Search

International audience; Speech signals can be corrupted by environmental noise as well as room reverberation which severely affects the speaker verification performance. In this paper, we propose to combine a multichannel pre-processing pipeline including filter-and-sum network (FaSnet), Rank-1 multichannel Wiener filter, and weighted prediction error as a front-end to speaker verification. Experimental evaluation shows that the pre-processing can improve the speaker verification performance as long as the enrollment files are processed similarly to the test data and that test and enrollment occur within similar SNR ranges. Our proposed pipeline is trained on synthetic data but generalizes to unseen, real recorded clips included in the VOiCES eval dataset and improves the speaker verification performance on all the noise conditions.

Gaussian-PLDA (G-PLDA) modeling for i-vector based speaker verification has proven to be competitive versus heavy-tailed PLDA (HT-PLDA) based on Student's t-distribution, when the latter is much more computationally expensive. However , its results are achieved using a length-normalization, which projects i-vectors on the non-linear and finite surface of a hypersphere. This paper investigates the limits of linear and Gaussian G-PLDA modeling when distribution of data is spherical. In particular, assumptions of homoscedasticity are questionable: the model assumes that the within-speaker variability can be estimated by a unique and linear parameter. A non-probabilistic approach is proposed, competitive with state-of-the-art, which reveals some limits of the Gaussian modeling in terms of goodness of fit. We carry out an analysis of residue, which finds out a relation between the dispersion of a speaker-class and its location and, thus, shows that homoscedasticity assumptions are not fulfilled.

International audience; In this paper, we describe the LIA system proposed for the MediaEval 2013 Soundtrack task. The aim is to predict the most suitable soundtrack from a list of candidate songs, given a TV commercial. The organizers provide a development dataset including multimedia features. The initial assumption of the proposed system is that commercials which sell the same type of product, also share the same music rhythm. A twofold system is proposed to provide a music for a commercial: find commercials with close subjects in order to determine the mean rhythm of this subset, and then extract from the candidate songs the music which better correspond to this mean rhythm.

International audience; In this paper, we describe the LIA system proposed for the MediaEval 2013 Spoken Web Search task. This multi-language task involves searching for an audio content query, in a database, with no training resources available. The participants must then find locations of each given query term within a large database of untranscribed audio files. For this task, we propose to build a language-independent audio search system using an i-vector based approach [2].

International audience; In this paper, we describe the LIA system proposed for the MediaEval 2013 Crowdsourcing for Social Multimedia task. The aim is to associate an accurate label to an image among multiple noisy labels collected from a crowdsourcing platform. In particular, the task participants have to predict two types of binary labels for each considered image. The first one mentions that an image is truly fashion-related or not, while the second label indicates that the fashion tag assigned to the image is related to the content of the image or not. The proposed system combines noisy crowdsourcing labels, image metadata and external resources.

In this paper, we propose a speaker verification system called m-vector system, where speakers are represented by uniform segmen-tation of their Maximum Likelihood Linear Regression (MLLR) super-vectors, denoted m-vectors. The MLLR super-vectors are extracted with respect to Universal Background Model (UBM) with MLLR adaptation using the speakers data. Two criterion are followed to segment the MLLR super-vector: one is disjoint segmen-tation technique and other one is overlapped windows. Afterward, m-vectors are conditioned by our recently proposed [1] session variability compensation algorithm before calculating score during test phase. However, the proposed method is not based on any total variability space concept and uses simple MLLR transformation for extracting m-vector without considering any transcription of the speech segment. The proposed system shows promising performance compared to the conventional i-vector system. This indicates that session variability compensation plays an important role in speaker verification. Speakers can be represented by simpler way instead of generating i-vector in conventional system and able to achieve performance comparable to the i-vector based system. Experiment results are shown on NIST 2008 SRE core condition.

International audience; I-vector extraction and Probabilistic Linear Discriminant Analysis (PLDA) has become the state-of-the-art configuration for speaker verification. Recently, Gaussian-PLDA has been improved by a preliminary length normalization of i-vectors. This normalization, known to increase the Gaussianity of the i-vector distribution, also improves performance of systems based on standard Linear Discriminant Analysis (LDA) and "two-covariance model" scoring. But this technique follows a standardization of the i-vectors (centering and whitening i-vectors based on the first and second order moments of the development data). We propose in this paper two techniques of normalization based on total, between-and within-speaker variance spectra 1. These "spectral" techniques both normalize the i-vectors length for Gaussianity, but the first adapts the i-vectors representation to a speaker recognition system based on LDA and two-covariance scoring when the second adapts it to a Gaussian-PLDA model. Significant performance improvements are demonstrated on the male and female telephone portion of NIST SRE 2010.

