Music structure estimation has recently emerged as a central topic within the field of music information retrieval. Indeed, as music is a highly structured information stream, knowledge of how a music piece is organized represents a key challenge to enhance the management and exploitation of large music collections. This paper focuses on the benefits that can be expected from a regularity constraint on the structural segmentation of popular music pieces. Specifically, here, we study how a constraint that favors structural segments of comparable size provides a better conditioning of the boundary estimation process. First, we propose a formulation of the structural segmentation task as an optimization process, which separates the contribution from the audio features and the one from the constraint. We illustrate how the corresponding cost function can be minimized using a Viterbi algorithm. We present briefly its implementation and results in three systems designed for and submitted to the MIREX 2010, 2011, and 2012 evaluation campaigns. Then, we explore the benefits of the regularity constraint as an efficient mean for combining the outputs of a selection of systems presented at MIREX between 2010 and 2015, yielding a level of performance competitive to that of the state-of-the-art on the “MIREX10” dataset (100 J-Pop songs from the RWC database).

International audience

In audio source separation, reference guided approaches are a class of methods that use reference signals to guide the separation. In prior work, we proposed a general framework to model the deformation between the sources and the references. In this paper, we investigate a specific scenario within this framework: music separation guided by the multitrack recording of a cover interpretation of the song to be processed. We report a series of experiments highlighting the relevance of joint Non-negative Matrix Factorization (NMF), dictionary transformation, and specific transformation models for different types of sources. A signal-to-distortion ratio improvement (SDRI) of almost 11 decibels (dB) is achieved, improving by 2 dB compared to previous study on the same data set. These observations contribute to validate the relevance of the theoretical general framework and can be useful in practice for designing models for other reference guided source separation problems.

This article proposes to analyze the structural regularities from the audio and video streams of TV-programs and explore their potential for the classification of videos into program collections. Our approach is based on the spectral analysis of distance matrices representing the short-and long-term dependancies within the audio and visual modalities of a video. We propose to compare two videos by their respective spectral features. We appreciate the benefits brought by the two modalities on the performances in the context of a K-nearest neighbor classification, and we test our approach in the context of an unsupervised clustering algorithm. These evaluations are performed on two datasets of French and Italian TV programs.

We present a general multi-channel source separation framework where additional audio references are available for one (or more) source(s) of a given mixture. Each audio reference is another mixture which is supposed to contain at least one source similar to one of the target sources. Deformations between the sources of interest and their references are modeled in a linear manner using a generic formulation. This is done by adding transformation matrices to an excitation-filter model, hence affecting different axes, namely frequency, dictionary component or time. A nonnegative matrix co-factorization algorithm and a generalized expectation-maximization algorithm are used to estimate the parameters of the model. Different model parameterizations and different combinations of algorithms are tested on music plus voice mixtures guided by music and/or voice references and on professionally-produced music recordings guided by cover references. Our algorithms improve the signal-to-distortion ratio (SDR) of the sources with the lowest intensity by 9 to 15 decibels (dB) with respect to original mixtures.

International audience

This paper describes and evaluates a computational architecture to discover and collect occurrences of speech repetitions, or motifs, in a totally unsupervised fashion, that is in the absence of acoustic, lexical or pronunciation modeling and training material. In the last few years, this task has known an increasing interest from the speech community because of a) its potential applicability in spoken document processing (as a preliminary step to summarization, topic clustering, etc.) and b) its novel methodology, that defines a new paradigm to speech processing that circumvents the issues common to all supervised, trained technologies. The contributions implied by the proposed system are two-fold: 1) the design of a discovery strategy that detects repetitions by extending matches of motif fragments, called seeds; 2) the implementation of template matching techniques to detect acoustically close segments, based on dynamic time warping (DTW) and self-similarity matrix (SSM) comparison of speech templates, in contrast to the decoding procedures of model-based recognition systems. The architecture is thoroughly evaluated on several hours of French broadcast news shows according to various parameter settings and acoustic features, namely mel-frequency cepstral coefficients (MFCCs) and different types of posteriorgrams: Gaussian mixture model (GMM)-based, and phone-based posteriors, in both language-matched and mismatched conditions. The evaluation highlights a) the improved robustness of the system that jointly employs DTW and SSM and b) the relevant impact of language-specific features to acoustic similarity detection based on template matching.

