Source-linked AI summary
ASSERT: Anti-Spoofing with Squeeze-Excitation and Residual neTworks
Cheng-I Lai, Nanxin Chen, Jesús Villalba, Najim Dehak
TL;DR
Spoofing attacks threaten automatic speaker verification across replay, text-to-speech, and voice conversion, motivating stronger anti-spoofing systems. ASSERT addresses this with a DNN pipeline combining feature engineering, squeeze-excitation and residual networks, optimization, and fusion, achieving more than 93% and 17% relative improvements over baselines on PA and LA. Its fixed-size-input design imposes a practical boundary for long utterances, although spoofed speech is mostly shorter than 10 seconds.
Problem
Automatic speaker verification is vulnerable to replay, text-to-speech, and voice-conversion spoofing attacks, which ASVspoof 2019 evaluates through PA and LA.
Method
ASSERT is a DNN-based anti-spoofing pipeline combining feature engineering, squeeze-excitation and residual network variants, network optimization, and system fusion.
Results
More than 93% and 17% relative improvements over baseline systems were achieved on the PA and LA sub-challenges, respectively.
Takeaways & Limitations
ASSERT's fusion system attained considerable improvements over baseline systems on the ASVspoof 2019 corpus and ranked among the top-performing systems.
Takeaways & Limitations
Dilated ResNet accepts only fixed-size input, which is impractical for long utterances, although spoofed speech is mostly shorter than 10 seconds.
Abstract
from arXiv · showhide
We present JHU's system submission to the ASVspoof 2019 Challenge: Anti-Spoofing with Squeeze-Excitation and Residual neTworks (ASSERT). Anti-spoofing has gathered more and more attention since the inauguration of the ASVspoof Challenges, and ASVspoof 2019 dedicates to address attacks from all three major types: text-to-speech, voice conversion, and replay. Built upon previous research work on Deep Neural Network (DNN), ASSERT is a pipeline for DNN-based approach to anti-spoofing. ASSERT has four components: feature engineering, DNN models, network optimization and system combination, where the DNN models are variants of squeeze-excitation and residual networks. We conducted an ablation study of the effectiveness of each component on the ASVspoof 2019 corpus, and experimental results showed that ASSERT obtained more than 93% and 17% relative improvements over the baseline systems in the two sub-challenges in ASVspooof 2019, ranking ASSERT one of the top performing systems. Code and pretrained models will be made publicly available.
1. Introduction
ASVspoof 2019 addresses replay, text-to-speech, and voice-conversion attacks through Physical Access and Logical Access sub-challenges. ASSERT develops a DNN-based anti-spoofing pipeline combining feature engineering, DNN models, network optimization, and system fusion, achieving substantial improvements over baselines.
- ASVspoof 2019 scope: ASVspoof 2019 covers replay, text-to-speech, and voice-conversion spoofing attacks through Physical Access and Logical Access sub-challenges.Logical Access considers TTS and VC, while Physical Access concerns replay.
- ASSERT pipeline: ASSERT combines feature engineering, DNN models, network optimization, and system fusion into a DNN-based anti-spoofing pipeline.The system was designed to identify core components of a working DNN-based anti-spoofing approach.
- ASSERT models: ASSERT evaluates variants of Squeeze-Excitation and residual networks for detecting replay, TTS, and VC spoofing attacks.The authors report introducing SENet and ResNet with statistical pooling for anti-spoofing.
- Results: More than 93% and 17% relative improvements were achieved over baseline systems on the PA and LA sub-challenges, respectively.The reported improvements were obtained on the ASVspoof 2019 corpus.
- Results: ASSERT was ranked 3rd in the PA sub-challenge and 14th in the LA sub-challenge.These rankings refer to the fusion system submission.
2. ASSERT
ASSERT combines acoustic feature engineering, squeeze-excitation and residual DNN models, optimization, and fusion for anti-spoofing. Its feature inputs include unified segmented maps or whole utterances, while its model variants and training schemes target spoofing classification.
