- The paper introduces a comprehensive multilingual dataset for singfake detection, integrating real and synthetic singing audios from diverse languages and music genres.
- It employs rigorous preprocessing and diverse generative models (GANs, diffusion, VC, TTM) to produce realistic deepfakes while enabling alternative fake detection and forensic source tracing.
- Experimental evaluations reveal that LFCC+ResNet excels in multilingual scenarios, highlighting the limitations of speech-centric models in capturing musical artifacts.
Authoritative Summary of "SingFox: A Multi-Lingual Singfake Detection Corpus" (2606.18985)
Motivation and Context
The emergence of highly realistic AI-generated audio content, especially in singing domains (singfakes), has prompted significant concerns regarding authenticity verification, copyright infringement, and security vulnerabilities. Unlike speech deepfakes, singfake detection introduces additional challenges due to musicality, pitch modulation, and vocal style artifacts that are absent in speech. Prevailing datasets such as ASVSpoof, SONICS, and WildSVDD either lack multilingualism, diverse generative paradigms, or real-life scenario modeling, which restricts the fidelity and robustness of singfake detectors in practical deployments. This work presents SingFox, a comprehensive dataset that bridges these deficiencies and establishes a rigorous benchmark for both singfake detection and source tracing.
Dataset Design and Architecture
SingFox comprises 113,802 audio clips across 20 languages and includes six tracks (T1-T6) reflecting real-world variations in linguistic diversity, music genre, and alternative fake constructs. The corpus integrates five distinct music types and covers both international and Indic languages, addressing the pronounced underrepresentation of the latter in previous research. Rigorous preprocessing (peak and RMS normalization, 4-second chunking, and careful bias avoidance) standardizes audio quality, ensuring the exclusion of shortcut or confounding cues.
The dataset encompasses both real and synthetic singing audios, with synthetic samples generated using a diverse set of state-of-the-art generative models: HiFi-GAN, BigVGAN, UnivNet (GANs); DiffSinger, DiffRhythm (diffusion models); So-VITS-SVC, RVC (voice conversion); and MusicGen (text-to-music models). Cross-model diversity is integral—GAN, diffusion, VC, and TTM approaches are all represented, mitigating bias and enhancing robustness in evaluation. Text symmetry is ensured by applying OpenAI Whisper for transcript alignment, countering dataset-induced shortcut biases and promoting artifact-level learning.
Tracks are tailored to scenario-specific testing:
- T1/T2: Global/Indic language coverage.
- T3: Genre-focused music deepfakes.
- T4: Multilingual mix for maximal generalization.
- T5: Alternative fakes, i.e., fake vocals with real music.
- T6: Source tracing for forensic attribution.
1150 distinct singers contribute to authentic diversity, with anonymized speaker metadata due to copyright and identity concerns.
Experimental Evaluation
Baseline models employ spectral features (LFCC, MFCC, GFCC) with ResNet, CNN, BiLSTM, and BiGRU classifiers. LFCC+ResNet consistently outperforms alternatives (e.g., GFCC), particularly evident in source tracing tasks where acoustic explainability is required. Model robustness is demonstrated across multilingual tracks: as language diversity increases, so does accuracy and cross-domain generalization.
In alternative fake detection (T5), accuracy dropped to 45.13% (LFCC+ResNet), evidencing the increased challenge when distinguishing fake vocals amidst real background music. For standard source tracing on T6, LFCC features achieved 89.06% accuracy, outperforming MFCC (88.71%) and GFCC (70.34%).
Cross-Dataset Validation
SingFox was used exclusively as a test set in cross-dataset scenarios, highlighting limitations of models trained on mismatched datasets such as CtrSVDD and WildSVDD, which achieved 46.06% and 54.17% accuracy respectively on SingFox. FMC-trained models performed best (77.84%), emphasizing the importance of diversified training data for real-world deployment.
Model-Wise and Perceptual Analysis
Objective and subjective metrics were employed for singfake sample quality evaluation:
- Objective: RVC achieved highest PESQ (4.64), STOI (1.0), and PCC (1.0), indicating high fidelity and intelligibility. BigVGAN produced the lowest model-wise detection accuracy (1.02%), signifying its generated samples were most difficult to distinguish from real recordings.
- Subjective: Mean Opinion Score (MOS) for generated deepfakes averaged 3.468, approaching the ground-truth MOS of 4.028. Notably, DiffRhythm and BigVGAN were perceptually favored over RVC, illustrating a divergence between acoustic metrics and human perceptual evaluation.
Source Tracing and Explainability
Track T6 introduces a protocol for forensic source tracing—closed-set and open-set conditions are simulated using enrollment and evaluation stages across multiple synthesis models. This enables robust attribution, rejection of unseen sources, and explainability analysis. LFCC+ResNet trained on SingFox performed reliably, with clear separation between known and unknown generative paradigms, advancing research on deepfake forensics and model accountability.
Practical and Theoretical Implications
SingFox’s multi-lingual, multi-model composition addresses critical gaps in existing resources, facilitating real-world robustness in singfake detection. It enables the benchmarking of emerging paradigms such as alternative fakes and rigorous forensic tracing under realistic, multilingual deployment environments. The dataset supports reproducibility (codes available) and encourages future investigation into specialized model architectures (e.g., cross-attention, graph-based approaches such as SingGraph).
Practically, SingFox can serve as a testbed for model explainability, forensic attribution, and robust evaluation of detectors deployed in global media, streaming, or security-sensitive scenarios. Theoretically, its diversity and protocol design inform future dataset construction, model generalization studies, and highlight the need for perceptually-informed evaluation metrics.
Conclusion
SingFox (2606.18985) represents a significant advancement in the domain of singing deepfake detection. Its methodological rigor, linguistic and generative diversity, and inclusion of novel tracks (alternative fakes and source tracing) collectively establish a challenging benchmark for evaluating detection and attribution models. The dataset supports robust cross-model and cross-lingual generalization and reveals the limitations of speech-centric training in the context of musical artifacts. Future research will benefit from SingFox’s structure, particularly in developing explainable, globally deployable singfake detection and forensic tracing frameworks.