Most audio source separation methods are developed for a particular scenario characterized by the number of sources and channels and the characteristics of the sources and the mixing process. In this paper, we introduce a general audio source separation framework based on a library of structured source models that enable the incorporation of prior knowledge about each source via user-specifiable constraints. While this framework generalizes several existing audio source separation methods, it also allows to imagine and implement new efficient methods that were not yet reported in the literature. We first introduce the framework by describing the model structure and constraints, explaining its generality, and summarizing its algorithmic implementation using a generalized expectation-maximization algorithm. Finally, we illustrate the above-mentioned capabilities of the framework by applying it in several new and existing configurations to different source separation problems. We have released a software tool named Flexible Audio Source Separation Toolbox (FASST) implementing a baseline version of the framework in Matlab.

Word discovery is the task of discovering and collecting occurrences of repeating words in the absence of prior acoustic and linguistic knowledge, or training material. The capability of extracting such patterns (or motifs) represents a preliminary step towards automatic mining of contentful information in spoken documents. The absence of modelling and training data, forces the use of direct pattern matching of speech templates, which, in turn, is sensitive to speech variability, like the inter-speaker one, for instance. In the present work, a variability tolerant pattern recognition technique is proposed that relies on the comparison of self similarity matrices of speech sequences. The joint use of such technique and a dynamic time warping dissimilarity measure, is shown to account for more variability with respect to the DTW-based system alone, as demonstrated on several hours of broadcast news shows.

International audience

spoken term detection, template matching, unsupervised learning, posterior features

We propose a new method for ventricular cancellation and atrial modelling in the ECG of patients suffering from atrial fibrillation. Our method is based on dictionary learning. It extends both the average beat subtraction and the sparse source separation approaches. Experiments on synthetic data show that this method can almost completely suppress the ventricular activity, but it generates some artifacts. Contrary to other ventricular cancellations methods, our approach also learns a model for the atrial activity.

We propose a variant of orthogonal matching pursuit (OMP), called LoCOMP, for scalable sparse signal approximation. The algorithm is designed for shift-invariant signal dictionaries with localized atoms, such as time-frequency dictionaries, and achieves approximation performance comparable to OMP at a computational cost similar to matching pursuit. Numerical experiments with a large audio signal show that, compared to OMP and gradient pursuit, the proposed algorithm runs in over 500 less time while leaving the approximation error almost unchanged.

The modeling of music as a language is a core issue for a wide range of applications such as polyphonic music retrieval, automatic style identification, audio to symbolic music transcription and computer-assisted composition. In this paper, we focus on the modeling of chord sequences by probabilistic N-grams. Previous studies using these models have achieved limited success, due to overfitting and to the use of a single chord labeling scheme. We investigate these issues using model smoothing and selection techniques initially designed for spoken language modeling. This approach is evaluated over a set of songs by The Beatles, considering several chord labeling schemes. Initial results show that the accuracy of N-grams is increased but that additional improvements may still be achieved in the future using more advanced, possibly music-specific, smoothing techniques.

Recently, we presented a rapid speaker adaptation technique, reference model interpolation (RMI), which is based on the linear interpolation of speaker-dependent models and the a posteriori selection of reference models. The approach uses the a priori knowledge provided by a set of representative speakers to guide the estimation of a new speaker model in the speaker space. RMI achieved rapid supervised adaptation in phoneme decoding tasks. In this paper, we present two new results of RMI: firstly, we apply the RMI technique in a practical large vocabulary continuous speech recognition (LVCSR) system with unsupervised instantaneous adaptation. Secondly, we propose an evolutional subspace scenario which integrates the slow update of reference models with RMI rapid adaptation to achieve incremental adaptation. The unsupervised adaptation experiments carried out on broadcast news transcription task show encouraging results for both instantaneous and incremental adapatation.

