Spoof Speech Research Articles

Time Series-Based Spoof Speech Detection Using Long Short-Term Memory and Bidirectional Long Short-Term Memory

Detecting fake speech in voice-based authentication systems is crucial for reliability. Traditional methods often struggle because they can't handle the complex patterns over time. Our study introduces an advanced approach using deep learning, specifically Long Short-Term Memory (LSTM) and Bidirectional LSTM (BiLSTM) models, tailored for identifying fake speech based on its temporal characteristics. We use speech signals with cepstral features like Mel-frequency cepstral coefficients (MFCC), Constant Q cepstral coefficients (CQCC), and open-source Speech and Music Interpretation by Large-space Extraction (OpenSMILE) to directly learn these patterns. Testing on the ASVspoof 2019 Logical Access dataset, we focus on metrics such as min-tDCF, Equal Error Rate (EER), Recall, Precision, and F1-score. Our results show that LSTM and BiLSTM models significantly enhance the reliability of spoof speech detection systems.

DeepDet: YAMNet with BottleNeck Attention Module (BAM) for TTS synthesis detection

Spoofed speeches are becoming a big threat to society due to advancements in artificial intelligence techniques. Therefore, there must be an automated spoofing detector that can be integrated into automatic speaker verification (ASV) systems. In this study, we recommend a novel and robust model, named DeepDet, based on deep-layered architecture, to categorize speech into two classes: spoofed and bonafide. DeepDet is an improved model based on Yet Another Mobile Network (YAMNet) employing a customized MobileNet combined with a bottleneck attention module (BAM). First, we convert audio into mel-spectrograms that consist of time–frequency representations on mel-scale. Second, we trained our deep layered model using the extracted mel-spectrograms on a Logical Access (LA) set, including synthesized speeches and voice conversions of the ASVspoof-2019 dataset. In the end, we classified the audios, utilizing our trained binary classifier. More precisely, we utilized the power of layered architecture and guided attention that can discern the spoofed speech from bonafide samples. Our proposed improved model employs depth-wise linearly separate convolutions, which makes our model lighter weight than existing techniques. Furthermore, we implemented extensive experiments to assess the performance of the suggested model using the ASVspoof 2019 corpus. We attained an equal error rate (EER) of 0.042% on Logical Access (LA), whereas 0.43% on Physical Access (PA) attacks. Therefore, the performance of the proposed model is significant on the ASVspoof 2019 dataset and indicates the effectiveness of the DeepDet over existing spoofing detectors. Additionally, our proposed model is robust enough that can identify the unseen spoofed audios and classifies the several attacks accurately.

INCREASING ROBUSTNESS OF I-VECTORS VIA MASKING: A CASE STUDY IN SYNTHETIC SPEECH DETECTION

Ensuring security in speaker recognition systems is crucial. In the past years, it has been demonstrated that spoofing attacks can fool these systems. In order to deal with this issue, spoof speech detection systems have been developed. While these systems have served with a good performance, their effectiveness tends to degrade under noise. Traditional speech enhancement methods are not efficient for improving performance, they even make it worse. In this research paper, performance of the noise mask obtained via a convolutional neural network structure for reducing the noise effects was investigated. The mask is used to suppress noisy regions of spectrograms in order to extract robust i-vectors. The proposed system is tested on the ASVspoof 2015 database with three different noise types and accomplished superior performance compared to the traditional systems. However, there is a loss of performance in noise types that are not encountered during training phase.

Improving self-supervised learning model for audio spoofing detection with layer-conditioned embedding fusion

The application of voice recognition systems has increased by a great deal with technology. This has allowed adversaries to falsely claim access to these systems by spoofing the identity of a target speaker. The existing supervised learning (SL)-based countermeasures are yet to provide a complete solution against the newly evolving spoofing attacks. To tackle this problem, we explore self-supervised learning (SSL)-based frameworks. At first, we implement widely used SSL frameworks, where our target is identifying spoofed speech. We report a considerable performance improvement over the SL state-of-the-art baseline as a whole. Then, we perform an attack-wise comparative analysis between SL and SSL frameworks. While the SSL performs better in most cases, there are certain attacks where the SL outperforms it. Hence, we hypothesize that there is scope to jointly utilize information effectively included by both these models for better performance. To do that, we first perform conventional weighted score fusion between the SL and best-performing SSL models, which reduces the EER, outperforming both the state-of-the-art SL and best-performing SSL framework. Then, we propose an embedding fusion scheme that minimizes the embedding distribution between the selected SL and SSL representations. To select the appropriate layers, we perform a comprehensive statistical analysis. The proposed fusion scheme outperforms the score fusion method and shows that the SSL performance can be improved by effectively including learned knowledge from the SL framework. The final EER achieved on the ASVspoof 2019 logical access (LA) database is 0.177%, a significant improvement over our baseline. Using the ASVspoof 2021 LA as a blind evaluation dataset, our proposed embedding fusion scheme reduces the EER to 2.666%.

