3 · p. 4Results
ALLM4ADD outperforms existing methods, achieving an EER below 5% and accuracy over 96% with around 200 training samples.
Source: §7 · p. 8Takeaways & Limitations
Abstract
from arXiv · showhide
Audio deepfake detection (ADD) has grown increasingly important due to the rise of high-fidelity audio generative models and their potential for misuse. Given that audio large language models (ALLMs) have made significant progress in various audio processing tasks, a heuristic question arises: \textit{Can ALLMs be leveraged to solve ADD?}. In this paper, we first conduct a comprehensive zero-shot evaluation of ALLMs on ADD, revealing their ineffectiveness. To this end, we propose ALLM4ADD, an ALLM-driven framework for ADD. Specifically, we reformulate ADD task as an audio question answering problem, prompting the model with the question: ``Is this audio fake or real?''. We then perform supervised fine-tuning to enable the ALLM to assess the authenticity of query audio. Extensive experiments are conducted to demonstrate that our ALLM-based method can achieve superior performance in fake audio detection, particularly in data-scarce scenarios. As a pioneering study, we anticipate that this work will inspire the research community to leverage ALLMs to develop more effective ADD systems. Code is available at https://github.com/ucas-hao/qwen_audio_for_add.git
1 Introduction
High-fidelity TTS and voice conversion increase the urgency of detecting audio deepfakes because they can be misused to spread misinformation and undermine trust. Although ALLMs have advanced in other audio tasks, their ADD capability was unexplored; this paper evaluates them and proposes ALLM4ADD.
1 Introduction
- High-fidelity TTS and voice conversion can be misused for misinformation, social unrest, and undermining trust in digital media, making ADD increasingly urgent.Source: §1 · p. 1
- Although ALLMs have progressed in audio captioning and speech recognition, their performance on ADD had remained unexplored.Source: §1 · p. 2
- ALLM4ADD reformulates ADD as audio question answering and uses supervised fine-tuning to train ALLMs to answer whether query audio is fake or real.The paper reports that it outperforms conventional pipeline and end-to-end models, particularly when data is scarce.Source: §1 · p. 2
- Current ALLMs perform poorly in zero-shot fake audio detection, which the authors attribute primarily to a mismatch between pretraining objectives and detection requirements.Source: §1 · p. 2
2 Related Work
Prior ADD methods use either a feature-extractor/classifier pipeline or end-to-end waveform processing. ALLMs have advanced in audio tasks, but their use for fake-audio detection remained unexplored.
2.1 Audio Deepfake Detection Methods
- Conventional ADD pipelines combine a front-end feature extractor with a back-end classifier, while end-to-end models jointly optimize feature extraction and classification from raw waveforms.Source: §2.1 · p. 2
- Pipeline feature extraction captures audio fake artifacts; features may be handcrafted or deep, and handcrafted representations can be biased by their design limitations.Wav2vec2 and Hubert are cited as widely used pretrained self-supervised speech models; LCNN is a widely used back-end classifier.Source: §2.1 · p. 2
- End-to-end ADD models process raw audio, with examples using Sinc-Layers, spectral and temporal sub-graphs, or convolutional and Transformer structures.Source: §2.1 · p. 2
2.2 Audio Large Language Model
- Audio LLM approaches have included using LLMs as interfaces to external tools, while end-to-end ALLMs support direct speech interaction; examples include SpeechGPT, LTU, and Qwen-Audio.LTU applies supervised fine-tuning to audio modules and LoRA adapters; Qwen-Audio develops a unified audio-text multilingual model.Source: §2.2 · p. 2
- Despite progress in audio captioning and speech recognition, ALLMs had not been explored for fake-audio detection; this paper formulates ADD as audio question answering.Source: §2.2 · p. 2
3 Can ALLMs detect fake audio zero-shot?
The paper evaluates zero-shot fake-audio detection by Qwen-audio and Qwen2-audio models on ASVspoof2019 across prompt templates. Both series perform poorly, including a Qwen-audio-chat model that labels every trial real.
3 Can ALLMs detect fake audio zero-shot?
- The evaluation categorizes predictions as true positives, true negatives, false positives, false negatives, or failures when the model cannot decide.Source: §3 · p. 3
- The evaluation tests Qwen-audio and Qwen2-audio checkpoints on ASVspoof2019 LA, reporting accuracy and modified F1 across prompt templates.The prompt templates ask the model to classify audio authenticity as fake or real; ambiguous responses are further assessed using gpt-3.5-turbo-0125.Source: §3 · p. 3
- Qwen-audio-base averages 11.34% accuracy and 4.60% mF1 across five prompts, while Qwen-audio-chat misclassifies all audio as real.The authors attribute the difficulty primarily to ALLMs not being designed for deepfake detection.Source: §3 · p. 3, Table 1 · p. 3
4 Method
ALLM4ADD formulates audio deepfake detection as audio question answering, mapping query audio and an instruction to a binary “Fake” or “Real” response. It uses a Qwen-Audio-like encoder–LLM architecture, supervised fine-tuning with LoRA, and point-wise scoring to evaluate authenticity.