International audience

We propose a new method, called DEMIX anechoic, to estimate the mixing conditions, i.e. number of audio sources plus attenuation and time delay of each sources, in an underdetermined anechoic mixture. The method relies on the assumption that in the neighborhood of some time-frequency points, only one source contributes to the mixture. Such time-frequency points, located with a local confidence measure, provide estimates of the attenuation, as well as the phase difference at some frequency, of the corresponding source. The time delay parameters are estimated, by a method similar to GCC-PHAT, on points having close attenuations. As opposed to DUET like methods, our method can estimate time-delay higher than only one sample. Experiments show that DEMIX anechoic estimates, in more than 65% of the cases, the number of directions until 6 sources and outperforms DUET in the accuracy of the estimation by a factor of 10.

Probabilistic approaches can offer satisfactory solutions to source separation with a single channel, provided that the models of the sources match accurately the statistical properties of the mixed signals. However, it is not always possible to train such models. To overcome this problem, we propose to resort to an adaptation scheme for adjusting the source models with respect to the actual properties of the signals observed in the mix. In this paper, we introduce a general formalism for source model adaptation which is expressed in the framework of Bayesian models. Particular cases of the proposed approach are then investigated experimentally on the problem of separating voice from music in popular songs. The obtained results show that an adaptation scheme can improve consistently and significantly the separation performance in comparison with nonadapted models.

In this paper, we present a speaker tracking system entirely based on anchor models approach. The aim of this article is to evaluate if the probabilistic anchor models approach, which models a speaker by a normal distribution in the anchor models space, gives good performances in speaker tracking and also to investigate how speaker segment cluster information can improve speaker tracking performances. Evaluation is done on the audio database of the ESTER evaluation campaign for the rich transcription of French broadcast news. Results show that deterministic metrics on anchor models are suitable for segmentation and clustering tasks, whereas the probabilistic approach on anchor models gives interesting results for speaker-tracking. It is also observed that tracking performances are improved when all segments of a cluster are pooled together prior to the classification process. This improvement manifests itself as an improved recall rate on short segments

In this paper, we address the problem of audio source separation with one single sensor, using a statistical model of the sources. The approach is based on a learning step from samples of each source separately, during which we train Gaussian scaled mixture models (GSMM). During the separation step, we derive maximum a posteriori (MAP) and/or posterior mean (PM) estimates of the sources, given the observed audio mixture (Bayesian framework). From the experimental point of view, we test and evaluate the method on real audio examples.

This paper presents an approach for speaker tracking in a large audio database. The system described is based on a speaker segmentation procedure consisting of a detection of statistical ruptures in the speech signal followed by a speaker detection procedure using anchor models. The technique of anchor modelling is presented and a new metric to compare speech segments based on the correlation coefficient is introduced. This novel metric is evaluated and compared to the classical Euclidean and angular metrics for the speaker detection task. Evaluation is carried out on the audio database of the ESTER evaluation campaign for the rich transcription of French broadcast news. The new metric appears to be more efficient than the classical metrics for the task of speaker detection.

We propose a new method to learn overcomplete dictionaries for sparse coding structured as unions of orthonormal bases. The interest of such a structure is manifold. Indeed, it seems that many signals or images can be modeled as the superimposition of several layers with sparse decompositions in as many bases. Moreover, in such dictionaries, the efficient block coordinate relaxation (BCR) algorithm can be used to compute sparse decompositions. We show that it is possible to design an iterative learning algorithm that produces a dictionary with the required structure. Each step is based on the coefficients estimation, using a variant of BCR, followed by the update of one chosen basis, using singular value decomposition. We assess experimentally how well the learning algorithm recovers dictionaries that may or may not have the required structure, and to what extent the noise level is a disturbing factor.

In this paper, we investigate the use of a distance between Gaussian mixture models for speaker detection. The proposed distance is derived from the KL divergences and is defined as an Euclidean distance in a particular model space. This distance is simply computable directly from the model parameters thus leading to a very efficient scoring process. This new framework for scoring is compared to the classical log likelihood ratio s-core approach on a speaker verification task of the NIST 2004 evaluation and on the speaker tracking task of the ESTER french evaluation. Results shows that the proposed approach is competitive and leads to computation times divided by a factor of more than 3. 1.

The Neologos project is a speech database creation project for the French language, resulting from a collaboration between universities and industrial companies and supported by the French Ministry of Research. The goal of Neologos is to re-think the design of the speech databases in order to enable the development of new algorithms in the field of speech processing. A general method is proposed to optimize the database contents in terms of diversity of the recorded voices, while reducing the number of recorded speakers.

Classification and regression trees (CART) are convenient for low complexity speaker recognition on embedded devices. However, former attempts at using trees performed quite poorly compared to state of the art results with Gaussian Mixture Mod-els (GMM). In this article, we introduce some solutions to im-prove the efficiency of the tree-based approach. First, we pro-pose to use at the tree construction level different types of infor-mation from the GMM used in state of the art techniques. Then, we model the score function within each leaf of the tree by a lin-ear score function. Considering a baseline state of the art system with an equal error rate (EER) of 8.6 % on the NIST 2003 eval-uation, a previous CART method provides typical EER ranging between 16 % and 18 % while the proposed improvements de-crease the EER to 11.5%, with a computational cost suitable for embedded devices. 1.

We introduce a MAP estimation of speaker models in automatic speaker verification with a distance constraint: the D-MAP adaptation. The D-MAP is based on the Kullback-Leibler distances and provides an easy way to automatically compute a speaker-dependent adaptation of the model parameters. We formulate a distance constrained MAP criterion and we show an equivalence between the D-MAP adaptation and the score normalization called D-norm. From the results obtained with the D-MAP technique, we show that this method gives better performance than a classical speaker-independent MAP adaptation. It is also found that the D-MAP based system without score normalization performs similarly to a classical MAP system with a model-based score normalization.

We propose a new method to perform the separation of two sound sources from a single sensor. This method generalizes Wiener filtering with locally stationary, non-Gaussian, parametric source models. The method involves a learning phase for which we propose three different algorithm. In the separation phase, we use a sparse non negative decomposition algorithm of our own. The algorithms are evaluated on the separation of real audio data.

The paper describes a method for analyzing audio signals with an adaptive "parametric dictionary". We use sliding frames to extract elementary signals or grains from the analysis signal. We search for similarities amongst the collected grains to form classes, which we then use to derive a signal model for each class. These signal models or prototypes, are used to decompose the audio signal and compute analysis parameters for each grain. As a preliminary evaluation, we tested the method with real-life, monophonic and monaural recordings and obtained encouraging results.

This paper focuses on the use of hidden Markov models (HMMs) for structure analysis of videos, and demonstrates how they can be efficiently applied to merge audio and visual cues. Our approach is validated in the particular domain of tennis videos. The basic temporal unit is the video shot. Visual features describe the audio events within a video shot. The video structure parsing relies on the analysis of the temporal interleaving of video shots, with respect to prior information about tennis content and editing rules. As a result, typical tennis scenes are identified. In addition, each shot is assigned to a level in the hierarchy described in terms of point, game and set.

Because of recent events and as members of the scientific community working in the field of speech processing, we feel compelled to publicize our views concerning the possibility of identifying or authenticating a person from his or her voice. The need for a clear and common message was indeed shown by the diversity of information that has been circulating on this matter in the media and general public over the past year. In a press release initiated by the AFCP and further elaborated in collaboration with the SpLC ISCA-SIG, the two groups herein discuss and present a summary of the current state of scientific knowledge and technological development in the field of speaker recognition, in accessible wording for nonspecialists. Our main conclusion is that, despite the existence of technological solutions to some constrained applications, at the present time, there is no scientific process that enables one to uniquely characterize a person’s voice or to identify with absolute certainty an individual from his or her voice. 1.

In this paper, we propose a new score normalization technique in Automatic Speaker Verification (ASV): the D-Norm. The main advantage of this score normalization is that it does not need any additional speech data nor external speaker population, as opposed to the state-of-the-art approaches. The D-Norm is based on the use of Kullback-Leibler (KL) distances in an ASV context. In a first step. we estimate the KL distances with a Monte-Carlo method and we experimentally show that they are correlated with the verification scores. In a second step, we use this correlation to implement a score normalization procedure, the D-Norm. We analyse its performance and we compare it to that of a conventional normalization, the Z-Norm. The results show that performance of the D-Norm is comparable to that of the Z-Norm. We then conclude about the results we obtain and we discuss the applications of this work.

We propose a formalism, called vector filtering of spectral trajectories, that allows the integration of a number of speech parameterization approaches (cepstral analysis, /spl Delta/ and /spl Delta//spl Delta/ parameterizations, auto-regressive vector modeling, ...) under a common formalism. We then propose a new filtering, called contextual principal components (CPC) or time-frequency principal components (TFPC). This filtering consists in extracting the principal components of the contextual covariance matrix, which is the covariance matrix of a sequence of vectors expanded by their context. We apply this new filtering in the framework of closed-set speaker identification, using a subset of the POLYCOST database. When using speaker-dependent TFPC filters, our results show a relative improvement of approximately 20% compared to the use of the classical cepstral coefficients augmented by their /spl Delta/-coefficients, which is significantly better with a 90% confidence level.

This article presents a new class of constrained and specialized autoregressive (AR) processes. They are derived from lattice filters where some reflection coefficients are forced to zero at a priori locations. Optimizing the filter topology allows to build parametric spectral models that have a greater number of poles than the number of parameters needed to describe their location. These NUT (non-uniform topology) models are assessed by evaluating the reduction of modeling error with respect to conventional AR models.

This paper deals with the tracking of speech segments in audio documents. We use a cepstral-based acoustic analysis and Gaussian mixture models for the representation of the training data. Three ways of scoring an audio document based on a frame-level likelihood calculation are proposed and compared. Our experiments are done on a database composed of television programs including news reports, advertisements, and documentaries. The best equal error rate obtained is approximately 12%.

Last year, SpLC - an ISCA Special Interest Group (SIG) centered around Speaker and Language Characterization was born. The aims of this paper are to present the SpLC SIG, its objectives, and the work done during the first year.

International audience

Classical adaptation approaches are generally used for speaker or environment adaptation of speech recognition systems. In this paper, we use such techniques for the incremental training of client models in a speaker verification system. The initial model is trained on a very limited amount of data and then progressively updated with access data, using a segmental-EM procedure. In supervised mode (i.e. when access utterances are certified), the incremental approach yields equivalent performance to the batch one. We also investigate on the impact of various scenarios of impostor attacks during the incremental enrollment phase. All results are obtained with the Picassoft platform-the state-of-the-art speaker verification system developed in the PICASSO project.

A particular form of constraint is incorporated to Linear Prediction lattice filter models in the form of unequal-length delays. This constraint amounts to reducing the number of intrinsic degrees of freedom defined by the reflection coefficients without modifying the LPC order of the corresponding transfer function. It can be optimized by a simple exhaustive search scheme. Preliminary results show that the prediction error is slightly decreased with respect to a conventional predictor using the same number of reflection coefficients.

In this article, we investigate on alternative speech features for speaker characterization. We study Line Spectrum Pairs features, Time-Frequency Principal Components and Discriminant Components of the Spectrum. These alternative features are tested and compared on a task of speaker verification. This task consists in verifying a claimed identity from a speech segment. Systems are evaluated on a subset of the evaluation data of the NIST 1999 speaker recognition campaign. The new speech features are also compared to the classical cepstral coefficients, which remain, in our experiments, the best performing features.

Keywords: speech Note: IDIAP-RR 00-48 Reference EPFL-CONF-82639 URL: http://publications.idiap.ch/downloads/reports/2000/rr00-48.pdf Related documents: http://publications.idiap.ch/index.php/publications/showcite/marietho:2000:rr00-48 Record created on 2006-03-10, modified on 2017-05-10

The problem of labelling speaker turns by automatically segmenting and clustering a continuous audio stream is addressed. A new clustering scheme is presented and evaluated using a clustering efficiency score which treats both agglomerative and divisive clustering strategies equally. Results show an efficiency of 70% can be obtained on both manually and automatically derived segments on the 1996 Hub4 development data. For the task of identifying potentially unknown anchor speakers within broadcast news shows, the frame classification error rate is very important. To reflect this, a frame-based cluster efficiency is defined and the results show a 90% frame-based efficiency can be achieved. Finally a frame-based comparison between the manually and automatically derived segment/cluster sets shows that approximately one third of the errors are introduced during segmentation and two-thirds during clustering.

