Performing speaker verification in the simultaneous presence of ageing progression and changing speech sample quality is an important, open problem. The issues of ageing and quality variation go hand in hand; the effect of ageing increases with time, while variations in quality are also more likely to be encountered as time passes. In this work we demonstrate the effect of ageing on speaker verification performance, and show the relationship between quality variation and verification score via a range of established quality measures. We employ a stacked classifier framework to combine the output of the baseline verification system with ageing information and quality measures. This new approach to long-term speaker verification allows for a multi-dimensional decision boundary that significantly improves upon the baseline performance. The proposed framework is evaluated on the Trinity College Dublin Speaker Ageing Database. Index Terms: speaker verification, ageing, stacked classifier

Age progression and other sources of variations such as head pose change have adverse effects on the performance of face verification systems. In this paper, we propose to manage the influence of both the age progression and head pose change on the adult face verification system by a general Q-stack age modeling technique, which uses the age and head pose as class-independent quality measures together with the scores from baseline classifiers, in order to obtain better verification performance. This allows for improved long-term class separation by introducing a 2D parameterized decision boundary in the scores-age space using a short-term enrollment model. This new method, based on the concept of classifier stacking with age- and head pose aware decision boundary compares favorably with the conventional face verification approach, which uses age- and head-pose-independent decision threshold calculated only in the score space at the time of enrollment. In this paper, we also show the advantages of user-specific approach for the face verification task over user-independent approach. The proposed approach is evaluated on the MORPH database.

Palm vein feature extraction from near infrared images is a challenging problem in hand pattern recognition. In this paper, a promising new approach based on local texture patterns is proposed. First, operators and histograms of multi-scale Local Binary Patterns (LBPs) are investigated in order to identify new efficient descriptors for palm vein patterns. Novel higher-order local pattern descriptors based on Local Derivative Pattern (LDP) histograms are then investigated for palm vein description. Both feature extraction methods are compared and evaluated in the framework of verification and identification tasks. Extensive experiments on CASIA Multi-Spectral Palmprint Image Database V1.0 (CASIA database) identify the LBP and LDP descriptors which are better adapted to palm vein texture. Tests on the CASIA datasets also show that the best adapted LDP descriptors consistently outperform their LBP counterparts in both palm vein verification and identification.

The problem of time validity of biometric models has received only a marginal attention from researchers. Actual and up-to-date at the time of their creation, extracted features and models relevant to a person's face may eventually become outdated, leading to a failure in the face identification task. If physical characteristics of the individual change over time, their classification model has to be updated. In this paper we present a mutli-classifier A-stack scheme, which is based on the concept of classifier stacking and makes use of the age information and scores of multiple baseline classifiers, in order to improve the identification performance during age progression. Our experiments on the MORPH database show that the use of the proposed multi-classifier stacking fusion allows for improving the identification accuracy as opposed to the baseline classifier and single-classifier A-stack method.

The influence of age progression on the performance of multi-classifier face verification systems is a challenging and largely open research problem that deserves more and more attention. In this paper, we propose to manage the aging influence on the adult face verification system by a multi-classifier Q-stack age modeling technique, which uses the age as a class-independent metadata quality measure together with scores from baseline classifiers, combining global and local patterns, in order to obtain better recognition rates. This allows for improved long-term class separation by introducing a 2D parameterized decision boundary in the scores-age space using a short-term enrollment model. This new method, based on the concept of classifier stacking and age-dependent decision boundary, compares favorably with the conventional face verification approach, which uses age-independent decision threshold calculated only in the score space at the time of enrollment. The proposed approach is evaluated on the MORPH database.