International audience; no abstract

Nous proposons une alternative aux méthodes état-de-l'art développées récemment pour la vérifi-cation du locuteur dans le cadre du paradigme de Variabilité Totale. Les expériences présentées montrent que l'utilisation de l'Analyse en Composante Principale (ACP) en remplacement du Factor Analysis (FA), pour la réduction de dimension des super-vecteurs, peut amener a des performances équivalentes. Ainsi l'extraction des i-vecteurs selon le critère du Maximum de Vraisemblance en utilisant la matrice des vecteurs propres obtenus par une ACP permet de surpasser un système état-de-l'art combinant Factor Analysis et Analyse Discriminante Linéaire Probabiliste dans 3 des 8 conditions de NIST-SRE08. Nous montrons également que des i-vecteurs obtenus par simple projection orthogonale sur la matrice produite par ACP peuvent surpasser l'approche état-de-l'art dans deux des conditions évaluées. Les résultats présentés dans cet article illustrent le potentiel d'une approche déterministe pour la vérification du locuteur. ABSTRACT Principal Component Analysis for i-vector extraction in speaker verification. In this work, we propose alternative algorithmic combinations for speaker verification based on the Total Variability paradigm. Experiments presented in this paper show that replacing Factor Analysis (FA) by a Principal Component Analysis (PCA) for super-vector dimensionality reduction can lead to state-of-the-art performance. Extracting the i-vectors according to the Maximum Likelihood criteria when using an Eigen Vector matrix resulting from a PCA outperforms a state-of-the-art system based on Factor Analysis and Probabilistic Linear Discriminant Analysis in 3 conditions of the NIST-SRE08 evaluation over 8. Computation of i-vectors by an orthognonal projection on the PCA matrix is also shown to outperform the state-of-the-art configuration in 2 of the 8 conditions. The results presented in this paper illustrate the potential of a Deterministic approach for speaker verification. MOTS-CLÉS : Vérification du locuteur, I-vecteurs, Réduction de dimension.

The total variability factor space in speaker verification system architecture based on Factor Analysis (FA) has greatly improved speaker recognition performances. Carrying out channel compensation in a low dimensional total factor space, rather than in the GMM supervector space, allows for the application of new techniques. We propose here new intersession compensation and scoring methods. Furthermore, this new approach contributes to a better understanding of the session variability characteristics in the total factor space.

During last decade, researchers in speaker recognition have been working over the same acoustic space, regardless of whether the data lie on a linear space or not. Our proposal is to take into account the inner geometric structure of the speech in order to obtain a new space with a better representation of the speech data. A topological approach based on manifolds obtained thanks to Laplacian and Isomap algorithms is proposed. In this first work, the proposal is evaluated in terms of dimension reduction of the supervector space, known to have a high redundancy. The experiments are done in the NIST-SRE framework. It appears that the proposed approach allows to reduce by a factor four the dimension of the supervector space without losses in terms of EER. This first result highlights the potential of topological approaches for speaker recognition.

Embedded speaker recognition in mobile devices involves a limited amount of computing resources. However , the performance of state-of-the-art systems are usually evaluated without any limitation of material resources. In this paper, we evaluate several downsca-led configurations of the LIA UBM/GMM speaker verification system. The impact of scalability is evaluated in terms of memory resources and computational time. We propose two downscale configurations which allow a good compromise between resource consumption and performance degradation. Experiments performed on the Banca database show that the memory allocation and the computational time decrease of about 70% and 88% when the error rate raises from 3, 48% to 4, 77% or still comparable to the baseline.

The Factor Analysis framework demonstrated its high power to model session variability during the past years. However, training the FA parameters implies to have a large amount of training data. When the size of the available database is limited, the number of components of the core statistical model, the UBM, is also limited as the UBM drives the dimension of the FA main matrix. As the size of the UBM gives directly the size of the speaker supervector (concatenation of the GMM mean parameters), it limits also the intrinsic capacity of the recognition system , reducing the performance expectation. This paper aims to withdraw this limitation by breaking the intrinsic link between the FA dimensionality and the UBM dimensionality. The session variability modelling is done on a smaller dimension compared to the UBM, which drives the discriminative power of the system. The first experimental results proposed in this paper, done using the NIST-SRE 2008 framework, are encouraging with a relative EER improvement of about 18% when a 512 components UBM is associated to a 32 components session variability modelling compared with a 32 components UBM associated with the same variability modelling.

