SLIM: Style-Linguistics Mismatch Model for Generalized Audio Deepfake Detection

Audio deepfake detection (ADD) is crucial to combat the misuse of speech synthesized from generative AI models. Existing ADD models suffer from generalization issues, with a large performance discrepancy between in-domain and out-of-domain data. Moreover, the black-box nature of existing models limits their use in real-world scenarios, where explanations are required for model decisions. To alleviate these issues, we introduce a new ADD model that explicitly uses the StyleLInguistics Mismatch (SLIM) in fake speech to separate them from real speech. SLIM first employs self-supervised pretraining on only real samples to learn the style-linguistics dependency in the real class. The learned features are then used in complement with standard pretrained acoustic features (e.g., Wav2vec) to learn a classifier on the real and fake classes. When the feature encoders are frozen, SLIM outperforms benchmark methods on out-of-domain datasets while achieving competitive results on in-domain data. The features learned by SLIM allow us to quantify the (mis)match between style and linguistic content in a sample, hence facilitating an explanation of the model decision.

Paper

Similar papers

Peer review

Reviewer 1UM96/10 · confidence 5/52024-07-09

Summary

This paper proposed a new approach named Style-Linguistic Mismatch (SLIM) for generalizable audio deepfake detection. The authors claimed that a certain dependency between linguistic information and style information can generalize well for audio anti-spoofing tasks. Additionally, the proposed method can also explain the final decision of the deep learning model. In order to prove the claim, a proof-of-concept experiment was conducted to show that real audio has a higher correlation coefficient between linguistics and style in Table 1. Based on this, the authors proposed a two-stage learning framework. In the first stage, the dependency is captured by two compression modules based on SSL features. In the second stage, a simple projection head is trained on the combination of the extracted dependency with the SSL features. The authors conducted sufficient experiments to demonstrate the generalization ability of the proposed method. Two in-domain evaluation datasets and two out-domain evaluation datasets were used for this point. Besides, the performance beats several SOTA baseline methods on these datasets, enhancing the reliability of this work. The authors also provided enough analysis and visualization to prove their hypothesis. Overall, it is a good paper for generalizable audio deepfake detection task.

Strengths

1. Explicitly explored the dependency between style and linguistics for audio deepfake detection 2. Pay attention to interpretable and generalizable audio deepfake detection simultaneously, which is novel to the community. 3. Good performance was shown in the experiments, especially significantly improving the performance on the out-domain datasets, which is quite crucial for anti-spoofing tasks.

Weaknesses

1. The first training stage adopts an idea from an abnormal detection task, in which only real audio (normal data) was used for training. However, the real normal data is of a much larger order of magnitude than that used in the experiment. These corner cases should be considered in the work.

Questions

I am curious whether the author applied their method to other unseen attacking scenarios. For example, in [1], the author proposed a new unseen scenario, where some audio genres in the evaluation dataset are not presented in the training dataset. I think if the author can provide some results on this scenario, it would be more convincing for their proposed method. [1] Zeng, Chang, et al. "Improving Generalization Ability of Countermeasures for New Mismatch Scenario by Combining Multiple Advanced Regularization Terms." arXiv preprint arXiv:2305.10940 (2023).

Rating

6

Confidence

5

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes.

Reviewer chQg7/10 · confidence 3/52024-07-12

Summary

The paper suggests a novel method for detecting synthesized speech. Namely, the framework introduced in the paper allows the detection of a statistically significant mismatch between the style (i. e. paralinguistic attributes) and linguistic characteristics of synthesized speech samples, which helps to differentiate them from the real speech samples. The framework is based on frozen pre-trained SSL encoders and relatively small learnable parts that make the experiments computationally feasible for a small cluster. The authors compare their method with several baselines, and it outperforms SoTa model in a cross-domain setup.

Strengths

- The idea of the proposed method is non-trivial and allows us to better understand the differences between real and synthesized speech. - The proposed method outperforms SoTa on out-of-domain data and is on par with SoTa on in-domain data. - The evaluation is good, and the method is compared with a number of very decent baselines. - The paper is well-written and easy to follow.

Weaknesses

- The Analysis section is somewhat limited (especially the "Interpretation of model decisions" part). The authors claim that the success of their method is connected with very particular artifacts in synthesized speech; however, this point is weakly supported since only four small cherry-picked correctly classified examples were provided (see the "Questions" part).