In order to create a useful, gracefully functioning system for travel arrangements, we have first observed the task as it is accomplished by a human.We then imitated the human while making the user believe he was dialoguing with an automatic system.As we gradually built our system, we devised ways to assess progress and to detect errors.The following described the manner in which the Carnegie Mellon Communicator was built, data collected, and assessment begun using these criteria.

Harmonic plus noise models have been successfully applied to a broad range of speech processing applications, including, among others, low bit-rate speech coding, and speech restoration and transformation. In conventional methods, the frequencies, the relative phases and the amplitudes of the pitch-harmonic components are assumed to be piecewise constants over an analysis frame. This assumption is inadequate in segments where fast variations of these parameters may occur, e.g. phoneme-to-phoneme boundaries or speech onsets. In this paper, a time-varying model of the pitch-harmonic parameter is presented. It is based on a basis expansion technique, consisting in representing the time-varying functions as a linear combination of a fixed basis function. An estimation procedure for the parameters of this expansion is presented. Results are provided to demonstrate the effectiveness of this approach.

The issue of a priori threshold setting in speaker verification is a key problem for field applications. In the context of the Caller Verification in Banking and Telecommunications (CAVE) project, we compared several methods for estimating speaker-independent and speaker-dependent decision thresholds. Relevant parameters are estimated from development data only, i.e. without resorting to additional client data. The various approaches are tested on the Dutch SESP database.

The problem of how to train variance parameters on scarce data is addressed in the context of text-dependent, HMM-based, automatic speaker verification. Three variations of variance flooring is explored as a means to prevent over-fitting. With the best performing one, the floor to a variance vector of a client model is proportional to the corresponding variance vector in a non-client multi-speaker model. It is also found that adapting the means and mixture weights from the non-client model while keeping variances constant works comparably to variance flooring and is much simpler. Comparisons are made on three large telephone quality corpora. 1. INTRODUCTION In practical applications, Automatic Speaker Verification (ASV) systems are generally used in contexts where very few client enrollment data are available. One problem with using small training data is the risk of over-training, that is, parameters of the client model are over-fitted to the particular training data. Especially var...

International audience

Language models are usually evaluated on test texts using the perplexity derived directly from the model likelihood function. In order to use this measure in the framework of a comparative evaluation campaign, we have developed an alternative scheme for perplexity estimation. The method is derived from the Shannon (1951) game and based on a gambling approach on the next word to come in a truncated sentence. We also use entropy bounds proposed by Shannon and based on the rank of the correct answer, in order to estimate a perplexity interval for non-probabilistic language models. The relevance of the approach is assessed on an example.

In the field of speech recognition, the patterns assumed to structure the speech material (phonemes, triphones, words...) are defined a priori according to a linguistic criterion, whereas the recognition criterion is based on an acoustic similarity measure. From this may result a lack of consistency for the recognition units. We explore the possibility of a more data-driven approach, where recognition units are derived according to an acoustic criterion, and then, mapped to variable length sequences of phonemes in an unsupervised way. Continuous speech recognition experiments are reported to evaluate the consistency of those units as opposed to linguistically defined units.

Cet article présente une méthode d'ajustement des seuils de vérification du locuteur basée sur un modèle Gaussien des distributions du logarithme du rapport de vraisemblance. L'article expose les hypothèses sous lesquelles ce modèle est valide, indique plusieurs méthodes d'ajustement des seuils, et en illustre les apports et les limites par des expériences de vérification sur une base de données de 20 locuteurs.

