Source-linked AI summary
Light Convolutional Neural Network with Feature Genuinization for Detection of Synthetic Speech Attacks
Zhenzong Wu, Rohan Kumar Das, Jichen Yang, Haizhou Li
TL;DR
Natural-sounding TTS and VC speech can threaten automatic speaker verification, and countermeasures may struggle with attacks from unseen sources. The paper learns a feature transformer from genuine speech and pairs it with an LCNN; on ASVspoof 2019 logical access, the system improves over the baseline and reported single-system comparisons.
Problem
Natural-sounding TTS and VC speech threatens ASV, while spoofing countermeasures may not generalize to synthetic speech from unseen sources.
Source: §1 · p. 1Method
A CNN-based transformer learned from genuine-speech features maps input features into a genuine-learned domain before LCNN classification.
Source: §1 · pp. 1–2, §3 · p. 3Results
FG-LCNN improves over baseline LCNN on the evaluation set containing unseen attacks and outperforms reported single systems in EER on ASVspoof 2019 logical access.
Source: §5 · pp. 3–4Takeaways & Limitations
The studies support feature genuinization with an LCNN as an effective approach for detecting synthetic speech attacks on the ASVspoof 2019 logical access corpus.
Source: §6 · p. 4
Abstract
from arXiv · showhide
Modern text-to-speech (TTS) and voice conversion (VC) systems produce natural sounding speech that questions the security of automatic speaker verification (ASV). This makes detection of such synthetic speech very important to safeguard ASV systems from unauthorized access. Most of the existing spoofing countermeasures perform well when the nature of the attacks is made known to the system during training. However, their performance degrades in face of unseen nature of attacks. In comparison to the synthetic speech created by a wide range of TTS and VC methods, genuine speech has a more consistent distribution. We believe that the difference between the distribution of synthetic and genuine speech is an important discriminative feature between the two classes. In this regard, we propose a novel method referred to as feature genuinization that learns a transformer with convolutional neural network (CNN) using the characteristics of only genuine speech. We then use this genuinization transformer with a light CNN classifier. The ASVspoof 2019 logical access corpus is used to evaluate the proposed method. The studies show that the proposed feature genuinization based LCNN system outperforms other state-of-the-art spoofing countermeasures, depicting its effectiveness for detection of synthetic speech attacks.
1. Introduction
Natural-sounding TTS and VC speech threatens ASV, while spoofing countermeasures can struggle with attacks from unseen sources. The paper proposes learning a genuine-speech domain and using it to distinguish synthetic speech.
1. Introduction
- Prior spoofing-countermeasure research has explored handcrafted front-end features and data-driven classifiers, motivating the need for both discriminative features and effective classification.The cited front ends include CFCCIF, LFCC, and CQCC, while classifier examples include LCNN, squeeze-excitation, and residual networks.Source: §1 · p. 1
- Genuine speech is described as less varied than synthetic speech, whose attacks can come from many TTS and VC algorithms.The authors hypothesize that genuine speech's consistent characteristics distinguish it from varied synthetic speech.Source: §1 · p. 1
- Existing countermeasures may lack generalization to synthetic speech from unseen sources, motivating a method based on genuine-speech characteristics.The authors also cite prior work that modifies spoofed-speech probabilities toward genuine speech as a related approach.Source: §1 · p. 1
- The proposed genuinization transformer is learned with a CNN from genuine-speech features and is intended to preserve genuine inputs while producing different outputs for spoof inputs.The authors hypothesize that the transform amplifies the difference between the two classes.Source: §1 · p. 1
- Feature genuinization projects feature representations onto a domain learned from genuine features, then uses the transformer with an LCNN to detect synthetic speech attacks.Source: §1 · p. 2
2. Feature Genuinization
Feature genuinization trains a transformer using only genuine-speech features, then applies it to features from either class. Its CNN architecture uses an encoder-decoder design to learn genuine characteristics.