4.1 Task Formulation
- The task is framed as audio question answering: given query audio and an instruction, the model outputs “Fake” or “Real.”During training, these two text options align with the original binary classification labels.Source: §4.1 · pp. 3–4
4.2 Model Architecture
- The architecture pairs an audio encoder that converts raw audio into hidden representations with an LLM that processes those representations and instructions to generate an authenticity response.The encoder output has length τ and hidden size d, with τ ≪ T; the described Qwen-Audio variants use Whisper encoders sampled at 16,000 Hz.Source: §4.2 · p. 4
4.3 Supervised Fine-tuning
- Supervised fine-tuning pairs each audio with an instruction and ground-truth “Real” or “Fake” response, then minimizes language-modeling loss.LoRA updates low-rank trainable matrices while the pretrained LLM weights remain fixed; the audio encoder may be trainable (ALLM4ADD△) or frozen (ALLM4ADD★).Source: §4.3 · p. 4
4.4 Evaluation
- Evaluation adds accuracy and AUC alongside EER because fake-sample predominance can let a model attain low EER by mostly predicting fake.For point-wise scoring, the method applies a bidimensional softmax to the “Fake” and “Real” answer-token scores and uses their probabilities to compute metrics.Source: §4.4 · pp. 4–5
5 Experiments
Experiments compare ALLM4ADD with conventional pipeline and end-to-end baselines on ASVspoof2019 LA, using full and sampled training sets. The trainable-encoder variant outperforms both baseline families and retains strong performance with limited data.
5.1.1 Datasets.
- ASVspoof2019 LA contains 19 spoofing attack algorithms across TTS and VC, with training, development, and evaluation subsets.The experiments also sample one quarter, one eighth, or one sixteenth of each real and fake class from the training set while keeping development and evaluation sets fixed.Source: §5.1.1 · p. 5
5.1.2 Baselines.
- The conventional pipeline baselines combine handcrafted or pretrained front-end features with GF or LCNN classifiers, while end-to-end baselines operate directly on raw audio.The end-to-end set comprises RawNet2, AASIST, RawGAT-ST, Rawformer, and RawBMamba.Source: §5.1.2 · p. 5
5.1.3 Evaluation Metric.
- EER, accuracy, and AUC evaluate fake-audio detection, with lower EER and higher accuracy and AUC indicating better performance.Baseline training follows the stated hyperparameters or official code, and no method uses data augmentation.Source: §5.1.3 · p. 5, §5.2 · p. 6
5.3 Experimental Results
- 0.41% EER: with full training data, ALLM4ADD△ outperforms the best end-to-end and conventional pipeline baselines, at 0.83% and 1.94% EER, respectively.At ASV@1/16, it achieves 2.45% EER and 98.52% accuracy.Source: §5.3 · p. 6, Table 3 · p. 6
5.2 Implementation Details
- The trainable-encoder ALLM4ADD△ has an average EER 0.81% lower than frozen-encoder ALLM4ADD★ across the four training-data settings.The authors report that performance drop as training data decreases is more moderate for ALLM4ADD than for end-to-end and conventional pipeline baselines.Source: §5.2 · pp. 5–6, §5.3 · p. 6
6 Ablation Study
The ablation study poses six questions covering generalization, prompts, LoRA rank, backbones, other fake-audio datasets, and extremely limited data. Its supplied setup describes evaluation scopes and table contents, but not the corresponding outcomes.
6 Ablation Study
- The ablation study asks six questions about generalization, prompt templates, LoRA rank, ALLM backbones, other fake-audio datasets, and extremely limited data.The supplied passages state the questions but do not provide their answers.Source: §6 · p. 6
6.1.1 Experimental Setup.
- The In-the-Wild evaluation uses models trained on ASV@full or ASV@1/16 and tests them on 19,963 genuine and 11,816 fake audio files.Table 4 reports EER, accuracy, and AUC for models and baselines on this dataset.Source: §6.1.1 · pp. 6–7, Table 4 · p. 7
- Figure 2 covers LoRA ranks 1, 4, 16, 64, and 256, with EER measured across four ASVspoof2019 LA training-data settings.The supplied caption identifies the rank and setting dimensions but gives no EER values.Source: Figure 2 · p. 7
- Table 5 reports EER for different prompt templates across ASV@full, ASV@1/4, ASV@1/8, and ASV@1/16.The supplied passage identifies the table scope but does not provide its values.Source: Table 5 · p. 7
6.2 Effect of Prompt Templates (Q2)
The prompt-template experiment isolates prompt interaction with the language model using a frozen audio encoder. Across the tested data settings, prompt performance varies slightly, and Prompt 1 is best except at ASV@1/16.
