An Initial Investigation on Optimizing Tandem Speaker Verification and Countermeasure Systems Using Reinforcement Learning

Any biometric recognizer is vulnerable to direct spoofing attacks and automatic speaker verification (ASV) is no exception; replay, synthesis and conversion attacks all provoke false acceptances unless countermeasures are used. We focus on voice conversion (VC) attacks. Most existing countermeasures use full knowledge of a particular VC system to detect spoofing. We study a potentially more universal approach involving generative modeling perspective. Specifically, we adopt standard ivector representation and probabilistic linear discriminant analysis (PLDA) back-end for joint operation of spoofing attack detector and ASV system. As a proof of concept, we study a vocoder-mismatched ASV and VC attack detection approach on the NIST 2006 speaker recognition evaluation corpus. We report stand-alone accuracy of both the ASV and countermeasure systems as well as their combination using score fusion and joint approach. The method holds promise. Index Terms: speaker recognition, spoofing, voice conversion attack, i-vector, joint verification and anti-spoofing

This article verifies the reliability of automatic speaker verification (ASV) systems on new synthesis methods based on deep neural networks. ASV systems are widely used and applied regarding secure and effective biometric authentication. On the other hand, the rapid deployment of ASV systems contributes to the increased attention of attackers with newer and more sophisticated spoofing methods. Until recently, speech synthesis of the reference speaker did not seriously compromise the latest ASV systems. This situation is changing with the deployment of deep neural networks into the synthesis process. Projects including WaveNet, Deep Voice, Voice Loop, and many others generate very natural and high-quality speech that may clone voice identity. We are slowly approaching an era where we will not be able to recognize a genuine voice from a synthesized one. Therefore, it is necessary to define the robustness of current ASV systems to new methods of voice cloning. In this article, well-known SVM and GMM as well as new CNN-based ASVs are applied and subjected to synthesized speech from Tacotron 2 with the WaveNet TTS system. The results of this work confirm our concerns regarding the reliability of ASV systems against synthesized speech.

In recent years, speaker verification technologies have received an extensive amount of attention. Designing and developing machines that could communicate with humans are believed to be one of the primary motivations behind such developments. Speaker verification technologies are applied to numerous fields such as security, Biometrics, and forensics. In this paper, the authors study the effects of different languages on the performance of the automatic speaker verification (ASV) system. The MirasVoice speech corpus (MVSC), a bilingual English and Farsi speech corpus, is used in this study. This study collects results from both an I-vector based ASV system and a GMM-UBM based ASV system. The experimental results show that a mismatch between the enrolled data used for training and verification data can lead to a significant decrease in the overall system efficiency. This study shows that it is best to use an i-vector based framework with data from the English language used in the enrollment phase to improve the robustness of the ASV systems. The achieved results in this study indicate that this can narrow the degradation gap caused by the language mismatch.

Children's speaker verification in low and zero resource conditions

Automatic Speaker Verification (ASV) systems are widely used for voice-based identity verification, but recent studies have shown that they are vulnerable to both over-the-line (OTL) and over-the-air (OTA) adversarial attacks. Although various detection methods have been proposed, their robustness against diverse OTA attack scenarios has not been systematically assessed. In this work, we present AdvSV 2.0, a large-scale, open-source dataset for adversarial attacks on ASV systems. AdvSV 2.0 contains 628k utterances (800 hours) spanning 8 ASV models, 4 attack methods, and 9 physical playback-recording device combinations, and includes both OTL and OTA adversarial examples. To better connect digital adversarial optimization with the physical OTA channel, we further propose an OTA attack generation method based on a Neural Replay Simulator (NRS), which simulates replay-channel distortions and improves OTA attack success rates by up to 33.4% after air transmission. To defend against these attacks, we propose CODA-OCC, a COntrastive Domain-Aligned One-Class Classifier. Built on a one-class classification paradigm, CODA-OCC is trained only on bona fide speech and does not rely on any white-box assumption about the ASV backend. Experimental results on AdvSV 2.0 show that CODA-OCC achieves an EER of 11.2% and an AUC of 0.95, outperforming the purification-based baseline by 8.6% absolute EER, and providing a robust foundation for studying OTA adversarial robustness in ASV.