2. Feature Genuinization
- The process has two stages: train the transformer on genuine-speech features, then use it to transform features so genuine and spoof speech are more discriminable.The stated aim is to leave genuine features' characteristics unchanged while projecting spoof features to different outputs.Source: §2 · p. 2
- The transformer uses a full-convolutional CNN with no fully connected layers and functions similarly to an autoencoder.The authors say this design focuses on temporal correlation and reduces training parameters and training time.Source: §2 · p. 2
- Its encoder compresses the input through strided convolutions and leaky ReLU; the decoder reverses the process with deconvolution and ReLU.The resulting autoencoder learns genuine-speech characteristics, which the paper says amplifies discrimination in the transformed domain.Source: §2 · p. 2
- After training, the transformer maps genuine or spoofed features into a domain learned from genuine-feature characteristics.Source: §2 · p. 2
3. LCNN with Feature Genuinization
The proposed system transforms speech features with the genuinization transformer before passing them to an LCNN. The LCNN is selected for compactness and anti-spoofing efficacy.
3. LCNN with Feature Genuinization
- The LCNN processes transformed features and is chosen for its compactness and efficacy in anti-spoofing.Source: §3 · pp. 2–3
- The system transforms log power spectrum features into genuinized features, which are then classified by the LCNN to detect spoofing attacks.Training uses the transformed features with their corresponding labels.Source: §3 · p. 3
- The LCNN uses Max-Feature-Map activation to learn compact features and suppress some neuron activations, reducing output dimensionality and producing more discriminative feature maps.The passage contrasts this with ReLU's sparse, high-dimensional features.Source: §3 · p. 3
4. Experiments
The experiments evaluate synthetic-speech detection on the ASVspoof 2019 logical access corpus, using its train, development, and evaluation partitions. The setup uses CQT-derived log power spectrum features and regularization during LCNN training.
4.1. Database
- The experiments use the ASVspoof 2019 logical access corpus, whose evaluation-set spoof examples come from methods unseen in training.Development attacks were created using two VC and four TTS methods; genuine speech comes from VCTK, with 107 speakers and no speaker overlap across subsets.Source: §4.1 · p. 3
- The corpus has train, development, and evaluation partitions, and benchmarking uses t-DCF as the primary metric and EER as the secondary metric.The t-DCF calculation combines the supplied ASV-system scores with spoofing-countermeasure scores.Source: §4.1 · p. 3
4.2. Experimental Setup
- Input features are CQT-derived log power spectra with 9 octaves, 96 bins per octave, and 863 × 256 dimensions per example.Files are padded with the last frame or cropped to 256 frames.Source: §4.2 · p. 3
- LCNN training uses dropout of 0.4 and weight decay of 2 × 10^-4 to address overfitting, with layer and node counts optimized on the development set.An extra batch-normalization step follows max pooling to increase stability and convergence speed.Source: §4.2 · p. 3
5. Results and Discussion
On ASVspoof 2019 logical access, FG-LCNN outperforms other reported single systems in EER and improves on the LCNN baseline, particularly on evaluation attacks of unseen nature.
5. Results and Discussion
- FG-LCNN improves on the LCNN baseline, with the clearest improvement on the evaluation set containing more challenging unseen attacks.The authors also report much better performance than the two ASVspoof 2019 challenge baselines.Source: §5 · p. 3, Table 2 · p. 4
- On the evaluation set, the feature-spoofing contrast system FS-LCNN degrades from baseline LCNN and performs worse than FG-LCNN.The comparison uses evaluation results because development-set results can show very accurate detection.Source: Table 3 · p. 4, §5 · p. 4
- FG-LCNN outperforms other reported single systems in EER on the ASVspoof 2019 logical-access evaluation set.LCNN-based systems were the best-performing single-system baseline in the comparison.Source: Table 4 · p. 4, §5 · p. 4
6. Conclusion
The conclusion presents feature genuinization followed by LCNN as a method for synthetic-speech detection, using genuine-speech characteristics to distinguish transformed genuine and spoof features. ASVspoof 2019 logical-access studies support its effectiveness, and replay-attack detection is left for future work.
6. Conclusion
- A CNN learns from genuine-speech characteristics to transform genuine features closer to the genuine distribution while producing different outputs for spoof features.Source: §6 · p. 4
- Studies on the ASVspoof 2019 logical-access corpus show the feature-genuinization LCNN system is effective for detecting synthetic-speech attacks.Source: §6 · p. 4
- The comparison presents the system as one of the strong single anti-spoofing systems.Source: §6 · p. 4
- Future work will extend the studies to replay-attack detection.Source: §6 · p. 4