SDI: A tool for speech differentiation in user identification

User identification (UI) with human speech is important in voice biometrics for user authentication. However, the recognized speech is at risk of being spoofed when an intruder attempts to access a voice-secured system. This can be done using a voice spoofing attack, which involves the use of manipulated speech signals for illegal purposes, including accessing authorized users’ data or sensitive information from a secured system. In recent years, different countermeasures have been developed to protect biometric systems from such spoofing attacks, but loopholes remain in such secured systems. Therefore, we urgently need a speech differentiation system that differentiates spoofed speech from bona fide speech in the process of UI. To this end, this paper proposes an anti-spoofing method for user authentication in voice biometrics. A refined mechanism for support vector machine (SVM) was designed; it implements threshold checking and frequency matching mechanisms, resulting in a speech differentiator with integrated SVM (SDI-SVM), or an integrated speech differentiator (SDI) for short. The mechanism was experimentally evaluated, and the results show that SVM outperformed its counterparts in the task of UI. SVM was upgraded into SDI-SVM, which can realize user authentication in voice biometrics by merging speech differentiation into UI. The experimental results also show that SDI-SVM produced better results than those demonstrated in recent works. Thus, the proposed solution can enhance the security of biometric technologies.

Local Self-attention-based Hybrid Multiple Instance Learning for Partial Spoof Speech Detection

The development of speech synthesis technology has increased the attention toward the threat of spoofed speech. Although various high-performance spoofing countermeasures have been proposed in recent years, a particular scenario is overlooked: partially spoofed audio, where spoofed utterances may contain both spoofed and bona fide segments. Currently, the research on partially spoofed speech detection is lacking. The existing methods either train with partially spoofed speech at utterance level, resulting in gradient conflicting at the segment level, or directly train with segment level data, which requires segment labels that are difficult to obtain in practice. In this study, to better detect partially spoofed speech when only utterance labels are available, we formulate partially spoofed speech detection into a multiple instance learning (MIL) problem. The typical MIL uses a pooling layer to fuse patch scores as a whole, and we propose a hybrid MIL (H-MIL) framework based on max and log-sum-exp pooling methods, which can learn better segment representations to improve partially spoofed speech detection performance. Theoretical and experimental verification shows that H-MIL can effectively relieve the gradient conflicting and gradient vanishing problems. In addition, we analyze the local correlations between segments and introduce a local self-attention mechanism to enhance segment features, which further promotes the detection performance. In our experiments, we provide not only detection results at the segment and utterance levels but also some detailed visualization analysis, including the effect of spoof ratio and cross-dataset detection. The experimental results demonstrate the effective detection performance of our method at both the utterance and segment levels, especially when dealing with low spoof ratio attacks. The results confirm that our approach can better deal with partially spoofed speech detection than previous methods.

Morse wavelet transform-based features for voice liveness detection

The need for Voice Liveness Detection (VLD) has emerged particularly for the security of Automatic Speaker Verification (ASV) systems. Existing Spoofed Speech Detection (SSD) systems rely on attack-specific approaches to detect spoofed speech. However, to safeguard ASV systems against all the kinds of spoofing attacks (known as well as unknown attacks), determining whether a speech is uttered live (genuine) or not, is important. To that effect, in this work, we propose the detection of pop noise using Morse wavelet for VLD task. Pop noise is a discriminative acoustic cue that is present in live speech and is absent/diminished in spoofed speech. It is captured by the microphone in the form of sudden bursts of air from a live speaker’s mouth due to the close proximity of the speaker with the microphone. To validate this hypothesis, we present an analysis of pop noise energy w.r.t. distance and found that it decreases exponentially with distance. Furthermore, pop noise is said to be present in very low frequency regions. To capture the pop noise effectively, we propose to exploit the excellent frequency resolution of Continuous Wavelet Transform (CWT) using Generalized Morse Wavelets (GMWs). GMWs are a superfamily of analytic wavelets. To that effect, in this work, we have analysed the suitability of GMWs for pop noise detection for VLD task using the POp noise COrpus (POCO). The wavelet parameters are fine-tuned according to the VLD task. Furthermore, the performance of VLD system is evaluated for various subband frequencies, and it is observed that the subband of 1 to 50Hz gives the best performance accuracy of 90.55% and 88.43% on the Dev and Eval sets, respectively. In addition, phoneme-based analysis shows the dependence of the performance of the VLD system on the type of phonemes in the utterances. It is shown that phonemes, such as plosives and fricatives show distinct pop noise as compared to other phonemes. Furthermore, the extension of the POCO dataset is used for experiments where simulated reverberation is added to spoofed signals, assuming the attacker (or the recording device) is positioned at various distances. This leads to the studying the effect of speaker-attacker distance. Similar to the previous results, it is observed that for the reverberated case too, the optimal frequency subband for VLD task is 1 to 50Hz, across all the distances. Furthermore, the proposed feature set is evaluated using three classifiers, namely, Convolutional Neural Network (CNN), Light CNN (LCNN), and Residual Neural Network (ResNet), for POCO dataset as well as reverberated POCO dataset. It is observed that CNN gives the highest accuracy of 88.43% on Eval set of the POCO dataset. Furthermore, the proposed features are also evaluated under the assumptions of two ideal scenarios — when the ASV system is strictly under attack, and when it is strictly not under attack. It is observed that the proposed Morse wavelet-based VLD system rejected 89% of the spoofed utterances, and accepted 88.30% of the genuine utterances.

Gaussian-Filtered High-Frequency-Feature Trained Optimized BiLSTM Network for Spoofed-Speech Classification.

Voice-controlled devices are in demand due to their hands-free controls. However, using voice-controlled devices in sensitive scenarios like smartphone applications and financial transactions requires protection against fraudulent attacks referred to as "speech spoofing". The algorithms used in spoof attacks are practically unknown; hence, further analysis and development of spoof-detection models for improving spoof classification are required. A study of the spoofed-speech spectrum suggests that high-frequency features are able to discriminate genuine speech from spoofed speech well. Typically, linear or triangular filter banks are used to obtain high-frequency features. However, a Gaussian filter can extract more global information than a triangular filter. In addition, MFCC features are preferable among other speech features because of their lower covariance. Therefore, in this study, the use of a Gaussian filter is proposed for the extraction of inverted MFCC (iMFCC) features, providing high-frequency features. Complementary features are integrated with iMFCC to strengthen the features that aid in the discrimination of spoof speech. Deep learning has been proven to be efficient in classification applications, but the selection of its hyper-parameters and architecture is crucial and directly affects performance. Therefore, a Bayesian algorithm is used to optimize the BiLSTM network. Thus, in this study, we build a high-frequency-based optimized BiLSTM network to classify the spoofed-speech signal, and we present an extensive investigation using the ASVSpoof 2017 dataset. The optimized BiLSTM model is successfully trained with the least epoch and achieved a 99.58% validation accuracy. The proposed algorithm achieved a 6.58% EER on the evaluation dataset, with a relative improvement of 78% on a baseline spoof-identification system.

The PartialSpoof Database and Countermeasures for the Detection of Short Fake Speech Segments Embedded in an Utterance

Automatic speaker verification is susceptible to various manipulations and spoofing, such as text-to-speech synthesis, voice conversion, replay, tampering, adversarial attacks, and so on. We consider a new spoofing scenario called <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">“Partial Spoof”</i> (PS) in which synthesized or transformed speech segments are embedded into a bona fide utterance. While existing countermeasures (CMs) can detect fully spoofed utterances, there is a need for their adaptation or extension to the PS scenario. We propose various improvements to construct a significantly more accurate CM that can detect and locate short-generated spoofed speech segments at finer temporal resolutions. First, we introduce newly developed self-supervised pre-trained models as enhanced feature extractors. Second, we extend our PartialSpoof database by adding segment labels for various temporal resolutions. Since the short spoofed speech segments to be embedded by attackers are of variable length, six different temporal resolutions are considered, ranging from as short as 20 ms to as large as 640 ms. Third, we propose a new CM that enables the simultaneous use of the segment-level labels at different temporal resolutions as well as utterance-level labels to execute utterance- and segment-level detection at the same time. We also show that the proposed CM is capable of detecting spoofing at the utterance level with low error rates in the PS scenario as well as in a related logical access (LA) scenario. The equal error rates of utterance-level detection on the PartialSpoof database and ASVspoof 2019 LA database were 0.77 and 0.90%, respectively.

Voice spoofing countermeasure for voice replay attacks using deep learning

In our everyday lives, we communicate with each other using several means and channels of communication, as communication is crucial in the lives of humans. Listening and speaking are the primary forms of communication. For listening and speaking, the human voice is indispensable. Voice communication is the simplest type of communication. The Automatic Speaker Verification (ASV) system verifies users with their voices. These systems are susceptible to voice spoofing attacks - logical and physical access attacks. Recently, there has been a notable development in the detection of these attacks. Attackers use enhanced gadgets to record users’ voices, replay them for the ASV system, and be granted access for harmful purposes. In this work, we propose a secure voice spoofing countermeasure to detect voice replay attacks. We enhanced the ASV system security by building a spoofing countermeasure dependent on the decomposed signals that consist of prominent information. We used two main features— the Gammatone Cepstral Coefficients and Mel-Frequency Cepstral Coefficients— for the audio representation. For the classification of the features, we used Bi-directional Long-Short Term Memory Network in the cloud, a deep learning classifier. We investigated numerous audio features and examined each feature’s capability to obtain the most vital details from the audio for it to be labelled genuine or a spoof speech. Furthermore, we use various machine learning algorithms to illustrate the superiority of our system compared to the traditional classifiers. The results of the experiments were classified according to the parameters of accuracy, precision rate, recall, F1-score, and Equal Error Rate (EER). The results were 97%, 100%, 90.19% and 94.84%, and 2.95%, respectively.

On significance of constant-Q transform for pop noise detection

Liveness detection has emerged as an important research issue for many biometrics, such as face, iris, hand geometry, etc. and significant research efforts are reported in the literature. However, less emphasis is given to liveness detection for voice biometrics or Automatic Speaker Verification (ASV). Voice Liveness Detection (VLD) can be a potential technique to detect spoofing attacks in ASV system. Presence of pop noise in the speech signal of live speaker provides the discriminative acoustic cue to distinguish between genuine vs. spoofed speech in the framework of VLD. Pop noise comes out as a burst at the lips, which is captured by the ASV system (since the speaker and microphone are close enough), indicating the liveness of the speaker and provides the basis of VLD. In this paper, we present the Constant-Q Transform (CQT) -based approach over the traditional Short-Time Fourier Transform (STFT) -based algorithm (baseline). With respect to Heisenberg’s uncertainty principle in signal processing framework, the CQT has variable spectro-temporal resolution, in particular, better frequency resolution for low frequency region and better temporal resolution for high frequency region, which can be effectively utilized to identify the low frequency characteristics of pop noise. We have also compared proposed algorithm with cepstral features, namely, Linear Frequency Cepstral Coefficients (LFCC) and Constant-Q Cepstral Coefficients. The experiments are performed on recently released POp noise COrpus (POCO) dataset with various statistical, discriminative, and deep learning-based classifiers, namely, Gaussian Mixture Model (GMM), Support Vector Machine (SVM), Convolutional Neural Networks (CNN), Light-CNN (LCNN), and Residual Network (ResNet), respectively. The significant improvement in performance, in particular, an absolute improvement of 14.23% and 10.95% in terms of percentage classification accuracy on development and evaluation set, respectively, is obtained for the proposed CQT-based algorithm along with SVM classifier, over the STFT-SVM (baseline) system. Similar trend of the performance improvement is observed for the GMM, CNN, LCNN, and ResNet classifiers for the proposed CQT-based algorithm vs. traditional STFT-based algorithm. The analysis is further extended by simulating the replay mechanism (in the standard framework of ASVSpoof-2019 PA challenge dataset) on the subset of POCO dataset in order to observe the effect of room acoustics onto the performance of the VLD system. By embedding the moderate simulated replay mechanism in POCO dataset, we obtained the percentage classification accuracy of 97.82% on evaluation set.

An Overview of Spoof Detection in ASV Systems

In current years, voice based application are used broadly in varied applications for speaker recognition. By and by, there is a wide work in the investigation of parodying and against mocking for Automatic Speaker Verification (ASV) framework. The current advancement within the ASV system ends up interest to secure these voice biometric systems for existent world applications. This paper provides the literature of spoofing detection, novel acoustic feature representations, deep learning, end-to-end systems, etc. Moreover, it conjointly summaries previous studies of spoofing attacks with stress on SS, VC, and replay alongside recent efforts to develop countermeasures for spoof speech detection and speech sound disorder tasks.

Effectiveness of energy separation-based instantaneous frequency estimation for cochlear cepstral features for synthetic and voice-converted spoofed speech detection

In this article, we propose Cochlear Filter Cepstral Coefficient-Instantaneous Frequency feature set using Energy Separation Algorithm (CFCCIF-ESA) feature set to detect the speech synthesis (SS) and voice conversion (VC)-based spoofing attacks. The SS- and VC-based spoof generation techniques predominantly uses the magnitude spectrum information, neglecting the phase information. Hence, SS and VC generated speech signal possess the distorted phase in time or frequency-domain. In this work, we exploit this anomaly in phase to efficiently detect the spoofing attack. Here, instantaneous frequency (IF) is utilized to represent the phase information as IF is nothing but the derivative of unwrapped instantaneous (analytic) phase. The experiments are performed on ASVSpoof-2015 challenge dataset, which is specifically designed to do Spoof Speech Detection (SSD) task for SS and VC. In ASVSpoof-2015 challenge during INTERSPEECH 2015, SSD system designed using Cochlear Filter Cepstral Coefficient-Instantaneous Frequency (CFCCIF) feature set was the relatively best performing system. The CFCCIF feature set composed of the information obtained from the magnitude envelope derived using cochlear filterbank and instantaneous frequency (IF) which is derived from Hilbert transform-based approach. However, Hilbert transform-based estimation requires a speech segment of 10–30 ms and thus, it limits time resolution of IF estimation and hence, defeats the key objective of IF estimation to be able to fit the frequency of a sinusoid (corresponding to a monocomponent signal) locally and almost instantaneously. Energy Separation Algorithm (ESA) is known to accurately estimate the modulation patterns due to their relatively low computational complexity, high time resolution, and instantaneously adapting nature. To that effect, we exploit the ESA instead of Hilbert transform to estimate the IFs of the subband filtered signal using cochlear filterbank. The significant improvement in performance, in particular, relative reduction of 51.21% and 46.87% in EER is observed on development and evaluation subsets, respectively, for CFCCIF-ESA feature set over its CFCCIF counterpart, using Gaussian Mixture Model (GMM)-based classifier. This improvement in the performance indicates that the IFs estimated using ESA-based approach are able to efficiently capture the artefacts produced in the instantaneous phase by the SS- and VC-based spoof signals. Furthermore, experiments are also performed with convolutional neural network (CNN) classifier, which further enhances the performance.

Amplitude and Frequency Modulation-based features for detection of replay Spoof Speech

Replay attack poses a great threat to the Automatic Speaker Verification (ASV) system. This paper introduces Amplitude Modulation and Frequency Modulation-based features for replay Spoof Speech Detection (SSD) task. In this context, we propose Instantaneous Amplitude (IA) and Instantaneous Frequency (IF) features using Energy Separation Algorithm (ESA). The speech signal is passed through bandpass (subband) filters to obtain narrowband components because speech is a combination of several monocomponent signals. To obtain a narrowband filtered signal, we have used linearly-spaced Butterworth and Gabor filterbanks. The instantaneous modulations helps to understand the local characteristics of a non-stationary signal. These IA and IF components are able to capture the information present in a slowly-varying amplitude envelope and fast-varying frequency. The slow-varying temporal modulations for replay speech have the distorted amplitude envelope, and the fast-varying temporal modulation do not preserve the harmonic structure compared to the natural speech signal. For replay speech signal, the intermediate device characteristics and acoustic environment distorts the spectral energy compared to the natural speech energy. Experiments were performed on the ASVspoof 2017 challenge version 2.0 database with Gaussian Mixture Model (GMM) as a classifier. When ESA-IACC and ESA-IFCC feature sets are fused with Constant Q Cepstral Coefficients (CQCC) feature set at the score-level, the % EER further reduces to 11.93% and 10.12%, respectively, on the evaluation set. In addition, for evaluation set, we have also studied the performance of proposed feature sets on different Replay Configurations (RC), namely, acoustic environments, playback, and recording devices. For all the levels of threat conditions (i.e., low, medium, and high-level) to the ASV system, the proposed feature sets performed better compared to the existing state-of-the-art feature sets. In addition to the ASVspoof 2017 Challenge database, we also performed experiments on other spoofing databases, namely, BTAS 2016, ASVspoof 2019 Challenge database, and Real PA of ASVspoof 2019 Challenge database. For all the spoofing databases used in this study, the proposed ESA-based feature sets perform significantly better than the other feature sets.

Detection of replay spoof speech using teager energy feature cues

The vulnerability of Automatic Speaker Verification (ASV) systems to spoofing or presentation attacks is still an open security issue. In this context, replay spoofing attacks pose a great threat to an ASV system since they can be easily performed (using a playback device, and without needing any technical skill). In this paper, we analyze replay speech signals in terms of reverberation that may occur during recording of the speech signal. Such reverberation introduces delay and changes in amplitude, producing close copies of speech signals, which significantly influences the replay components. To that effect, we propose to exploit the capabilities of the Teager Energy Operator (TEO) to compute a running estimate of subband energies for replay vs. genuine signals. We have used a linearly-spaced Gabor filterbank to obtain a narrowband filtered signal. The TEO has the ability to track the instantaneous changes of a signal. Experiments are performed on the ASVspoof 2017 Challenge version 2.0 database using a Gaussian Mixture Model (GMM) as pattern classifier. Furthermore, we compared our results with state-of-the-art feature sets, namely, Constant Q Cepstral Coefficients (CQCC), Linear Frequency Cepstral Coefficients (LFCC), Mel Frequency Cepstral Coefficients (MFCC), and used their score-level fusion with the proposed feature sets, i.e., Teager Energy Cepstral Coefficients (TECC), in order to obtain possible complementary information that further reduces the Equal Error Rate (EER). Relatively low EERs are obtained with score-level fusion of CQCC, MFCC, LFCC, and TECC feature sets, resulted in 6.68% and 10.45% on development and evaluation sets, respectively. Moreover, for the evaluation dataset, we also studied the performance of the TECC feature set on different Replay Configurations (RC), namely, for acoustic environments, playback, and recording devices. For all the levels of threat conditions (i.e., low, medium, and high-level) to an ASV system, the proposed feature set performed better compared to existing state-of-the-art feature sets. In addition to the ASVspoof 2017 Challenge database, we also performed experiments on other spoofing databases, namely, the ASVspoof 2015 Challenge, BTAS 2016, and ASVspoof 2019 Challenge databases. For all the spoofing databases used in this study, the proposed TECC feature set perform significantly better than the other feature sets.

Combination of Amplitude and Frequency Modulation Features for Presentation Attack Detection

In this paper, we propose the combination of Amplitude Modulation and Frequency Modulation (AM-FM) features for replay Spoof Speech Detection (SSD) task. The AM components are known to be affected by noise (in this case, due to replay mechanism). In particular, we exploit this damage in AM component to corresponding Instantaneous Frequency (IF) for SSD task. Thus, the novelty of proposed Amplitude Weighted Frequency Cepstral Coefficients (AWFCC) feature set lies in using frequency components along with squared weighted amplitude components that are degraded due to replay noise. The AWFCC feature set contains the information of both AM and FM components together and hence, gave discriminatory information in the spectral characteristics. The experiments were performed on publicly available ASVspoof 2017 challenge version 1.0 and 2.0 databases using AWFCC feature set. We have compared results of proposed feature set with the other state-of-the-art feature set, such as Constant Q Cepstral Coefficients (CQCC), Linear Frequency Cepstral Coefficients (LFCC), Mel Frequency Cepstral Coefficients (MFCC) and using a simple Gaussian Mixture Model (GMM) classifier. The individual performance of AWFCC feature set obtained lower % EER than the other feature sets on both version 1.0 and 2.0 databases. Furthermore, we used score-level fusion in order to obtain the possible complementary information of two feature sets to reduce the % EER further. To that effect, the score-level fusion of CQCC and AWFCC feature sets gave 5.75 % and 10.42 % EER on development and evaluation sets, respectively, of ASVspoof 2017 version 2.0 database. Moreover, for evaluation dataset, we have also studied the performance of proposed feature set on different Replay Configurations (RC), namely, acoustic environments, playback, and recording devices. For all the levels of threat conditions (i.e., low, medium, and high) to the ASV system, the proposed feature set performed better compared to the existing state-of-the-art feature sets.

Spoofing Detection in Automatic Speaker Verification Systems Using DNN Classifiers and Dynamic Acoustic Features.

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.

Significance of Source–Filter Interaction for Classification of Natural vs. Spoofed Speech

Countermeasures used to detect synthetic and voice-converted spoofed speech are usually based on excitation source or system features. However, in the natural speech production mechanism, there exists nonlinear source–filter (S–F) interaction as well. This interaction is an attribute of natural speech and is rarely present in synthetic or voice-converted speech. Therefore, we propose features based on the S–F interaction for a spoofed speech detection (SSD) task. To that effect, we estimate the voice excitation source (i.e., differenced glottal flow waveform, $\dot{g} (t))$ and model it using the well-known Liljencrants–Fant model to get coarse structure, $g_{c}(t)$ . The residue or difference, $g_{r}(t)$ , between $\dot{g} (t)$ and $g_{c}(t)$ is known to capture the nonlinear S–F interaction. In the time domain, the $L^{2}$ norm of $g_{r}(t)$ in the closed, open, and return phases of the glottis are considered as features. In the frequency domain, the Mel representation of $g_{r}(t)$ showed significant contribution in the SSD task. The proposed features are evaluated on the first ASVspoof 2015 challenge database using a Gaussian mixture model based classification system. On the evaluation set, for vocoder-based spoofs (i.e., S1 – S9 ), the score-level fusion of residual energy features, Mel representation of the residual signal, and Mel frequency cepstral coefficients (MFCC) features gave an equal error rate (EER) of 0.017%, which is much less than the 0.319% obtained with MFCC alone. Furthermore, the residues of the spectrogram (as well as the Mel-warped spectrogram) of estimated $\dot{g} (t)$ and $g_{c}(t)$ are also explored as features for the SSD task. The features are evaluated for robustness in the presence of additive white, babble, and car noise at various signal-to-noise-ratio levels on the ASVspoof 2015 database and for channel mismatch condition on the Blizzard Challenge 2012 dataset. For both cases, the proposed features gave significantly less EER than that obtained by MFCC on the evaluation set.

Cochlear Filter and Instantaneous Frequency Based Features for Spoofed Speech Detection

Vulnerability of voice biometrics systems to spoofing attacks by synthetic speech (SS) and voice converted (VC) speech has arose the need of standalone spoofed speech detection (SSD) systems. This paper is an extension of our previously proposed features (used in relatively best performing SSD system) at the first ASVspoof 2015 challenge held at INTERSPEECH 2015. For the challenge, the authors proposed novel features based on cochlear filter cepstral coefficients (CFCC) and instantaneous frequency (IF), i.e., CFCCIF. The basic motivation behind this is that human ear processes speech in subbands. The envelope of each subband and its IF is important for perception of speech. In addition, the transient information also adds to the perceptual information that is captured. We observed that subband energy variations across CFCCIF when estimated by symmetric difference (CFCCIFS) gave better discriminative properties than CFCCIF. The features are extracted at frame level and the Gaussian mixture model based classification system was used. Experiments were conducted on ASVspoof 2015 challenge database with MFCC, CFCC, CFCCIF, and CFCCIFS features. On the evaluation dataset, after score-level fusion with MFCC, the CFCCIFS features gave an overall equal error rate (EER) of 1.45% as compared to 1.87% and 1.61% with CFCCIF and CFCC, respectively. In addition to detecting the known and unknown attacks, intensive experiments have been conducted to study the effectiveness of the features under the condition that either only SS or only VC speech is available for training. It was observed that when only VC speech is used in training, both VC, as well as SS, can be detected. However, when only SS is used in training, VC speech was not detected. In general, amongst vocoder-based spoofs, it was observed that VC speech is relatively difficult to detect than SS by the SSD system. However, vocoder-independent SS was toughest with highest EER (i.e., > 10%).