Video genre classification is a challenging task in a global context of fast growing video collections available on the Internet. This paper presents a new method for video genre identification by audio analysis. Our approach relies on the combination of low and high level audio features. We investigate the discrimi-native capacity of features related to acoustic instability, speaker interactivity, speech quality and acoustic space characterization. The genre identification is performed on these features by using a SVM classifier. Experiments are conducted on a corpus composed from cartoons, movies, news, commercials and musics on which we obtain an identification rate of 91%. Index Terms: video genre classification, audio-based video processing

In speaker recognition, performance of a system is usually estimated globally on a large set of tests, even if it is well known that some subsets of tests could show a very different behavior from the complete set. In fact, a small subset of tests could represent the main part of the reported errors. In this work, we highlight a such subset of tests, for which impostors obtain very high recognition scores. We evaluate if the problem comes from the involved speakers, from the voice excerpts or from the client model estimation technique. We also propose a strategy in order to minimize the effects of the observed phenomena on the overall performance of the system.

With the purpose of improving Automatic Speech Recognition (ASR) systems performance, many different approaches on combining them have been studied. In this paper, a combination of state a posteriori probabilities given by different feature sets is proposed. In order to perform a coherent combination of state posterior probabilities, the acoustic models trained on different feature sets must have the same topo-logy (i.e. same set of states). For this purpose, a fast and efficient twin model training protocol is proposed. Then, two different strategies for combining probabilities are presented : the linear and the log linear interpolation. By using log linear interpolation, a relative Word Error Rate (WER) reduction of about 15% on MEDIA and 14% on ESTER corpora have been respectively observed.

International audience; This paper presents the ALIZE/SpkDet open source software packages for text independent speaker recognition. This software is based on the well-known UBM/GMM approach. It includes also the latest speaker recognition developments such as Latent Factor Analysis (LFA) and unsupervised adaptation. Discriminant classifiers such as SVM supervectors are also provided , linked with the Nuisance Attribute Projection (NAP). The software performance is demonstrated within the framework of the NIST'06 SRE evaluation campaign. Several other applications like speaker diarization, embedded speaker recognition , password dependent speaker recognition and pathological voice assessment are also presented.

National audience; This paper investigates the effect of a transfer function-based voice transformation on automatic speaker recognition system performance. We focus on increasing the impostor acceptance rate, by modifying the voice of an impostor in order to target a specific speaker. This paper is based on the following idea : in several applications and particularly in forensic situations, it is reasonable to think that some organizations have a knowledge on the speaker recognition method used and could impersonate a given, well known speaker. We also evaluate the effect of the voice transformation when the transformation is applied both on client and impostor trials.

National audience; Nous présentons le système de transcription automatique d'émissions radiophoniques du LIA. Il s'agit d'un système complet, intégrant une première phase de segmentation et d'extraction d'informations de haut niveau ; cette étape préliminaire est suivie du processus de décodage lui même, réalisé par le système de reconnaissance grand vocabulaire du LIA (Speeral). Les configurations du système non contrainte (tâche TRS) puis temps réel (tâche TTR) qui ont été évaluées lors de la campagne ESTER sont décrites successivement. Les résultats obtenus sont satisfaisants ; ils valident les choix qui ont été faits et permettent d'identifier des axes de travail prioritaires.

The speaker-dependent HMM-based recognizers gives lower word error rates in comparison with the corresponding speaker-independent recognizers. The aim of speaker adaptation techniques is to enhance the speaker-independent acoustic models to bring their recognition accuracy as close as possible to the one obtained with speaker-dependent models. In this paper, we propose a method using test and training data for acoustic model adaptation. This method operates in two steps. The first one performs an a priori adaptation using the transcribed training data of the closest training speakers to the test speaker. This adaptation is done with MAP procedure allowing reduced variances in the acoustic models. The second one performs an a posteriori adaptation using the MLLR procedure on the test data, allowing mapping of Gaussians means to match the test speaker's acoustic space. This adaptation strategy was evaluated in a large vocabulary speech recognition task. Our method leads to a relative gain of 15% with respect to the baseline system and 10% with respect to the conventional MLLR adaptation .