Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

LJ-Spoof: A Generatively Varied Corpus for Audio Anti-Spoofing and Synthesis Source Tracing

Surya Subramani, Hashim Ali, Hafiz Malik

TL;DR

The paper tackles robustness of speaker-specific anti-spoofing against diverse generative synthesis by introducing LJ-Spoof, a single-speaker, generatively diverse corpus spanning 30 TTS families, 500+ variant subsets, and over 3 million utterances. It presents a reproducible generation workflow, comprehensive metadata, and a strategic split protocol to enable fine-grained analysis of synthesis artifacts and source tracing. By varying training regimes, inputs, generative parameters, and neural post-processing, LJ-Spoof provides a controlled benchmark for evaluating spoof detection under realistic adversarial conditions. The authors position LJ-Spoof as a practical resource for training and benchmarking anti-spoofing systems, with plans to extend to in-the-wild data and multiple speakers, enhancing generalization and traceability in synthesis-based threats.

Abstract

Speaker-specific anti-spoofing and synthesis-source tracing are central challenges in audio anti-spoofing. Progress has been hampered by the lack of datasets that systematically vary model architectures, synthesis pipelines, and generative parameters. To address this gap, we introduce LJ-Spoof, a speaker-specific, generatively diverse corpus that systematically varies prosody, vocoders, generative hyperparameters, bona fide prompt sources, training regimes, and neural post-processing. The corpus spans one speakers-including studio-quality recordings-30 TTS families, 500 generatively variant subsets, 10 bona fide neural-processing variants, and more than 3 million utterances. This variation-dense design enables robust speaker-conditioned anti-spoofing and fine-grained synthesis-source tracing. We further position this dataset as both a practical reference training resource and a benchmark evaluation suite for anti-spoofing and source tracing.

LJ-Spoof: A Generatively Varied Corpus for Audio Anti-Spoofing and Synthesis Source Tracing

TL;DR

The paper tackles robustness of speaker-specific anti-spoofing against diverse generative synthesis by introducing LJ-Spoof, a single-speaker, generatively diverse corpus spanning 30 TTS families, 500+ variant subsets, and over 3 million utterances. It presents a reproducible generation workflow, comprehensive metadata, and a strategic split protocol to enable fine-grained analysis of synthesis artifacts and source tracing. By varying training regimes, inputs, generative parameters, and neural post-processing, LJ-Spoof provides a controlled benchmark for evaluating spoof detection under realistic adversarial conditions. The authors position LJ-Spoof as a practical resource for training and benchmarking anti-spoofing systems, with plans to extend to in-the-wild data and multiple speakers, enhancing generalization and traceability in synthesis-based threats.

Abstract

Speaker-specific anti-spoofing and synthesis-source tracing are central challenges in audio anti-spoofing. Progress has been hampered by the lack of datasets that systematically vary model architectures, synthesis pipelines, and generative parameters. To address this gap, we introduce LJ-Spoof, a speaker-specific, generatively diverse corpus that systematically varies prosody, vocoders, generative hyperparameters, bona fide prompt sources, training regimes, and neural post-processing. The corpus spans one speakers-including studio-quality recordings-30 TTS families, 500 generatively variant subsets, 10 bona fide neural-processing variants, and more than 3 million utterances. This variation-dense design enables robust speaker-conditioned anti-spoofing and fine-grained synthesis-source tracing. We further position this dataset as both a practical reference training resource and a benchmark evaluation suite for anti-spoofing and source tracing.
Paper Structure (15 sections, 1 figure, 2 tables)

This paper contains 15 sections, 1 figure, 2 tables.

Table of Contents

  1. Introduction
  2. Audio Anti-Spoofing Datasets and Shortcomings
  3. LJ-Spoof
  4. Bonafide and Default Subsets
  5. Variant Subsets
  6. Training Categories Variants
  7. Acoustic--Vocoder Variants
  8. Input Variants
  9. Generative-Parameter Variants
  10. Neural Post-Processing Variants
  11. Metadata Preparation
  12. Analysis Scope and Discussion
  13. Why Variants?
  14. Split Protocol
  15. Conclusion and Future Scope

Figures (1)

  • Figure 1: Dataset Generation Workflow for Matcha-TTS. A mel-spectrogram $S$ is synthesized from text $T$ using default generative parameters (sampling temperature $\tau=0.5$, peak rate $R=1.0$, ODE solver steps $D=10$). $S$ is then decoded by multiple vocoders (blue), including the default HiFiGAN (purple), to produce diverse waveform subsets. Finally, the default waveform undergoes post-processing via re-vocoding (yellow) and re-codec (blue).