A new multimodal biometric database designed and acquired within the framework of the European BioSecure Network of Excellence is presented. It is comprised of more than 600 individuals acquired simultaneously in three scenarios: 1) over the Internet, 2) in an office environment with desktop PC, and 3) in indoor/outdoor environments with mobile portable hardware. The three scenarios include a common part of audio/video data. Also, signature and fingerprint data have been acquired both with desktop PC and mobile portable hardware. Additionally, hand and iris data were acquired in the second scenario using desktop PC. Acquisition has been conducted by 11 European institutions. Additional features of the BioSecure Multimodal Database (BMDB) are: two acquisition sessions, several sensors in certain modalities, balanced gender and age distributions, multimodal realistic scenarios with simple and quick tasks per modality, cross-European diversity, availability of demographic data, and compatibility with other multimodal databases. The novel acquisition conditions of the BMDB allow us to perform new challenging research and evaluation of either monomodal or multimodal biometric systems, as in the recent BioSecure Multimodal Evaluation campaign. A description of this campaign including baseline results of individual modalities from the new database is also given. The database is expected to be available for research purposes through the BioSecure Association during 2008.

The impact of feature correlations on class separation has received limited attention from researchers. Previous reports treat the problem from the viewpoint of multi-classifier fusion and are partially inconsistent in their conclusions. In this paper we show that these ambiguities are the result of incompatible basic assumptions, and that the conclusions from prior art hold only for specific configurations of class-conditional distributions. We show that the impact of feature correlations on class separation between two bivariate normal distributions can be positive or negative, and that it can only be gauged in the context of the parameters of involved marginals. The findings reported in this paper are of importance for the practice of feature extraction, feature selection, and in multi-classifier fusion.

In pattern recognition, the need to quantify the quality of a classifier's output has gained importance in the past years. Speaker verification is no exception. This paper presents a probabilistic reliability framework incorporating signal-domain information into the confidence estimation and contrasts this method with classical approaches to estimating the confidence in a given speaker verification classifier output. We show that the method proposed can deal with adverse acoustic conditions for a wide range of signal-to-noise ratios, does not depend on a Gaussian assumption for impostor and client score distributions, and presents benefits in terms of scalability and interpretability of the measure. We contrast reliability and confidence approaches, and evaluate performance on a degraded version of the 295-users XM2VTS database

In this paper, a probabilistic measure for reliability of speaker verification under noisy acoustic conditions is proposed. A Bayesian network is used to estimate a probability for verification errors, given the GMM-based speaker verification system output and additional information about the level of acoustic noise. In particular, the log-likelihood ratio and a signal-to-noise related feature are used to account for the adverse acoustic conditions. The probabilistic measure is subsequently employed in governing a repair sequence of trials for acquiring additional speech presentations which are less likely to lead to unreliable verification. The potential of the proposed method is tested through cross-validation experiments. Finally, the benefits of the repair sequence in terms of verification accuracy is evaluated on a noisy environment speaker verification task.

In this paper we propose a methodology for selecting the most discriminative features in a set for online signature verification. We expose the difference in the definition of class between signature verification and other pattern recognition tasks, and extend the classical Fisher ratio to make it more robust to the small sample sizes typically found when dealing with global features and client enrollment time constraints for signature verification systems. We apply our methodology to global and local features extracted from a 50-users database, and find that our criterion agrees better with classifier error rates for local features than for global features. We discuss the possibility of performing feature selection without having forgery data available.

The main task of a voice-enabled tour-guide robot in mass exhibition setting is to engage visitors in dialogue and provide as much exhibit information as possible in a limited time. In managing such a dialogue, extracting the user (visitor) goal or intention at each dialogue state is the key issue. In mass exhibition conditions uncooperative visitors and speech recognition limitations in noisy acoustic conditions may jeopardize user goal identification. In this paper, we introduce the use of sequential dialogue repair techniques, exploiting the inherent multimodality of the tour-guide robot, in order to reduce the risk of the resulting communication failures. Bayesian networks fusing acoustic and non-acoustic modalities during user goal identification serve as input to graphical models known as decision networks. Decision networks allow the definition of dialogue repair sequences as actions, and provide a decision-theoretic utility-based strategy for selecting actions. The benefits of the proposed repair strategies are demonstrated through experiments with the dialogue system of RoboX, a tour-guide robot successfully deployed at the Swiss National Exhibition (Expo. 02).