- Feature Engineering: ASSERT uses low-level acoustic features with either unified feature maps or whole-utterance inputs, followed by squeeze-excitation and residual DNN variants.The model set includes SENet34, SENet50, Mean-Std ResNet, Dilated ResNet, and Attentive-Filtering Network.
- Feature Engineering: CQCC features have 30 dimensions and logspec features have 257 dimensions; neither uses voice activity detection or normalization.The authors report empirically better results without these processing steps.
- Feature Engineering: Unified feature maps extend utterances to multiples of M, split them into overlapping segments, and average segment-level DNN outputs per utterance.For ASVspoof 2019, M is 400 and overlap L is either 0 or 200.
- DNN Models: SENet variants append channel-wise transforms to residual backbones, while Mean-Std ResNet pools frame-level features using timestep-wise mean and standard deviation.Mean-Std ResNet is trained with whole-utterance input because its pooling layer supports variable lengths.
- DNN Models: Dilated ResNet uses dilated convolution within residual blocks and accepts only fixed-size input because it has no pooling layer.Attentive-Filtering adds attention-based feature masking before Dilated ResNet using downsampling, upsampling, bilinear interpolation, and skip connections.
- Optimization and Fusion: Training uses binary or spoof-condition multi-class classification, Adam optimization with warm-up scheduling, and greedy logistic-regression fusion with PA and LA effective priors.The effective priors are 0.672 for PA and 0.707 for LA.
3. Experiments
Experiments compared ASSERT’s DNN systems with official LFCC-GMM, CQCC-GMM, and i-vector baselines on ASVspoof 2019 using t-DCF and EER. ASSERT improved substantially over baselines on both PA and LA, while results also exposed overfitting on LA evaluation data.
- Experimental setup: Experiments used LFCC-GMM and CQCC-GMM official baselines alongside implemented i-vector systems on the ASVspoof 2019 PA and LA subsets.The evaluation used minimum normalized t-DCF and EER as spoofing-countermeasure metrics.
- Experimental setup: Table 3 reports an ablation study of single-system results, listing the top two systems for each DNN model under space constraints.Training objectives include multi-class cross entropy and binary cross entropy, with accuracy or EER used for epoch-level model selection.
- Single-system results: 92% and 94% relative improvements over CQCC-GMM were obtained for PA dev t-DCF and EER, while LA achieved 100% relative improvement against CQCC-GMM.The best single system used SENet34 with overlapping unified logspec features, binary cross-entropy training, and development-accuracy model selection.
- Evaluation results: The primary system fused five models based on SENet34, Mean-Std ResNet, SENet50, and Dilated ResNet, with CQCC and logspec variants.The single system was based on SENet34 with logspec features.
- Evaluation results: On evaluation data, ASSERT improved over CQCC-GMM by 93% and 95% on PA t-DCF and EER, and over LFCC-GMM by 27% and 17% on LA t-DCF and EER.ASSERT generalized well across development and evaluation for PA but overfit on development data for LA.
4. Conclusions
The paper introduced ASSERT as a DNN-based anti-spoofing system combining squeeze-excitation and residual networks with feature engineering, optimization, and fusion. Its fusion system improved over baseline systems, while the authors identify metadata analysis and LA model refinement as directions for further study.
- Contributions: ASSERT combines squeeze-excitation and residual-network variants with feature engineering, network optimization, and system-fusion schemes for anti-spoofing.The paper presents this work as a preliminary study of DNN-based countermeasures for speech spoofing attacks.
- Findings: The fusion system attained considerable improvement over baseline systems on the ASVspoof 2019 corpus.The conclusion summarizes the system-level outcome without reporting an additional metric.
- Future work: Metadata analysis and model refinements on LA remain areas for further investigation.This scope boundary is stated as a future direction in the conclusion.