Citation Context

... trained on the same genuine speech as the second classifier and the 9892 converted voice utterances in the development set respectively. All SVM classifiers are implemented using the LIBSVM5 library =-=[38]-=- and are tuned using only geunine speech or converted voices in the development set. 4.5. Results Here we report the performance of the baseline system, the effect on performance of spoofing attacks w...

Citation Context

...recognition. New contributions are two-fold. First, we introduce a new feature for spoofing detection based on the analysis of conventional speech parameterisations using Local Binary Patterns (LBPs) =-=[21]-=- and second, we present the first one-class classification approach to ASV spoofing detection. Experiments usinghal-00869893, version 1 - 4 Oct 2013 a state-of-the-art i-vector / probabilistic linear...

Citation Context

...on (MLSA) filters. They correspond to STRAIGHT Mel-cepstral coefficients and are driven by a mixed excitation signal and waveforms reconstructed using the pitch synchronous overlap add (PSOLA) method =-=[27]-=-. 2.3. Artificial signals Artificial signal attacks are based on the algorithm reported in [13]. It is based on a modification of the voice conversion algorithm presented in Section 2.1. Let S = {c1, ...

Citation Context

...hniques such as nuisance attribute projection (NAP) and Factor Analysis (FA). New to this contribution is the consideration of an i-vector system, the current state of the art in speaker verification =-=[31]-=-. The i-vector system uses FA to model session and speaker variability at the front-end by means of a so-called total variability matrix. The setup involves mixtures of 1024 Gaussian components and i-...

Citation Context

...rmed using the ALIZE toolkit [32] and the LIA-RAL framework [33]. Unwanted variability is handled through Probabilistic Linear Discriminant Analysis (PLDA) compensation [34] with length normalization =-=[35]-=-. 4.2. Protocols and metricshal-00869893, version 1 - 4 Oct 2013 Dataset NIST’05 (dev) NIST’06 (eval) Speakers 201 298 Client tests 984 1344 Impostor tests 9862 12648 Table 1. Number of target and im...

Citation Context

...g F0) along with their delta and acceleration coefficients. Acoustic spectral characteristics and duration probabilities are modeled using multispace distribution hidden semi-Markov models (MSD-HSMM) =-=[25]-=-. Speaker dependent excitation, spectral and duration models are adapted from corresponding independent models according to a speaker adaptation strategy referred to as constrained structural maximum ...

Citation Context

...and duration models are adapted from corresponding independent models according to a speaker adaptation strategy referred to as constrained structural maximum a posteriori linear regression (CSMAPLR) =-=[26]-=-. Finally, time domain signals are synthesized using a vocoder based on Mel-logarithmic spectrum approximation (MLSA) filters. They correspond to STRAIGHT Mel-cepstral coefficients and are driven by a...

Citation Context

...22, 23]. 2.2. Speech synthesis Spoofing attacks with speech synthesis were implemented using the hidden Markov model (HMM)-based Speech Synthesis System (HTS) 1 and the specific approach described in =-=[24]-=-. Parameterisations include STRAIGHT (Speech Transformation and Representation using Adaptive Interpolation of weiGHTed spectrum) features, Melcepstrum coefficients and the logarithm of the fundamenta...

Citation Context

...nd i-vector extraction is performed using the ALIZE toolkit [32] and the LIA-RAL framework [33]. Unwanted variability is handled through Probabilistic Linear Discriminant Analysis (PLDA) compensation =-=[34]-=- with length normalization [35]. 4.2. Protocols and metricshal-00869893, version 1 - 4 Oct 2013 Dataset NIST’05 (dev) NIST’06 (eval) Speakers 201 298 Client tests 984 1344 Impostor tests 9862 12648 T...

Citation Context

...on performance is then comparable to that expected by chance. Recently the research community has started to investigate spoofing actively. Although there is one notable exception in face recognition =-=[14]-=-, due to the novelty of such work there are currently no standard large-scale datasets, protocols or metrics for the evaluation of spoofing countermeasures. In consequence, it is still common practice...