Questions

Questions: - Is the confidence of your detector connected with the severity of the artifacts in the synthesized speech samples? Do the mel spectrograms of the most confidently correctly detected synthesized samples contain the most clearly visible artifacts? - Which types of TTS models are easier to detect, and which are harder to detect by your method? - You provided examples of correctly classified speech samples. What about incorrectly classified ones? Can you share an explanation or guess what are the peculiarities of these "complicated" samples that prevent them from being correctly classified by your method? - Do you plan to upload your code to Git Hub? Suggestions for the current paper: - I suggest adding the visual analysis of mel spectrograms of the **incorrectly** classified speech samples to the paper. Suggestions for future research: - Your current research makes a step forward toward more explainable synthesized speech detection. Another step in the same direction was made in Topological Data Analysis for Speech Processing by Tulchinskii et al. (2023). They have shown that synthetic and real speech samples can be separated using the barcodes of the attention maps of the HuBERT model. The sum of the bars in the barcode for real speech is bigger than the sum of bars in the barcode for the synthesized speech - at least on some attention heads of the HuBERT. It would be interesting to investigate in more detail, how the topological characteristics (i. e. barcodes or other characteristics) of the speech embeddings are connected with the linguistic and paralinguistic properties of the speech.

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors addressed the limitations adequately.

Reviewer 4aqZ7/10 · confidence 3/52024-07-18

Summary

"SLIM: Style-Linguistics Mismatch Model for Generalized Audio Deepfake Detection" describes a motivation and systematic approach to disentangling different components of speaking characteristics, in order to perform audio deepfake detection. This paper demonstrates a working 2-stage training pipeline, numerous ablations, and metrics over several English language focused datasets. In addition, qualitative demonstrations of the space of features learned and used by the model form a study which backs the philosophical approach of the paper with regard to disentangling stylistic characteristics of real speakers, in order to have a generalized defense against audio deepfakes.

Strengths

The overall approach taken here is useful, the problem is relevant, and the overall experiments cover a portion of the necessary ground to match the claims. Qualitative studies both in the main body of the paper, and the appendix are generally interesting and it is worth considering if these experiments can be fit into the unified paper "flow" for this publication, especially given final conclusion writing. Overall ablation study and description of methods are well done, and form a high quality "core" to the paper. Architecture figures and ablation tables are both outstanding in terms of clearly communicating approach, and variations in the results. Use of open source toolkits, and sharing of hyperparameters should help reproducibility.

Weaknesses

The portion of the title "... Generalized Audio Deepfake Detection" claims a robust and general method for "audio deepfake detection". However, there are a broad range of prosodic styles from people with impaired speech, new learners of languages, children developing their ability to communicate and so on. For a paper about generalized deepfake detection, where such detection keys particularly on the prosody of at least one example, I would like to see a larger study and example of voices in-the-wild, as opposed to the current examples which seem to be reasonably fluent speakers and performers, who may have "trained" speech patterns by and large. This is despite the use of the "in-the-wild" dataset, which doesn't seem truly wild in terms of robustness testing. Defense approaches should have some example study and discussion of False Positives beyond pure metrics (though the metrics discussions here are well done) - particularly when out-of-domain speaking patterns may have heavy overlap with "deepfake" data and the features used for classification, demonstrated as part of the paper. The examples in Figure 4. along with accompanied writing are a start down this road but not sufficient. Mozilla Common Voice has a decent amount of this type of truly-in-the-wild speech for some qualitative study, and there are existing papers which use the same dataset for few-and-zero-shot TTS and voice conversion. The "in-the-wild" dataset here seems to largely focus on imitative TTS and voice conversion, and their "real" counterparts, which would generally point to celebrities, politicians and other public figures who (very likely) do not have the types of speech patterns mentioned previously. Though dysarthric speech is mentioned briefly in the limitations section, the issues which crop up from study of dysarthric speech are also found to some extent in many "typical" speakers as well, in more subtle ways so directly addressing this with some examples would strengthen the core claim of the paper in regards to "generalized detection". MLAAD is multilingual, but some details of the dataset construction lead to limitations in its testing (outside the scope of this paper, beyond the continued critique that broader datasets and synthetic generation methods are needed to test generalization). However, here only the EN subset appears to be used - which again reduces the claims from the title since it means the bulk of testing is on English locales. This is not a problem in terms of the experiments, but the writing and claims of the paper should be limited around this fact. Additionally, these are speech deepfakes not the broader category of "audio" per-se, so maybe some further adjustment is warranted, though other papers in this subarea tend to use "audio deepfake" to describe speech deepfakes. As it stands, the examples shown do not convince me that the "attacks" used here are sufficiently high-quality to claim a generalized defense, though the developed method seems to perform well on the datasets used, and the overall scientific study (though limited) is well done.