With the growing number of devices that use voice biometric system for verification purposes, it becomes mandatory to accurately verify the target speaker. Moreover, intruders use different spoofing attacks to fool an automatic speaker verification (ASV) system. Voice replay attack being the easiest to generate among other spoofing attacks is frequently employed in front of the ASV systems of these devices to get access to the system for malicious purposes. Thus, we need to develop robust systems that are capable of classifying the target speaker and differentiate between the bonafide and replay voice. To combat the vulnerabilities of ASV systems, we proposed an automatic speaker verification and replay attack detection (ASVRAD) system. For this purpose, we proposed a novel audio feature i.e., Glottal Flow Gammatone Cepstrum Coefficients (GLGTCC), an enhanced version of Glottal Flow Cepstrum Coefficients (GLFCC) for accurate classification of the bonafide speakers and voice replay attacks detection. Our GLGTCC features can reliably capture the speaker-specific attributes from their vocal excitation patterns in the audio. Later, we employed a multiclass support vector machine (SVM) to classify the bonafide speakers, and a binary SVM for classification of bonafide and replay audios. Experimental results of our system on the standard ASVspoof 2019 PA dataset achieved the lowest EER of 3.00% for speaker verification and 18.38% on voice replay attacks over the comparative methods.

Voice phishing (vishing) is increasingly popular due to the development of speech synthesis technology. In particular, the use of deep learning to generate an arbitrary-content audio clip simulating the victim's voice makes it difficult not only for humans but also for automatic speaker verification (ASV) systems to distinguish. Countermeasure (CM) systems have been developed recently to help ASV combat synthetic speech. In this work, we propose BTS-E, a framework to evaluate the correlation between Breathing, Talking (speech), and Silence sounds in an audio clip, then use this information for deepfake detection tasks. We argue that natural human sounds, such as breathing, are hard to synthesize by Text-to-speech (TTS) system. We conducted a large-scale evaluation using ASVspoof 2019 and 2021 evaluation set to validate our hypothesis. The experiment results show the applicability of the breathing sound feature in detecting deepfake voices. In general, the proposed system significantly increases the performance of the classifier by up to 46%.

Deep generative variational autoencoding for replay spoof detection in automatic speaker verification

This paper explores the application of active learning to enhance machine learning classifiers for spoofing detection in automatic speaker verification (ASV) systems. Leveraging the ASVspoof 2019 database, we integrate an active learning framework with traditional machine learning workflows, specifically focusing on Random Forest (RF) and Multilayer Perceptron (MLP) classifiers. The active learning approach was implemented by initially training models on a small subset of data and iteratively selecting the most uncertain samples for further training, which allowed the classifiers to refine their predictions effectively. Experimental results demonstrate that while the MLP initially outperformed RF with an accuracy of 95.83% compared to 91%, the incorporation of active learning significantly improved RF's performance to 94%, narrowing the performance gap between the two models. After applying active learning, both classifiers showed enhanced precision, recall, and F1-scores, with improvements ranging from 3% to 5%. This study provides valuable insights into the role of active learning in boosting the efficiency of machine learning models for dynamic spoofing scenarios in ASV systems. Future research should focus on designing advanced active learning techniques and exploring their integration with other machine learning paradigms to further enhance ASV security.