Citation Context

...nger [2]. Automatic speaker verification (ASV) systems are also vulnerable to spoofing attacks with varying degrees of sophistication. Impersonation [3, 4], replayed speech [5, 6], synthesised speech =-=[7, 8]-=-, voice conversion [9–12] and artificial signals [13] have all been shown to provoke significant increases in the false acceptance rate of state-of-the-art ASV systems. More often than not, authentica...

Citation Context

...matrix. The setup involves mixtures of 1024 Gaussian components and i-vectors with 400 dimensions. The total variability matrix estimation and i-vector extraction is performed using the ALIZE toolkit =-=[32]-=- and the LIA-RAL framework [33]. Unwanted variability is handled through Probabilistic Linear Discriminant Analysis (PLDA) compensation [34] with length normalization [35]. 4.2. Protocols and metrics...

Citation Context

...asr. Hx(f) is estimated from xfil using an LPCC-to-LPC transformation and a time-domain signal is synthesized from converted frames with a standard overlap-add technique. Full details can be found in =-=[11, 22, 23]-=-. 2.2. Speech synthesis Spoofing attacks with speech synthesis were implemented using the hidden Markov model (HMM)-based Speech Synthesis System (HTS) 1 and the specific approach described in [24]. P...

Citation Context

...oofed with a fake, gummy finger [2]. Automatic speaker verification (ASV) systems are also vulnerable to spoofing attacks with varying degrees of sophistication. Impersonation [3, 4], replayed speech =-=[5, 6]-=-, synthesised speech [7, 8], voice conversion [9–12] and artificial signals [13] have all been shown to provoke significant increases in the false acceptance rate of state-of-the-art ASV systems. More...

Citation Context

...are also vulnerable to spoofing attacks with varying degrees of sophistication. Impersonation [3, 4], replayed speech [5, 6], synthesised speech [7, 8], voice conversion [9–12] and artificial signals =-=[13]-=- have all been shown to provoke significant increases in the false acceptance rate of state-of-the-art ASV systems. More often than not, authentication performance is then comparable to that expected ...

Citation Context

...rformance in detecting attacks with synthesized speech and artificial signals. No knowledge of these algorithms was used intentionally during development. 3.1. LBP features Based on our previous work =-=[28]-=-, we hypothetise that genuine speech can be distinguished from spoofed speech according to differences in the spectro-temporal ‘texture’. The motivation stems from the assumption that, while lower-lev...

Citation Context

...nition systems can be spoofed with a fake, gummy finger [2]. Automatic speaker verification (ASV) systems are also vulnerable to spoofing attacks with varying degrees of sophistication. Impersonation =-=[3, 4]-=-, replayed speech [5, 6], synthesised speech [7, 8], voice conversion [9–12] and artificial signals [13] have all been shown to provoke significant increases in the false acceptance rate of state-of-t...

Citation Context

...oofed with a fake, gummy finger [2]. Automatic speaker verification (ASV) systems are also vulnerable to spoofing attacks with varying degrees of sophistication. Impersonation [3, 4], replayed speech =-=[5, 6]-=-, synthesised speech [7, 8], voice conversion [9–12] and artificial signals [13] have all been shown to provoke significant increases in the false acceptance rate of state-of-the-art ASV systems. More...

Citation Context

... this approach to classification, among others. In all cases the proposed countermeasure is integrated with the ASV system as an independent post processing step, in equivalent fashion to the work in =-=[8, 17, 28]-=-. 4. Experimental work Here we report experiments to assess the performance of the new countermeasure using a state-of-the-art ASV system and public, standard datasets. 4.1. ASV systems In previous wo...

Citation Context

...nger [2]. Automatic speaker verification (ASV) systems are also vulnerable to spoofing attacks with varying degrees of sophistication. Impersonation [3, 4], replayed speech [5, 6], synthesised speech =-=[7, 8]-=-, voice conversion [9–12] and artificial signals [13] have all been shown to provoke significant increases in the false acceptance rate of state-of-the-art ASV systems. More often than not, authentica...