Questions

What are the systems tested in Table 1? Either by name, or citation? What is the source of the speakers? If these have PII, a description of the speakers broad categorizations is sufficient. If pulled from an existing dataset, speaker ids would be good. As it stands this table is largely uninformative, without any material information beyond a general design motivation for follow-on work (since CCA shows some behavior differences between methods). Given the importance of both sample rate and noise in audio, it would be very useful to test this approach under those forms of degradation - e.g. does the method scale down to data of narrow bandwidth, at low samplerate or under the presence of additive noise / background sound (such as music, crowd noise, applause, and so on). The prosodic example may hold under reasonable conditions, but how many detections are relying on prosodic features versus simpler acoustic artifacts? Figure 2. hints at this to some extent, but some explicit description and study would be useful. Generally the data examples shown are extremely noisy, and the synthesis methods are not particularly high quality. Testing on both clean audio, and higher quality synthesis, as well as under controlled degradations could raise my score. After all it is plausible an attacker may use telephony as a transmission channel - especially if the degradations imposed by the channel give the attacker a further advantage. As a general direction - it may be useful to directly answer some of the questions posed by the titles of the citations in this paper e.g. "Does audio deepfake detection generalize?" - the claim here being "yes", but demonstrations being limited to existing datasets rather than further tests with recently developed technologies / APIs and so on. "Does deepfake detection rely on artifacts?" - the claim here is also (somewhat) "yes", which hurts the counterclaim of being generalized to some degree, unless these artifacts are general across a broad swath of methods, which would be a surprising finding given existing demonstrations. The primary concern in order to raise my score would be a more proper scoping of the generalization claims, and the domain claims around this method given limitations of the testing datasets. The conclusion also discusses a fair bit about qualitative analyses which are largely relegated to the appendix, so there is further mismatch between the chosen title and the final claim. Larger and more diverse datasets (multi-lingual being one option, more unusual speaker styles would be another), or more particularly use of a variety of recent, high performing methods would raise my score if the writing is mostly unchanged. Some of these methods may only be available by API, which is unfortunate but perhaps necessary - additionally TortoiseTTS and spinoffs should have specific, stronger synthesis exemplars than those demonstrated, especially under the assumption an attacker may be doing manual selection given a corpus of intermediate generations to choose the best final result.

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors have addressed some limitations of their work, however this review is partly hinging on the gap between claims, and the effective limitations and demonstrated results. More writing on the limitations, and particularly potential harms of deploying unbalanced "defense" methods in terms of accessibility would be beneficial.

Reviewer DiH36/10 · confidence 3/52024-07-25

Summary

This paper proposes a new method for audio deepfake detection by first employing self-supervised pre-training on real samples only and then used to do real/fake classification. The proposed method achieves SOTA performance in both within-domain and cross-domain scenarios.

Strengths

1. The proposed technique is sound and reasonable. Learning the correlation between style and linguistics for detecting deepfakes makes sense to me. 2. The comprehensive ablation studies in Table 2 further verify the effectiveness of the proposed method. 3. The analysis (Interpretation of model decisions) and visualization (Mel-spectrograms) are reasonable.

Weaknesses

1. The idea of capturing the mismatch between style and linguistics is promising, but it's unclear how this mismatch correlates with deepfake samples. More intuitive illustrations and examples are needed to better validate this motivation. 2. In Figure 2 (training framework for ADD), the source of supervision for style and linguistics is not apparent. How do you ensure that each encoder learns the corresponding features? Additionally, how do you achieve perfect disentanglement between the style and linguistics encoders? 3. It appears that the latest dataset used is ASVspoof2021, which is quite old. Why not incorporate more recent and advanced deepfake datasets for evaluation?

Questions

See the Weaknesses part.

Rating

6

Confidence

3

Soundness

3

Presentation

4

Contribution

3

Limitations

I did not see any obvious limitations for this work. It is a fairly good paper but not very impressive to me.

Reviewer 4aqZ2024-08-12

Reply to Authors

The additional analysis, study under various degradations, clarifications around data, and particularly addition of ASVspoof5 data cover a reasonable portion of my prior concerns. Combined with writing which more accurately scopes the claims made, this has strengthened the impact of the work and should be of interest to a larger audience, now and into the future.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC