[Paper Review] UR Channel-Robust Synthetic Speech Detection System for ASVspoof 2021

Xinhui Chen, You Zhang+2 more|arXiv (Cornell University)|Jul 26, 2021

Speech Recognition and Synthesis32 references4 citations

TL;DR

This paper presents the UR-AIR system for the ASVspoof 2021 challenge, introducing a channel-robust synthetic speech detection framework using codec and impulse response augmentation to improve generalization. It employs ECAPA-TDNN with one-class learning and model fusion, achieving EER 5.46% and min-tDCF 0.3094 on the LA track, and EER 20.33% on the DF track.

ABSTRACT

In this paper, we present UR-AIR system submission to the logical access (LA) and the speech deepfake (DF) tracks of the ASVspoof 2021 Challenge. The LA and DF tasks focus on synthetic speech detection (SSD), i.e. detecting text-to-speech and voice conversion as spoofing attacks. Different from previous ASVspoof challenges, the LA task this year presents codec and transmission channel variability, while the new task DF presents general audio compression. Built upon our previous research work on improving the robustness of the SSD systems to channel effects, we propose a channel-robust synthetic speech detection system for the challenge. To mitigate the channel variability issue, we use an acoustic simulator to apply transmission codec, compression codec, and convolutional impulse responses to augmenting the original datasets. For the neural network backbone, we propose to use Emphasized Channel Attention, Propagation and Aggregation Time Delay Neural Networks (ECAPA-TDNN) as our primary model. We also incorporate one-class learning with channel-robust training strategies to further learn a channel-invariant speech representation. Our submission achieved EER 20.33% in the DF task; EER 5.46% and min-tDCF 0.3094 in the LA task.

Motivation & Objective

• Address the generalization gap in synthetic speech detection (SSD) systems under unseen spoofing and channel conditions.
• Improve robustness to transmission and audio compression variability in anti-spoofing systems.
• Develop a channel-invariant speech representation through data augmentation and training strategies.
• Achieve state-of-the-art performance on both the LA and DF tracks of the ASVspoof 2021 challenge.
• Investigate the effectiveness of ECAPA-TDNN in anti-spoofing tasks and model fusion for performance gain.

Proposed method

• Apply an acoustic simulator to augment training data with transmission codecs (landline, cellular, VoIP), compression codecs (MP3, AAC), and device impulse responses (IRs).
• Use a multi-stage data augmentation pipeline: first apply codec degradation, then optionally convolve with device IRs to simulate real-world channel effects.
• Employ ECAPA-TDNN as the primary neural network backbone, with variants using different channel dimensions (C=512, 1024) and architectural modifications.
• Integrate one-class learning with channel-robust training to learn invariant representations across diverse channel conditions.
• Fuse multiple ECAPA-TDNN variants to improve detection performance, leveraging ensemble learning for better generalization.
• Optimize the system using min-tDCF and EER metrics on augmented development sets before final submission.

Experimental results

Research questions

• RQ1 How effective is codec and impulse response-based data augmentation in improving SSD generalization under unseen channel conditions?
• RQ2 Can ECAPA-TDNN be effectively adapted for anti-spoofing tasks, and how does its architecture scale with model size?
• RQ3 To what extent does one-class learning combined with channel-robust training enhance model invariance to channel variability?
• RQ4 How does model fusion improve performance on unseen spoofing and compression conditions in the DF and LA tracks?
• RQ5 What is the impact of over-parameterization on ECAPA-TDNN performance in anti-spoofing, particularly in terms of overfitting?

Key findings

• The UR-AIR system achieved an EER of 5.46% and min-tDCF of 0.3094 on the LA track, outperforming all baseline systems.
• On the DF track, the system achieved an EER of 20.33%, demonstrating strong performance on general audio compression conditions.
• Model fusion significantly improved detection performance, with the fused system outperforming all individual models on the augmented development set.
• Smaller ECAPA-TDNN variants (C=512) outperformed larger ones (C=1024), indicating overfitting in larger models despite higher parameter counts.
• The integration of one-class learning and channel-robust training strategies led to more consistent performance across diverse codec and channel conditions.
• The use of device impulse responses in data augmentation effectively reduced dataset bias from the ASVspoof 2019 training set.