With the development of speech synthesis technology, automatic speaker verification (ASV) systems have encountered the serious challenge of spoofing attacks. In order to improve the security of ASV systems, many antispoofing countermeasures have been developed. In the front-end domain, much research has been conducted on finding effective features which can distinguish spoofed speech from genuine speech and the published results show that dynamic acoustic features work more effectively than static ones. In the back-end domain, Gaussian mixture model (GMM) and deep neural networks (DNNs) are the two most popular types of classifiers used for spoofing detection. The log-likelihood ratios (LLRs) generated by the difference of human and spoofing log-likelihoods are used as spoofing detection scores. In this paper, we train a five-layer DNN spoofing detection classifier using dynamic acoustic features and propose a novel, simple scoring method only using human log-likelihoods (HLLs) for spoofing detection. We mathematically prove that the new HLL scoring method is more suitable for the spoofing detection task than the classical LLR scoring method, especially when the spoofing speech is very similar to the human speech. We extensively investigate the performance of five different dynamic filter bank-based cepstral features and constant Q cepstral coefficients (CQCC) in conjunction with the DNN-HLL method. The experimental results show that, compared to the GMM-LLR method, the DNN-HLL method is able to significantly improve the spoofing detection accuracy. Compared with the CQCC-based GMM-LLR baseline, the proposed DNN-HLL model reduces the average equal error rate of all attack types to 0.045%, thus exceeding the performance of previously published approaches for the ASVspoof 2015 Challenge task. Fusing the CQCC-based DNN-HLL spoofing detection system with ASV systems, the false acceptance rate on spoofing attacks can be reduced significantly.

This paper addresses the design and implementation of automatic speaker verification (ASV) systems. There is great interest in developing and increasing the performance of ASV applications, taking into account the advantages offered when compared to other biometrical methods. State-of-the-art speaker recognizers are based on statistical models such as GMM, HMM, SVM, ANN or hybrid models. This work reports experiments on prompted text speaker verification on a Romanian corpus. First, a Hidden Markov Model (HMM) based continuous speech recognizer is built at context independent, single mixture, monophone level. Then the client model and the world model are built using appropriate speech data. The text-prompted speaker verification system is based on sentence HMMs, constructed for the key text by concatenating corresponding acoustic models. In the verification stage, the normalized log-likelihood is derived as the difference between the log-likelihood obtained through Viterbi forced alignment of the client model and world model, respectively. Finally a procedure used to determine the verification performances is presented.

In the last few years, we have witnessed an exponential growth in voice spoofing attacks. The intruders employ different types of attacks such as speech synthesis where they use the machine generated speech against any target person to fool the automatic speaker verification (ASV) systems for various tasks i.e. home control, bank account access, etc. The availability of modern-day advanced tools has made it convenient to launch such types of voice spoofing attacks. To overcome the challenges associated with bypassing the security of ASV systems using the synthetic speech, we propose an effective synthetic speech detector using a fusion of spectral features. More specifically, we propose a fused feature vector consisting of MFCC, GTCC, Spectral Flux, and Spectral Centroid for audio signal representation. This fused feature set is capable of capturing the traits of speech variation attributes of genuine signal and algorithmic artifacts of synthetic signals. These features are further used to train the bilstm to classify the signal as genuine or spoof. The proposed framework is capable of detecting both the voice conversion and synthetic speech attacks on ASV systems. Performance of our framework is evaluated on ASVspoof 2019 LA dataset. Our experimental results illustrate the effectiveness of the proposed framework for logical access attacks (voice conversion and cloned/synthetic voice) detection.

Voice-based phishing and wire-fraud attacks have become a topical problem in the recent years due to the emergence of advanced AI-based speech synthesis models. These models can generate realistic speech signal of a known target that is difficult to differentiate from a bonafide voice of real human. This proved to be an issue by the recent security reports related to vishing attacks on bank call centers, fraud or pranking of public figures, and spoofing of voice authentication systems. Current approaches to address voice fraud issue are based on applying an Automatic Speaker Verification (ASV) system. In most cases, these systems are tuned on datasets that consist of wideband quality bonafide and spoofed voice samples. This makes ASV systems vulnerable to speech signal degradation caused by voice encoding in cellular network and Voice over IP (VoIP). However, performance evaluation of ASV (namely Equal Error Rate (EER) estimation) is almost exclusively available only for the cellular networks. Thus, performance of ASV systems for modern VoIP applications remains unclear. In this paper, we evaluate the modern ASV systems on audio compressed with codecs used in both cellular networks (AMR and GSM codecs) and VoIP applications (G.711, G.722, AAC, Lyra and Opus codecs). In addition, ASV performance was tested using popular VoIP application (Discord). Obtained results have shown that codec application results in considerable (up to two times) EER increase compared to the baseline results. Moreover, we observed up to three times increase in EER on data transmitted using Discord. We propose to apply hard samples mining to the training process in order to improve the accuracy of ASV systems on compressed voice samples. It allows to reduce EER from 21% down to 16% even for the most distorted samples obtained after aggressive voice compression by GSM codecs. Note, that improvement for real VoIP application is even higher - with EER on Discord data decrease from 35% to 20%.

Factor analysis (FA) is one of the key advances presented in recent speaker verification evaluations. This technique is able to successfully remove session variability effects and it is currently used in many state-of-the-art automatic speaker verification systems. This paper addresses several practical issues in using an FA model in order to speed up model training and to achieve good performance. A parallelized training algorithm as well as maximum-likelihood estimation are proposed for fast training. The front-end feature normalization techniques are also investigated in the context of FA model. We demonstrate that factor analysis is very robust, and can be successfully applied to various kinds of feature normalization. Moreover, the proposed parallelized MLE implementation speeds up the training procedure from several days to several hours without sacrificing the performance.

The extraction of speaker-related features through speaker representations or utterance embeddings has been extensively studied for years. Convolutional Neural Networks (CNNs) are the most commonly used backbone network for feature extraction, and deep Residual Networks (ResNets), which incorporate residual connectivity into CNNs, achieve effective frame-level feature extraction inside the same input samples. ResNets have demonstrated impressive performance as speaker embedding extractors. With emerging this backbone and its advanced extensions, such as Res2Net, as dominant backbone networks in speaker recognition, and despite its widespread adoption, a significant gap in current speaker embeddings is the unmodified use of these successful backbones from the visual domain, without considering the specific adaptation requirements for the speech field. Tailoring the backbone to account for the unique characteristics of speech representations ensures effective learning in Automatic Speaker Verification (ASV) systems. This thesis concentrates on the backbone network, the fundamental and most critical component of Deep Neural Networks (DNNs), responsible for extracting features from input data and mapping them into distinct feature representations. One of the key questions addressed in the thesis is the following: Which backbone is most robust and yields the most generalized speaker representation, with the potential to be applied to many existing ASV systems?. Through a strategic evaluation, the study analyzed the performance of integrating different residual-based backbones into the most popular ResNet-based fundamental baselines, ResNet-34 and ECAPA-TDNN based models, with highly different model complexity and CNN stem representations. To this end, the proposed architectures were evaluated on the publicly available VoxCeleb dataset. This thesis introduces an innovative approach to speaker modeling by incorporating scale and cardinality dimensions for the first time. Experimental findings reveal that utilizing multi-scale and multi-branch aggregated residual transformations within the Res2NeXt backbone yields considerable performance improvements. Notably, 6s8g and 8s8g variants of Res2NeXt, prove particularly beneficial in improving the representational capability of learned features for speaker embedding. The consistent superiority, excellent compatibility, and robustness of Res2NeXt compared to the current popular ResNet and Res2Net backbones across various architectures and conditions demonstrate its broad applicability and effectiveness as a backbone in speaker modeling. To further evaluate the Robustness and Generalization (R&G) capabilities of the proposed speaker embeddings, the thesis expanded its scope to cross-domain applications, specifically Speech Emotion Recognition (SER), one of the most challenging downstream tasks. A comparative analysis of the proposed speaker embeddings was conducted using two benchmark datasets, IEMOCAP and CREMA-D. The overarching goal of the thesis is to contribute to the development of general-purpose and universal speech representations that generalize effectively across multiple speech-related tasks. By examining the alignment and divergence of embeddings for ASV and SER, the study seeks to answer a fundamental question: Can a single representation framework cater to both speaker verification and emotion recognition or are task-specific embeddings inevitable?. The results confirm that speaker embeddings derived from the Res2NeXt backbone are also highly effective for the SER task, maintaining their performance across diverse evaluation protocols and data variations. This demonstrates the generalizability of Res2NeXt embeddings in the cross-domain task. The study attributes their success to the ability of Res2NeXt-based models to capture rich speech features, including tone, pitch, and phonation patterns, during speaker recognition training—features that are also vital for emotion recognition.