Language Orthogonalization for Zero-Shot Cross-Lingual Audio Deepfake Detection

1. INTRODUCTION

• Detectors must transfer to languages with no bonafide or spoofed speech available for adaptation or calibration, while labeled anti-spoofing resources remain limited relative to multilingual synthesis.The introduction notes that synthesis spans over a thousand languages, while standard benchmarks remain largely English-centric and multilingual resources cover only a fraction of that breadth.Source: §1 · p. 1
• S3Ms encode language identity, so detectors can confuse language-conditioned acoustics with spoof cues and degrade especially when transferring across linguistic boundaries.The passage states that both bonafide and spoofed speech retain linguistic structure.Source: §1 · p. 1
• The method fits a ridge map from continuous LID embeddings to S3M representations using source-language bonafide speech, then subtracts the mapped language component without target data.The approach is motivated by prior findings that S3M phonetic variations occupy approximately linear subspaces.Source: §1 · p. 1
• Across six South and Southeast Asian languages and six S3M backbones, cross-lingual EER increases with LID distance.The authors use continuous LID space as a representation of phonological, geographic, and genealogical language relationships [13].Source: §1 · p. 1
• Language orthogonalization consistently reduces EER across six S3Ms and all Leave-N-Out settings, with the largest gains on linguistically distant transfers.The paper also reports gains under constrained source setups.Source: §1 · p. 1

2.1. Feature Representations

• Six frozen S3Ms produce 2048-dimensional utterance embeddings through mean-plus-standard-deviation pooling across hidden layers, depthwise layer averaging, and L2 normalization.This fixed pipeline avoids backbone-specific layer selection.Source: §2.1 · pp. 1–2
• The figure depicts removing language-predictable variation and retaining a residual space that aligns bonafide distributions and improves transfer across linguistic groups.Its schematic begins with similar languages placed nearby in S3M representation space.Source: Figure 1 · p. 2
• A VoxLingua107-pretrained ECAPA-TDNN extracts continuous 256-dimensional LID embeddings that place utterances in a shared language space, enabling language-distance analysis and transformation of target languages unseen during training.Source: §2.1 · p. 2

2.2. Language Orthogonalization

• Language orthogonalization fits its mapping exclusively on source-language bonafide speech, requiring neither class from target languages and excluding spoof artifacts from estimation.The transformation subtracts the language-predictable component from an utterance representation.Source: §2.2 · p. 2

3.1. Dataset

• SEA-Spoof contains bonafide and spoofed speech from six languages, grouped into South Asia, Mainland Southeast Asia, and Maritime Southeast Asia.The groups are Hindi and Tamil; Thai and Vietnamese; and Indonesian and Malay, respectively.Source: §3.1 · p. 2

3.2. Cross-Lingual Evaluation

• Leave-N-Out evaluation holds out every combination of N=1–5 languages and trains on the remaining 6−N, with held-out languages entirely unseen and no language-specific adaptation.Raw and orthogonalized representations are compared using EER averaged across language combinations at each N; evaluation uses class-balanced logistic regression.Source: §3.2 · pp. 2–3
• Across six S3Ms, greater LID distance is consistently associated with higher EER, while intra-group language pairs are closer than MSEA–MRT and SEA–SA pairs.The figure caption identifies r as the Pearson correlation between distance and EER.Source: Figure 3 · p. 3
• Table 1 covers cross-lingual EER (%) under Leave-N-Out evaluation, with the remaining 6−N languages used for training and parentheses showing relative EER reduction from Raw.Its caption reports that orthogonalization improves every backbone and N.Source: Table 1 · p. 3

4.1. Language Distance and Cross-Lingual Difficulty

• LID distance reflects linguistic-area structure and correlates with cross-lingual EER across all six S3M backbones.Transfer difficulty aligns with linguistic distance from the training language.Source: §4.1 · p. 3

4.2. Overall Leave-N-Out Performance

• Orthogonalization reduces EER by 9–17% on average across all backbones and Leave-N-Out setups.Table 1 covers N = 1 to 5 unseen languages.Source: §4.2 · p. 3

4.3. Effect of Linguistic-Area Composition

• Single-language transfers gain most across distant areas, with up to 22% relative EER reduction on XLS-R; two-language training cuts EER up to 41% and 45% in constrained compositions.For two-language training, the reported XLS-R reductions apply to intra-group and MRT × MSEA compositions, respectively.Source: §4.3 · pp. 3–4

4.4. Representation Geometry after Orthogonalization

• Orthogonalization aligns bonafide representations across source and unseen languages, reducing bonafide language-centroid distances across all six backbones.Reported distance reductions range from 44% to 70%.Source: Figure 4 · p. 4, §4.4 · p. 4

• The proposed language orthogonalization strategy fits an LID-to-S3M mapping on bonafide speech.Source: §5 · p. 4
• Across multiple S3M backbones and unseen-language setups, the approach consistently lowers EER, with larger gains on linguistically distant transfers.Source: §5 · p. 4
• Continuous distance in LID space reflects linguistic-area structure and characterizes transfer difficulty.Source: §5 · p. 4