Andreas Nautsch

Robustness of Quality-based Score Calibration of Speaker Recognition Systems with respect to low-SNR and short-duration conditions.

Degraded signal quality and incomplete voice probes have severe effects on the performance of a speaker recognition system. Unified audio characteristics (UACs) have been proposed to quantify multi-condition signal degradation effects into posterior probabilities of quality classes. In previous work, we showed that UAC-based quality vectors (q-vectors) are efficient at the score-normalization stage. Hence, we motivate qvector based calibration by using functions of quality estimates (FQEs). In this work, we examine the robustness of calibration approaches to low-SNR and short-duration conditions utilizing measured and estimated quality indicators. Thereby, comparisons are drawn to quality measure functions (QMFs) employing oracle SNRs and sample duration. In the robustness study, low-SNR and short-duration conditions are excluded from calibration training. The present analysis provides insights on the behavior of calibration schemes in combined conditions of high signal degradation and short segment duration regarding accurate approximation of idealized calibration. We seek calibration methods in order to parsimonious preserve robustness against unseen data. A separate analysis is provided on durationand noise-only scenarios as well as on combined duration and noise scenarios. QMFs and FQE reduce Cmc costs down to 5 – 6% of conventional calibration schemes if all conditions are known, and to 10 – 12% in the presence of unseen conditions.

Biometric Systems under Morphing Attacks: Assessment of Morphing Techniques and Vulnerability Reporting

With the widespread deployment of biometric recognition systems, the interest in attacking these systems is increasing. One of the easiest ways to circumvent a biometric recognition system are so-called presentation attacks, in which artefacts are presented to the sensor to either impersonate another subject or avoid being recognised. In the recent past, the vulnerabilities of biometric systems to so-called morphing attacks have been unveiled. In such attacks, biometric samples of multiple subjects are merged in the signal or feature domain, in order to allow a successful verification of all contributing subjects against the morphed identity. Being a recent area of research, there is to date no standardised manner to evaluate the vulnerability of biometric systems to these attacks. Hence, it is not yet possible to establish a common benchmark between different morph detection algorithms. In this paper, we tackle this issue proposing new metrics for vulnerability reporting, which build upon our joint experience in researching this challenging attack scenario. In addition, recommendations on the assessment of morphing techniques and morphing detection metrics are given.

Multi-Bit Allocation: Preparing Voice Biometrics for Template Protection.

Technologies of biometric template protection grant a significant improvement in data privacy and increase the likelihood that the general public will effectively consent in the biometric system usage. Focusing on speaker recognition this area of research is still in its infancy. Previously proposed voice biometric template protection schemes fail in guaranteeing required properties of irreversibility and unlinkability without significantly degrading the recognition accuracy. A crucial step for accurate and secure template protection schemes is the feature type transformation which might be required to binarize extracted feature vectors. In this paper we introduce a binarization technique for voice biometric features called multi-bit allocation. The proposed scheme, which builds upon a GMM-UBM-based speaker recogniton system, is designed to extract discriminative compact binary feature vectors to be applied in a voice biometric template protection scheme. In a preliminary experimental study we show that the resulting binary representation causes only a marginal decrease in biometric performance compared to the baseline system, confirming the soundness and aplicability of the proposed scheme.

Bridging Gaps: An application of feature warping to online signature verification

The use of (online) signatures for the purpose of verifying a subjects identity is highly accepted within society and perceived as a noninvasive and nonthreatening biometric characteristic by most users. However, signature biometrics is typically characterized by a high intra-class variability, being influenced by several physical and emotional conditions, i.e. identity verification based on online signature biometrics represents an extremely challenging task. Online signature verification systems mainly utilize time-discrete signal processing techniques for biometric signature authorship verification. The vast majority of state-of-the-art approaches to online signature verification construct subject-specific probabilistic models during feature extraction, e.g. Gaussian Mixture Models (GMMs). Focusing on the construction of these models feature normalization turns out to be vital in order to achieve robustness against noise. In this work we propose the very first application of a feature normalization technique, referred to as Feature Warping (FW), which is well-established within the speaker recognition community, to a GMM-based online signature verification system. Experimental evaluations, which are carried out on the MCYT signature corpus, demonstrate that the presented adaptation of FW significantly improves the biometric performance of the underlying online signature verification system, achieving relative gains of approximately 47% in terms of equal error rates.