The aim of this paper is to reduce the effect of mismatch in recording conditions due to the transmission channel and recording device, using conditional dependencies of prosodic and spectral envelope features. The developed system is based on a Bayesian network framework which combines statistical models of the pitch and spectral envelope features. This approach is applied to forensic automatic speaker recognition, where mismatched recording conditions pose a serious problem to the accurate estimation of the strength of voice evidence. The method is evaluated using a forensic speaker recognition database that contains three different recording conditions typical to forensic tasks. The performance of the system is evaluated using both speaker verification as well as forensic speaker recognition measures.

In forensic speaker recognition, the strength of evidence is estimated using the likelihood ratio, which is the relative probability of observing the evidence, given the hypothesis that the suspect is the source of the questioned recording and the hypothesis that anyone else in a relevant potential population is its source. In order to calculate the likelihood ratio we use two approaches; one, directly using the likelihoods returned by the Gaussian Mixture Models (GMMs), and the other by modeling the distributions of these likelihood scores and then deriving the likelihood ratio on the basis of these score distributions. The former approach is used implicitly in speaker verification systems, although in forensic speaker recognition, the latter is preferred as it does not depend on the automatic speaker recognition technique used. However, both these methods have their advantages and disadvantages. In this paper, we propose statistical representations in order to evaluate the strength of evidence in each of these two methods.

The goal of this paper is to establish a robust methodology for forensic automatic speaker recognition (FASR) based on sound statistical and probabilistic methods, and validated using databases recorded in real-life conditions. The interpretation of recorded speech as evidence in the forensic context presents particular challenges. The means proposed for dealing with them is through Bayesian inference and corpus based methodology. A probabilistic model -- the odds form of Bayes' theorem and likelihood ratio -- seems to be an adequate tool for assisting forensic experts in the speaker recognition domain to interpret this evidence. In forensic speaker recognition, statistical modelling techniques are based on the distribution of various features pertaining to the suspect's speech and its comparison to the distribution of the same features in a reference population with respect to the questioned recording. In this paper, the state-of-the-art automatic, text-independent speaker recognition system, using Gaussian mixture model (GMM), is adapted to the Bayesian interpretation (BI) framework to estimate the within-source variability of the suspected speaker and the between-sources variability, given the questioned recording. This doublestatistical approach (BI-GMM) gives an adequate solution for the interpretation of the recorded speech as evidence in the judicial process.

In this paper, we propose a method based on Bayesian networks for interpretation of multimodal signals used in the spoken dialogue between a tour-guide robot and visitors in mass exhibition conditions. We report on experiments interpreting speech and laser scanner signals in the dialogue management system of the autonomous tour-guide robot RoboX, successfully deployed at the Swiss National Exhibition (Expo.02). A correct interpretation of a user s (visitor s) goal or intention at each dialogue state is a key issue for successful voice-enabled communication between tour-guide robots and visitors. To infer the visitors goal under the uncertainty intrinsic to these two modalities, we introduce Bayesian networks for combining noisy speech recognition with data from a laser scanner, which is independent of acoustic noise. Experiments with real data, collected during the operation of RoboX at Expo.02 demonstrate the effectiveness of the approach.

In this paper, we propose a Bayesian network framework for managing interactivity between a tour-guide robot and visitors in mass exhibition conditions, through robust interpretation of multi-modal signals. We report on methods and experiments interpreting speech and laser scanner signals in the spoken dialogue management system of the autonomous tour-guide robot RoboX, successfully deployed at the Swiss National Exhibition (Expo.02). A correct interpretation of a user's (visitor's) goal or intention at each dialogue state is a key issue for successful speech-based interaction in voice-enabled communication between robots and visitors. We introduce a Bayesian network approach for combining noisy speech recognition results with noise-independent data from a laser scanner, in order to infer the visitors' goal under the uncertainty intrinsic to these two modalities. We demonstrate the effectiveness of the approach by simulation based on real observations during experiments with the tour-guide robot RoboX at Expo.02.

This paper considers design methodologies in order to develop voice-enabled interfaces for tour-guide robots to be deployed at the Robotics Exposition of the Swiss National Exhibition (Expo.02). Human-robot voice communication presents new challenges for the design of fully autonomous mobile robots, in that interactivity must be robot-initiated in conversation and within a dynamic adverse environment. We approached these general problems for a voice enabled interface, tailored to limited computational resources of one on-board processor, when integrating smart speech signal acquisition, automatic speech recognition and synthesis, as well as dialogue system into the multi-modal, multisensor interface for the Expo tour-guide robot. We also focus on particular issues that need to be addressed in voice-based interaction when planning specific tasks and research experiments for Expo.02 where tour-guide robots will interact with hundred of thousands of visitors for six months, seven days a week, ten hours per day.

Keywords: LTS1 Reference LTS-CONF-2001-011 Record created on 2006-06-14, modified on 2016-08-08

Keywords: LTS1 Reference LTS-CONF-2001-027 Record created on 2006-06-14, modified on 2017-01-14

Keywords: LTS1 Reference LTS-CONF-2001-039 Record created on 2006-06-14, modified on 2016-08-08

This paper addresses the problem of robust speech recognition in noisy conditions in the framework of hidden Markov models (HMMs) and missing feature techniques. It presents a new statistical approach to detection and estimation of unreliable features based on a probabilistic measure and Gaussian mixture model (GMM). In the estimation process, the GMM is compensated using parameters of the statistical model of additive background noise. The GMM means are used to replace the unreliable features. The GMM based technique is less complex than the corresponding HMM based estimation and gives similar improvement in the recognition performance. Once unreliable features are replaced by the estimated clean speech features, the entire set of spectral features can be transformed to the other feature domain characterized by higher baseline recognition rate (e.g. MFCCs) for final recognition using continuous density hidden Markov models (CDHMMs) with diagonal covariance matrices.

Relations between non-adjacent parts of an utterance are commonly regarded as an important source of information for speech recognition. However, they have not been very much used in speech recognition systems. In this paper, we include this information by joint distributions of pairs of phones occurring in the same utterance. In addition to relations between acoustic events, we also have incorporated relations between spectral and prosodically oriented information, such as phone duration, position in utterance and funda-mental frequency. Preliminary recognition results on N-best rescoring show 10% word error reduction compared to a baseline Viterbi decoder.

This paper describes a unifying framework for both formant tracking and speech synthesis using Hidden Markov Models (HMM). The feature vector in the HMM is composed by the first three formant frequencies, their bandwidths and their delta with time. Speech is synthesized by generating the most likely sequence of feature vectors from a HMM, trained with a set of sentences from a given speaker. Higher formant tracking accuracy can be achieved by finding the most likely formant track given a distribution of the formants of every sound. This data-driven formant synthesizer bridges the gaps between rulebased formant synthesizers and concatenative synthesizers by synthesizing speech that is both smooth and resembles the speaker in the training data.

In the framework of Gaussian mixture models (GMMs), we present a new approach towards robust automatic speaker verification (SV) in adverse conditions. This new and simple approach is based on the combination of speech enhancement using traditional spectral subtraction, and missing feature compensation to dynamically modify the probability computations performed in GMM recognizers. The identity of the spectral features missing due to noise masking is provided by the spectral subtraction algorithm. Previous works have demonstrated that the missing feature modeling method succeeds in speech recognition with some artificially generated interruptions, filtering and noise. We show that this method also improves noise compensation techniques used for speaker verification in more realistic conditions.

Wiener filter based postfiltering has shown its usefulness in microphone array speech enhancement systems. In our earlier work, we developed a postfilter in the wavelet domain where better performance has been obtained compared to the algorithms developed in the Fourier domain. Furthermore, considerable computational savings are provided thanks to the multi-resolution and multi-rate analysis. This contribution shows that the coherence function, calculated between the beamforming output signal and the reference microphone output signal using the wavelet transform, provides a relevant and exploitable information for further noise suppression. Thus, a nonlinear coherence filtering and a linear Wiener filtering are combined in the wavelet transform domain to improve the performance of the Wiener filter based postfilter, especially during pauses. Evaluations of the new algorithm confirm that the speech quality is indeed improved with significantly reduced distortions. Finally, the results of the objective measures are presented.

