A cross-vocoder study of speaker independent synthetic speech detection using phase information

Jon Sanchez, Ibon Saratxaga, Inma Hernaez, Eva Navas, Daniel ErroInterspeech 2014

TL;DR

This paper investigates a speaker-independent method for detecting synthetic speech, focusing on how well it generalizes across different speech synthesis techniques (vocoders). They found that using phase information (RPS) significantly improves detection robustness compared to standard MFCCs, especially when the specific synthesis method is unknown.

What Problem Does This Paper Address?

The increasing sophistication of speech synthesis and voice conversion technologies poses a significant threat to speaker verification systems. It's becoming easier to create synthetic voices that can fool these systems. While synthetic speech detection (SSD) methods exist, many are either speaker-dependent (requiring specific models for each user) or dependent on the specific synthesis technique used to create the spoof. This paper tackles the challenge of creating a universal synthetic speech detector that is both speaker-independent and robust to various synthesis methods, without prior knowledge of the spoofing technique.

Methods

The core of the proposed system is a statistical classifier that models both human and synthetic speech. The key innovation lies in the parameterization of the speech signals:

  • MFCCs (Mel-Frequency Cepstral Coefficients): Used as a standard baseline for comparison. These capture spectral envelope information.
  • RPS (Relative Phase Shifts): This novel parameterization focuses on the phase information of the speech signal. The authors hypothesize that many synthesis techniques disregard phase, leading to detectable differences between natural and synthetic speech. RPS captures harmonic phase information, and for modeling, it's further processed using DCT-mel transformation.

The system trains Gaussian Mixture Models (GMMs) on these parameters for both human and synthetic speech. A decision threshold, set at the Equal Error Rate (EER), is used to classify incoming speech. The study explores training models with varying numbers of speakers to assess speaker independence and then evaluates the system's performance across different vocoders.

Datasets

The WSJ database was used for this study. Specifically, the SI-284 set, known for its large number of speakers and high signal-to-noise ratio, was utilized. Sentences were copy-synthesized using three different vocoders (MLSA, STRAIGHT, and AHOCODER) to create synthetic speech. The dataset was then partitioned into training and testing sets with varying numbers of speakers (30-SPK, 50-SPK, 140-SPK) to evaluate speaker independence.

Key Insights and Ideas

  • Phase Matters: The most striking finding is that phase information (RPS) is a much more robust feature for synthetic speech detection than spectral information (MFCCs), especially when dealing with diverse synthesis methods.
  • Vocoder Dependency is Real: The system's performance drops significantly when a model trained on one vocoder is used to detect speech synthesized by a different vocoder. This highlights a major challenge for universal SSD.
  • RPS Generalizes Better: While both parameterizations suffer from vocoder dependency, RPS shows a greater ability to generalize and perform better in cross-vocoder scenarios.
  • Multi-Vocoder Models Offer Limited Improvement: Training models with data from multiple vocoders doesn't consistently improve detection rates for known spoofing vocoders. However, for unknown spoofing vocoders, multi-vocoder RPS models show some benefit.

Key Results

  • Speaker Independence Achieved: The speaker-independent classifiers performed well across all tested vocoders, with RPS consistently achieving low error rates (well below 1% for matched vocoders).
  • Cross-Vocoder Performance: When testing against a different vocoder than the one used for training, EERs increased dramatically. For example, with RPS, detecting AHOCODER speech using a STRAIGHT model yielded an EER of 2.80%, while MFCCs resulted in 25.59%.
  • RPS vs. MFCC: RPS parameterization consistently outperformed MFCCs in cross-vocoder tests and when dealing with the combined "all vocoders" test set (T). For instance, in Table 3 (cross-vocoder), RPS generally showed much lower EERs than MFCCs.
  • Multi-Vocoder Models: While not a silver bullet, aggregating vocoders into RPS models led to slight improvements in cross-vocoder detection, especially for the "all vocoders" test set, suggesting a path towards more universal detection.

Limitations and Future Work

  • Vocoder Dependency: The paper clearly demonstrates that current SSD systems are highly dependent on the specific vocoder used. This is a significant limitation for real-world applications where the spoofing technique is unknown.
  • Future Work: The authors suggest that a practical, universal SSD would require more extensive research. They propose that a system incorporating information from all available vocoders and using RPS parameterization could offer some success against unknown spoofing attacks. Further investigation into robustness against newer, non-minimum phase vocoders is also mentioned.

Summary updated

BibTeX
@inproceedings{sanchez2014cross,
  author = {Jon Sanchez and Ibon Saratxaga and Inma Hernaez and Eva Navas and Daniel Erro},
  title = {{A cross-vocoder study of speaker independent synthetic speech detection using phase information}},
  booktitle = {Interspeech},
  year = {2014},
}
Figure