Unit-Selection Attack Detection Based on Unfiltered Frequency-Domain Features.

Modern text-to-speech algorithms pose a vital threat to the security of speaker identification and verification (SIV) systems, in terms of subversive usage, i.e. generating presentation attacks. In order to distinguish between presentation attacks and bona fide authentication attempts, presentation attack detection (PAD) subsystems are of utmost importance. Until now, the vast majority of introduced spoofing countermeasures rely on speech production and perception based features. In this paper, we utilize the complete frequency band without further filterbank processing in order to detect non-smooth transitions in the full and high frequency domain caused by unit-selection attacks. For the purpose of especially detecting unit selection attacks, the applicability of Fast Fourier Transformation (FFT) and Discrete Wavelet Transformation (DWT) is examined regarding non-smooth transitions in the full and high frequency domain, excluding filter-bank analyses. Gaussian Mixture Model (GMM) and Support Vector Machine (SVM) classifiers are trained on the German Speech Data Corpus (GSDC) and validated on the standard ASVspoof 2015 corpus resulting in EERs of 7.1% and 11.7%, respectively. Despite language and data shifts, the proposed unit-selection PAD scheme achieves promising biometric performance and hence, introduces a new direction to voice PAD.

Entropy analysis of i-vector feature spaces in duration-sensitive speaker recognition

The vast majority of speaker recognition cross-entropy evaluations are focused on score domain. By examining the generalized relative distance between genuine and impostor sub-spaces, biometric characteristics become comparable to other authentication approaches. In this paper we demonstrate that the i-vector feature spaces biometric information measured by relative entropy is comparable to e.g., knowledge-based mechanisms or face recognition. Examining NIST SRE 2004-2010 corpora, short samples of e.g, 5 seconds duration, comprise already 127 bits in a text-independent scenario. Further, the vast majority of short samples does not fall below 50% of the biometric information of samples having a duration of more than 40 seconds. The generalized i-vector feature space entropy of long samples corresponds to 182.1 bits, and the highest lower entropy bound of a subject was observed at 471.6 bits.

Deep Quality-Informed Score Normalization for Privacy-Friendly Speaker Recognition in Unconstrained Environments

In scenarios that are ambitious to protect sensitive data in compliance with privacy regulations, conventional score normalization utilizing large proportions of speaker cohort data is not feasible for existing technology, since the entire cohort data would need to be stored on each mobile device. Hence, in this work we motivate score normalization utilizing deep neural networks. Considering unconstrained environments, a quality-informed scheme is proposed, normalizing scores depending on sample quality estimates in terms of completeness and signal degradation by noise. Utilizing the conventional PLDA score, comparison i-vectors, and corresponding quality vectors, we aim at mimicking cohort based score normalization optimizing the minCllr discrimination criterion. Examining the I4U data sets for the 2012 NIST SRE, an 8.7% relative gain is yielded in a pooled 55-condition scenario with a corresponding condition-averaged relative gain of 6.2% in terms of minCllr. Robustness analyses towards sensitivity regarding unseen conditions are conducted, i.e. when conditions comprising lower quality samples are not available during training.

Decision Robustness of Voice Activity Segmentation in Unconstrained Mobile Speaker Recognition Environments

Voice activity detection (VAD) is an essential segmentation process in speaker recognition systems, seperating speech and non-speech segments of voice samples. In speaker recognition, references are modelled purely by concerning speech segments. Different VAD segmentations lead to variations in biometric models, and consequently in system performance. Thus, VAD decisions need to be robust among different conditions. In this paper, the decision robustness of different VAD algorithms is examined on mobile data by simulating different environmental noise conditions for which we propose a Hamming distance based analysis. By examining speech and speaker recognition based VADs, we further propose to extend a well- performing VAD algorithm, which is based on likelihood ratio comparison of speech to non-speech models, by including most dominant frequency component (MDFC) features for selection of model training segments. Thereby, more robust VAD decisions are conducted by 7%, while sustaining an average EER SNR-sensitivity of 0.76% per dB SNR.