The SJTU Robust Anti-Spoofing System for the ASVspoof 2019 Challenge

Yexin Yang, Hongji Wang, Heinrich Dinkel, Zhengyang Chen, Shuai Wang, Yanmin Qian, Kai YuInterspeech 2019 · OpenReview

Abstract

The robustness of an anti-spoofing system is increasingly critical for developing reliable speaker verification systems. This paper presents the SJTU system submitted to the ASVspoof 2019 challenge, addressing both logical and physical access spoofing. By combining Log-CQT features with advanced deep learning architectures—specifically Gated Linear Units (GLU) and Variational Autoencoders (VAE)—the proposed system achieves consistent performance improvements. Our primary submissions reached 5th and 8th positions for logical and physical access, respectively, with a contrastive submission for the physical access task performing at a level comparable to the 3rd position.

Introduction

Automatic Speaker Verification (ASV) systems are vulnerable to spoofing attacks, which generally fall into two categories: Logical Access (LA), involving synthetic speech from Text-to-Speech (TTS) or Voice Conversion (VC) systems, and Physical Access (PA), involving replay attacks. The core challenge in anti-spoofing is generalization—the ability to detect new, unseen attack types. The SJTU team aimed to build a robust system capable of identifying fundamental spoofing signatures across diverse conditions.

Methodology

The SJTU system utilizes a multi-pronged approach involving specialized feature engineering and novel neural architectures.

Feature Engineering

Instead of standard Fourier transforms, the team employed the Log-Constant-Q Transform (Log-CQT) for its superior frequency resolution. They also incorporated phase features to capture timing information and VAE log-CQT features, where a Variational Autoencoder learns the distribution of genuine speech to better identify anomalies.

Architectural Innovations

A key component of the system is the Context Gate CNN (CGCNN), which utilizes Gated Linear Units (GLU). Unlike standard activation functions, GLUs act as a dynamic gate, allowing the network to focus on the most relevant parts of the audio signal for spoofing detection.

The Context Gate mechanism using Gated Linear Units (GLU) to dynamically weight audio features for better spoofing detection.

Furthermore, the system integrates traditional speaker modeling techniques by concatenating 400-dimensional i-vectors into ResNet-based models. This helps the system distinguish between speaker-specific characteristics and artifacts introduced by spoofing.

System architecture illustrating the fusion of deep embeddings from residual layers with i-vectors for final classification.

Training Strategies

For the Physical Access task, the team used Multi-Task Learning, training the model to identify specific replay configurations (e.g., recording distance) alongside the primary spoofing detection task. They also employed heavy data augmentation using simulated room impulse responses (RIRs) from the OpenSLR26 database to improve robustness against diverse acoustic environments.

Results and Discussion

The SJTU system demonstrated high competitiveness in the ASVspoof 2019 challenge:

  • Logical Access (LA): The fusion system achieved an Equal Error Rate (EER) of 0.90% on the development set and ranked 5th overall.
  • Physical Access (PA): While the primary system ranked 8th, the contrastive submission—leveraging CGCNN fusion and extensive RIR augmentation—achieved a development EER of 0.16%, matching the performance of the 3rd-place system in the challenge.

These results underscore the importance of phase information and learned gating mechanisms in capturing the nuances of spoofed audio. The success of the contrastive PA submission highlights that simulating diverse acoustic environments is essential for generalizing to unseen replay attacks.

Conclusion

The SJTU anti-spoofing system provides a robust framework for securing ASV systems. While the current multi-model fusion approach is highly accurate, future work will explore reducing computational complexity for real-time applications and further refining the boundaries of genuine speech using uncertainty-driven models.

Summary updated

Abstract

The robustness of an anti-spoofing system is progressively more important in order to develop a reliable speaker verification system. Previous challenges and datasets mainly focus on a specific type of spoofing attacks. The ASVspoof 2019 edition is the first challenge to address two major spoofing types — logical and physical access. This paper presents the SJTU’s submitted anti-spoofing system to the ASVspoof 2019 challenge. Log-CQT features are developed in conjunction with multi-layer convolutional neural networks for robust performance across both subtasks. CNNs with gradient linear units (GLU) activations are utilized for spoofing detection. The proposed system shows consistent performance improvement over all types of spoofing attacks. Our primary submissions achieve the 5th and 8th positions for the logical and physical access respectively. Moreover, our contrastive submission to the PA task exhibits better generalization compared to our primary submission, and achieves a comparable performance to the 3rd position of the challenge.
BibTeX
@inproceedings{yang2019sjtu,
  author = {Yexin Yang and Hongji Wang and Heinrich Dinkel and Zhengyang Chen and Shuai Wang and Yanmin Qian and Kai Yu},
  title = {{The SJTU Robust Anti-Spoofing System for the ASVspoof 2019 Challenge}},
  booktitle = {Interspeech},
  year = {2019},
  url = {https://openreview.net/forum?id=OtTpQWg5xo},
}
Figure