This paper presents a new wide-band speech coding system based on a fast wavelet packet transform algorithm as well as a formulation of temporal and spectral psychoacoustic models of masking. The proposed FFT-like overlapped block orthogonal transform allows us to approximate the auditory critical band decomposition in an efficient manner, which is a major advantage over previous approaches that used uniform filter banks. As a result of such a decomposition, the perceptually tuned time-frequency structure of the original speech signal is preserved. This allows us to make use of the temporal and spectral properties of the human auditory system to decrease the average bit rate of the encoder, while perceptually hiding the quantization error.

This paper addresses the problem of merging speech enhancement and coding in the context of an auditory modeling. The noisy signal is first processed by a fast wavelet packet transform algorithm to obtain an auditory spectrum, from which a rough masking model is estimated. Then, this model is used to refine a subtractive-type enhancement algorithm. The enhanced speech coefficients are then encoded in the same time-frequency transform domain using masking threshold constraints for quantization noise. The advantage of the proposed method is that both enhancement and coding are performed with the transform coefficients, without making use of the additional FFT processing.

Keywords: LTS1 Reference LTS-CONF-1997-045 Record created on 2006-06-14, modified on 2016-08-08

In this paper we present orthogonal overlapped block transforms as a frame synchronized signal analysis tool with the capability of arbitrary multiresolution time-spectral decomposition of speech signals.Our prime interest is in the representation of nonstationary discrete-time signals in terms of wavelet packets, and we concentrate on their fast transform algorithms.Wavelet packet representations provide a local time-spectral description which reveals the nonstationary nature of a signal.They allow the speech signal to be accurately parameterised for such applications as speech and speaker recognition, where a front-end is responsible for the frame synchronized feature extraction.In this case the fast overlapped block transform algorithms represent an elegant and ecient solution to the implementation of wavelet packet transforms, since their FFT-like lattice block structure provides all possible multiresolution timespectral coecients.The frame synchronization is also preserved in subbands which allows a new subband-based approach for speech recognition.

Orthogonal wavelet packet transforms, presented recently as powerful tools for nonstationary signal processing, suffered from the lack of efficient and time-domain transparent algorithms for their implementation. We present an overlapped block lattice structure that allows fast short-time wavelet packet transform algorithms to be built in an equivalent manner to other short-time transforms, such as the short-time Fourier transform (STFT) based on the fast Fourier transform (FFT) algorithms. The outputs obtained with the overlapped block lattice are equivalent to those achieved by the well known tree-structured filter bank approach, though the implementation algorithms are different.

Keywords: LTS1 Reference LTS-CONF-1995-037 Record created on 2006-06-14, modified on 2016-08-08

Keywords: LTS1 Reference LTS-CONF-1995-038 Record created on 2006-06-14, modified on 2016-08-08

Keywords: LTS1 Reference LTS-CONF-1993-052 Record created on 2006-06-14, modified on 2016-08-08

Spectral analysis/synthesis ideas that are common for orthogonal transforms, multichannel and multirate filtering, and wavelet transforms are discussed and generalized. Some recently developed unconventional applications of the butterfly orthogonal decomposition technique are reviewed and its usefulness in developing efficient multiresolution digital signal processing systems is discussed. A generalized multirate filtering structure is developed that is based on fast algorithms of orthogonal transforms and their orthogonal subtransforms. In particular the structural subband decomposition of a discrete signal in sequency and frequency spectral domains is given. A generalized butterfly tree structure with all-pass branches and arbitrary weighting constants as well as its multilevel filter application is discussed. Wavelet filter bank realizations appear as a subset of presented structures.>