6.2.1 Experimental Setup.
- The experiment compares prompts 1 through 5, omitting the final sentence “Answer fake or real.”Source: §6.2.1 · p. 7
- The audio encoder is frozen and only the LLM is fine-tuned, isolating prompt-template interaction with the LLM.Source: §6.2.1 · p. 7
- This frozen-encoder setup corresponds to ALLM4ADD★ and uses LoRA for LLM fine-tuning.Source: §6.2.1 · p. 7
- The prompt comparison reports EER on the ASVspoof2019 LA evaluation set.Source: §6.2.1 · p. 7
- The study reports equal error rate (EER, %) on the ASVspoof2019 LA evaluation set across different settings, using a frozen audio encoder and LoRA fine-tuning of the LLM.Source: §6.2.1 · p. 7
6.2.2 Experimental Results.
- Prompt 1 is best in every training-data setting except ASV@1/16.The experiment reports slight performance variations among templates overall.Source: §6.2.2 · p. 7
- The results indicate that no single prompt is best across all four training-data settings.Prompt 1 is the exception at ASV@1/16.Source: §6.2.2 · p. 7
6.3 Effect of LoRA Rank (Q3)
The experiments examine LoRA rank and backbone choice for ALLM4ADD. Higher ranks reduce EER, while Qwen-audio-chat outperforms Qwen-audio-base across the tested data settings and encoder configurations.
6.3.1 Experimental Setup.
- The rank experiment tests values 1, 4, 16, 64, and 256 using ALLM4ADD★ on the ASVspoof2019 LA evaluation set across four training-data sizes.EER is the reported metric; the four settings are ASV@full, ASV@1/4, ASV@1/8, and ASV@1/16.Source: §6.3.1 · p. 7
6.3.2 Experimental Results.
- Increasing LoRA rank progressively decreases EER, with a more pronounced performance effect at ASV@full.The authors attribute the improvement to more trainable parameters enabling finer fake-audio distinctions.Source: §6.3.2 · p. 7
6.4.1 Experimental Setup.
- The backbone comparison covers Qwen-audio-chat and Qwen-audio-base at ASV@full and ASV@1/16; further ALLM evaluation is reserved for future work.ALLM4ADD★ freezes the audio encoder, whereas ALLM4ADD△ makes it trainable.Source: §6.4.1 · p. 7
6.4.2 Experimental Results.
- Qwen-audio-chat consistently outperforms Qwen-audio-base at ASV@full and ASV@1/16, whether the audio encoder is frozen or trainable.With a trainable encoder at ASV@full, chat achieves 0.41% EER versus 1.29% for base.Source: §6.4.2 · p. 7, Table 6 · p. 8
6.5 Performance on other Fake Datasets (Q5)
ALLM4ADD is evaluated on SceneFake and EmoFake, which represent acoustic-scene and emotional-characteristic changes. It achieves the lowest EER on both datasets at full and 1/16 training-data settings.
6.5.1 Experimental Setup.
- SceneFake changes an utterance’s acoustic scene, while EmoFake modifies its emotional characteristics; each 1/16 setting samples real and fake files from its training set.Source: §6.5.1 · p. 8
- ALLM4ADD achieves the lowest EER on SceneFake and EmoFake at both full and 1/16 training-data settings.This comparison includes both ALLM4ADD★ and ALLM4ADD△ against several baselines.Source: §6.5.1 · p. 8
- At the full-data setting, ALLM4ADD△ records 8.14% EER on SceneFake and 0.86% on EmoFake, versus competing end-to-end models’ best EERs of 11.38% and 1.22%, respectively.Source: §6.5.1 · p. 8
6.6 Performance on extremely limited data (Q6)
The experiments test ALLM4ADD as training data shrinks to fractions as small as 1/128. Performance tends to decline with less data, but at 1/128 the trainable-encoder version retains below 5% EER and above 96% accuracy.
6.6.1 Experimental Setup.
- The tested training-set fractions range from the full dataset down to 1/128; the 1/128 set contains 178 fake and 20 real audio samples.For 1/32, 1/64, and 1/128, models are trained for 15 epochs with learning rates adjusted to data volume.Source: §6.6.1 · p. 8
- Figure 3 compares ALLM4ADD△ and ALLM4ADD★ across the data proportions.Source: Figure 3 · p. 8, §6.6.1 · p. 8
6.6.2 Experimental Results.
- At ASV@1/128, ALLM4ADD△ maintains EER below 5% and accuracy above 96%.The authors report that performance tends to decline as the data ratio decreases.Source: §6.6.2 · p. 8
7 Conclusion
The paper evaluates ALLMs for audio deepfake detection and introduces ALLM4ADD, which frames detection as audio question answering. Its experiments report superior performance to existing methods, particularly with scarce data, and the authors identify explanation-capable detection as future work.
7 Conclusion
- The paper’s zero-shot evaluation finds limitations in ALLMs’ fake-audio detection capabilities.Source: §7 · p. 8
- ALLM4ADD reformulates audio deepfake detection as an audio question-answering problem and is reported to outperform existing methods, particularly in data-scarce scenarios.The conclusion reports below 5% EER and above 96% accuracy with around 200 training samples.Source: §7 · p. 8
- The authors plan to develop an audio deepfake detection system that also provides explanations.Source: §7 · p. 8