EmoAnti: Emotion-Guided Audio Anti-Deepfake

• EmoAnti is an audio anti-deepfake system that integrates high-level emotional cues with conventional acoustic features to detect spoofed speech.
• It employs a fine-tuned wav2vec2 backbone, a convolutional residual feature extractor, and attentive fusion to refine multi-layer transformer outputs.
• Performance evaluations show state-of-the-art results on ASVspoof2019 LA and ASVspoof2021 LA benchmarks, while highlighting limitations in diverse, in-the-wild conditions.

Searching arXiv for the specified EmoAnti paper and closely related work to ground the article in current literature. EmoAnti is an audio anti-deepfake detection system that uses refined emotion-guided representations to detect spoofed speech by incorporating high-level emotional cues alongside speech representations. It is introduced in "Emoanti: audio anti-deepfake with refined emotion-guided representations" (Li et al., 13 Sep 2025), where the central premise is that most detection systems primarily rely on low-level acoustic features or pretrained speech representations and frequently neglect emotional information that may provide complementary anti-deepfake evidence. In this formulation, emotional structure is treated not as an auxiliary annotation target but as a representational prior for robust spoof detection.

1. Problem setting and conceptual premise

EmoAnti is positioned within audio spoofing countermeasures for settings in which synthetic or manipulated speech has become sufficiently sophisticated that robust detection is difficult. The method is explicitly motivated by the hypothesis that current deepfake techniques struggle to convincingly model or mimic genuine human emotional expression in speech (Li et al., 13 Sep 2025). On that basis, the system targets subtle, high-level discrepancies in emotional rendering rather than depending only on conventional acoustic front ends.

The approach therefore departs from anti-deepfake systems that use pretrained speech models only as generic feature extractors. Its stated objective is to derive emotion-guided representations and then refine them so that paralinguistic characteristics become discriminative for bonafide-versus-spoof classification. This design choice also carries an interpretability claim: because the backbone is adapted through speech emotion recognition, the learned embedding space partially aligns with human-perceivable emotion categories.

2. Emotion-specialized backbone

The first principal module is a pretrained Wav2Vec2 model fine-tuned for emotion recognition. EmoAnti uses wav2vec2-large-robust, described as a model with a 7-layer Conv1d frontend, followed by a 12-layer transformer encoder with 16 heads per layer, and an optional quantization module mainly used during self-supervised pretraining (Li et al., 13 Sep 2025). Rather than using this backbone unchanged, the system fine-tunes it for Speech Emotion Recognition (SER).

In the fine-tuning stage, the final transformer output $H^{12}$ is fed to an emotion classifier that predicts one of four emotion labels: "angry," "happy"/"excited," "sad," and "neutral." The emotion fine-tuning dataset is IEMOCAP, and macro F1 and accuracy are tracked for emotion recognition. The methodological significance of this stage is that it endows the representations with emotion-awareness before the anti-deepfake head is trained.

A key architectural rationale is that intermediate transformer layers often retain richer speaker and paralinguistic cues than the final layer, whose optimization is typically more closely tied to speech recognition. EmoAnti therefore does not restrict itself to a single terminal representation. Instead, it uses outputs from all transformer layers, denoted $H=\{H^l\}_{l=1}^{12}$, as the input substrate for subsequent refinement.

3. CRFE, attentive fusion, and classification

The second principal module is the Convolutional Residual Feature Extractor (CRFE). Each CRFE block consists of two Conv1d layers with kernel size $3$ and padding $1$, followed by batch normalization and ReLU, with a residual connection. If input and output dimensions differ, the residual branch uses a $1 \times 1$ Conv1d; otherwise it is the identity (Li et al., 13 Sep 2025). Four such convolutional residual blocks are stacked to extract and refine emotional characteristics from multi-layer transformer outputs.

The CRFE is intended to capture temporal local structure while preserving cross-layer paralinguistic detail. In the system description, this refinement stage is justified as a way to surface inconsistencies in emotion rendering that may be hard for deepfake generators to reproduce. This suggests that EmoAnti treats emotion not as a single global utterance label but as a distributed temporal pattern whose local regularities may differ between bonafide and spoofed speech.

After CRFE processing, EmoAnti applies a Feature Fusion Module based on temporal attention. For each CRFE output $f_i$, the temporal attention sub-network is defined as

$$e_{i,t} = \text{Linear}_2\bigl(\text{ReLU}(\text{Linear}_1(f_{i,t}))\bigr)$$

$$\alpha_{i,t} = \frac{\exp(e_{i,t})}{\sum_{t'=1}^{T}\exp(e_{i,t'})}$$

$$\hat{f}_i = \sum_{t=1}^T \alpha_{i,t} \odot f_{i,t}$$

and the fused representation is

$$F_{\mathrm{fused}} = [\hat{f}_1,\, \hat{f}_2,\, \dots,\, \hat{f}_i].$$

The classification head then passes the fused feature vector through two linear layers with ReLU and dropout to produce a binary bonafide-versus-spoof prediction. Architecturally, the overall pipeline thus combines an emotion-specialized SSL backbone, convolutional residual refinement over transformer-layer outputs, and attentive aggregation before final decision making.