Citation Context

...ures of 1024 Gaussian components and i-vectors with 400 dimensions. The total variability matrix estimation and i-vector extraction is performed using the ALIZE toolkit [32] and the LIA-RAL framework =-=[33]-=-. Unwanted variability is handled through Probabilistic Linear Discriminant Analysis (PLDA) compensation [34] with length normalization [35]. 4.2. Protocols and metricshal-00869893, version 1 - 4 Oct...

Citation Context

...eat is common to all biometric modalities. For example, face recognition systems can be spoofed with a photograph [1], whereas fingerprint recognition systems can be spoofed with a fake, gummy finger =-=[2]-=-. Automatic speaker verification (ASV) systems are also vulnerable to spoofing attacks with varying degrees of sophistication. Impersonation [3, 4], replayed speech [5, 6], synthesised speech [7, 8], ...

Citation Context

...nition systems can be spoofed with a fake, gummy finger [2]. Automatic speaker verification (ASV) systems are also vulnerable to spoofing attacks with varying degrees of sophistication. Impersonation =-=[3, 4]-=-, replayed speech [5, 6], synthesised speech [7, 8], voice conversion [9–12] and artificial signals [13] have all been shown to provoke significant increases in the false acceptance rate of state-of-t...

Citation Context

...sis attacks were implemented using the voice cloning toolkit2 with a default configuration. We used standard speakerindependent models provided with the toolkit which were trained on the EMIME corpus =-=[37]-=-. Synthesized speech is generated using the transcripts of the original impostor utterances. While it is admittedly not representative of real scenarios, we assess countermeasure performance in a wors...

Citation Context

... known and then the performance of existing countermeasures in practical scenarios is unknown. Recent work shows the potential impact of using prior knowledge. For instance, de Freitas Pereira et al. =-=[15]-=- showed that state-of-the-art spoofing countermeasures for face recognition do not generalise well to forms of spoofing not considered during development. Similar behaviour can be expected in speaker ...

Citation Context

...n do not generalise well to forms of spoofing not considered during development. Similar behaviour can be expected in speaker recognition. Countermeasures based on phase [16–18] and prosodic features =-=[19, 20]-=- can be used very successfully to detect voice conversion and speech synthesis attacks. It is likely, however, that they will be overcome by the particular approach to voice conversion investigated in...

Citation Context

...pecific attack or spoofing indicator. The same technique may also be used to enhance robustness to a single, specific attack e.g. the combination of motion and texture analysis for face anti-spoofing =-=[30]-=-. Binary spoofing detectors are generally independent of the biometric system and typically trained using both genuine data (negative samples) and examples of spoofed data (positive samples). The main...

Citation Context

...can be used very successfully to detect voice conversion and speech synthesis attacks. It is likely, however, that they will be overcome by the particular approach to voice conversion investigated in =-=[11]-=- which modifies only the spectral slope of a speech utterance while retaining the original phase and pitch of the original, genuine speech signal. Spoofing thus remains very much an open problem. This...

Citation Context

...n do not generalise well to forms of spoofing not considered during development. Similar behaviour can be expected in speaker recognition. Countermeasures based on phase [16–18] and prosodic features =-=[19, 20]-=- can be used very successfully to detect voice conversion and speech synthesis attacks. It is likely, however, that they will be overcome by the particular approach to voice conversion investigated in...

Citation Context

...asr. Hx(f) is estimated from xfil using an LPCC-to-LPC transformation and a time-domain signal is synthesized from converted frames with a standard overlap-add technique. Full details can be found in =-=[11, 22, 23]-=-. 2.2. Speech synthesis Spoofing attacks with speech synthesis were implemented using the hidden Markov model (HMM)-based Speech Synthesis System (HTS) 1 and the specific approach described in [24]. P...

Citation Context

...s the biometric system is equipped with appropriate spoofing countermeasures, this threat is common to all biometric modalities. For example, face recognition systems can be spoofed with a photograph =-=[1]-=-, whereas fingerprint recognition systems can be spoofed with a fake, gummy finger [2]. Automatic speaker verification (ASV) systems are also vulnerable to spoofing attacks with varying degrees of sop...