4. Training protocol and evaluation setting

EmoAnti is evaluated on standard anti-spoofing benchmarks: ASVspoof2019 LA, ASVspoof2021 LA, and ASVspoof2021 DF (Li et al., 13 Sep 2025). In the supplied description, ASVspoof2019 LA is characterized as the classic logical-access benchmark; ASVspoof2021 LA is described as more difficult due to new speakers and microphone, channel, and transmission artifacts; and ASVspoof2021 DF focuses on deepfakes with compression artifacts and in-the-wild conditions.

Training uses Adam or AdamW with low learning rates, cosine annealing, early stopping based on validation loss or F1, and gradient accumulation. For anti-deepfake detection the loss function is Cross-Entropy Loss. The principal evaluation metrics are Equal Error Rate (EER, %) and minimum normalized tandem detection cost function (min t-DCF). EER measures the operating point at which false positive and false negative rates intersect, while min t-DCF evaluates the effect of the countermeasure in a full speaker verification pipeline; lower values are better in both cases.

The benchmark results reported for EmoAnti are summarized below.

Benchmark	EmoAnti result	Comparison context
ASVspoof2019 LA	EER 0.44%, min t-DCF 0.0139	Lower than LFCC-GMM, RawNet2, and W2V2 (fixed)+LCNN+BLSTM
ASVspoof2021 LA	EER 4.62%, min t-DCF 0.2920	Lower than LFCC-GMM, ECAPA-TDNN, W2V2 (fixed)+LCNN+BLSTM, and W2V2 (finetuned)+LCNN+BLSTM
ASVspoof2021 DF	EER 13.72%	Reported as strong generalization, though not the lowest EER among all compared systems

These results are framed in the paper as state-of-the-art on ASVspoof2019 LA and ASVspoof2021 LA, together with strong generalization on ASVspoof2021 DF.

5. Empirical findings, ablations, and limitations

On ASVspoof2019 LA, EmoAnti reports an EER of 0.44% and a min t-DCF of 0.0139, compared with 8.09 / 0.212 for LFCC-GMM, 5.64 / 0.130 for RawNet2, and 1.47 / 0.105 for W2V2 (fixed)+LCNN+BLSTM (Li et al., 13 Sep 2025). On ASVspoof2021 LA, it reports 4.62% EER and 0.2920 min t-DCF, outperforming the listed baselines. These are the strongest results presented in the comparison tables and are the basis for the paper’s state-of-the-art claim on those two benchmarks.

On ASVspoof2021 DF, the system achieves 13.72% EER. The paper describes this as strong generalization, but also notes an important qualification: in wild settings, some W2V2-based methods not leveraging emotion, but using different fusion or classifier heads, achieve lower EERs. This is a central limitation rather than a peripheral caveat. It indicates that emotion guidance improves performance substantially in the reported LA settings, but does not by itself guarantee superiority under all degradations or distribution shifts.

The ablation studies attribute performance gains to both major design decisions. Removing emotion-aware fine-tuning worsens detection performance, and removing the CRFE also worsens performance. The summary states that both emotional embedding enrichment and feature refinement are crucial for strong anti-deepfake generalization. It also reports a more nuanced outcome on ASVspoof2021 DF: removing only emotion fine-tuning hurts most, while removing both emotion fine-tuning and CRFE can yield unexpected gains, which the authors suggest may reflect overfitting. This makes the paper’s overall claim more specific than a generic “emotion always helps” argument; the benefit depends on how emotional supervision interacts with downstream refinement and domain conditions.

6. Research significance and terminological context

Within anti-spoofing research, EmoAnti is significant because it formalizes emotional information as a first-class source of discriminative evidence rather than as a downstream descriptive label. The method integrates fine-tuned emotion-aware Wav2Vec2 embeddings, a convolutional residual extractor over multi-layer transformer states, and attentive fusion for classification (Li et al., 13 Sep 2025). Its contribution is therefore architectural as well as conceptual: the system operationalizes the claim that high-level paralinguistic cues can complement conventional speech representations in spoof detection.

The name EmoAnti also has a broader terminological context. In a distinct line of work on online messaging and Affect Control Theory, the term appears in a summary as "Emotional Anticipation/Anti-deflection", referring to monitoring deflection trends to identify and minimize emotional discord in chatbot interactions (Mostafavi et al., 2021). That usage concerns emotional transition modeling in text-and-emoji dialogue rather than audio anti-deepfake detection. This suggests that the label is not yet terminologically stable across subfields.

In the narrower and now better-defined sense established by (Li et al., 13 Sep 2025), however, EmoAnti denotes an audio anti-deepfake framework in which emotion recognition is repurposed as an inductive bias for spoof detection. The paper’s results support the view that refined emotion-guided representations can improve benchmark performance and generalization, while its DF results also make clear that emotional cues remain one component of a broader anti-spoofing design space rather than a complete solution on their own.