This paper summarizes the main results from the Speaker Verification (SV) research pursued so far in the CAVE project. Different state-of-the art SV algorithms were implemented in a common HMM framework and compared on two databases : YOHO (office environment speech) and SESP (telephone speech). This paper is concerned with the different design issues for LR-HMM-based SV algorithms which emerged from our investigations and which led to our current SV system, which delivers Equal Error Rates below 0.5 % on a very realistic telephone speech database. 1. CONTEXT The CAVE project (CAller VErification in Banking and Telecommunications) is a 2 year project supported by the Language Engineering Sector of the Telematics Applications Programme of the European Union, and for the Swiss partners by the Office F'ed'eral de l'Education et de la Science (Bundesamt fur Bildung und Wissenschaft). The partners are Dutch PTT Telecom, KUN, KTH, ENST, UBILAB, IDIAP, VOCALIS, TELIA and Swiss Telecom PTT. ...

In this paper, we investigate on the role of dynamic information on the performances of AR-vector models for speaker recognition. To this purpose, we design an experimental protocol that destroys the time structure of speech frame sequences, which we compare to a more conventional one, i.e., keeping the natural time order. These results are also compared with those obtained with a (single) Gaussian model. Several measures are systematically investigated in the three cases, and different ways of symmetrisation are tested. We observe that the destruction of the time order can be a factor of improvement for the AR-vector models, and that results obtained with the Gaussian model are merely always better. In most cases, symmetrisation is beneficial.

The aim of this paper is to describe how the combination of speaker verification algorithms with a priori decision thresholds can improve the overall robustness of a real application. The evaluation is performed in the context of a field application where each client is verified from a 7 digit pin code. This paper demonstrate that it is possible to increase the global performances of the system on combining the result of several algorithms.

The multigram model assumes that language can be described as the output of a memoryless source that emits variable-length sequences of words. The estimation of the model parameters can be formulated as a maximum likelihood estimation problem from incomplete data. We show that estimates of the model parameters can be computed through an iterative expectation-maximization algorithm and we describe a forward-backward procedure for its implementation. We report the results of a systematical evaluation of multigrams for language modeling on the ATIS database. The objective performance measure is the test set perplexity. Our results show that multigrams outperform conventional n-grams for this task.

International audience

The conventional n-gram language model exploits dependencies between words and their fixed-length past. This letter presents a model that represents sentences as a concatenation of variable-length sequences of units and describes an algorithm for unsupervised estimation of the model parameters. The approach is illustrated for the segmentation of sequences of letters into subword-like units. It is evaluated as a language model on a corpus of transcribed spoken sentences. Multigrams can provide a significantly lower test set perplexity than n-gram models.>

Theoretical aspects and practical applications are reported of two variants of the AR-vector modeling technique: the standard AR-vector model and the target driven (or multistep excited) AR-vector model. The standard version supposes a white excitation, while the target driven model assumes a piecewise constant input. The standard AR-vector model turns out to be extremely efficient for speaker recognition, since, for a set of 420 different speakers, the recognition score ranges from 93% to 100%, depending on the duration of the test speech sample. The target driven AR-vector model shows very interesting properties for speech analysis and segmentation. There exists a strong correspondence between the steps in the input function and the underlying phonetic content of speech. Moreover, under some normalization, the values of the steps can be interpreted as acoustic targets.>

Two models, the temporal decomposition and the multivariate linear prediction, of the spectral evolution of speech signals capable of processing some aspects of the speech variability are presented. A series of acoustic-phonetic decoding experiments, characterized by the use of spectral targets of the temporal decomposition techniques and a speaker-dependent mode, gives good results compared to a reference system (i.e., 70% vs. 60% for the first choice). Using the original method developed by Laforia, a series of text-independent speaker recognition experiments, characterized by a long-term multivariate auto-regressive modelization, gives first-rate results (i.e., 98.4% recognition rate for 420 speakers) without using more than one sentence. Taking into account the interpretation of the models, these results show how interesting the cinematic models are for obtaining a reduced variability of the speech signal representation.>

The technique of temporal decomposition is used to initialize continuous density hidden Markov models. The temporal decomposition process produces a representation of each word in terms of a set of target vectors and interpolation functions. Roughly speaking, the target vectors represent the centers of the important acoustic events, and the interpolation functions describe a spectral path between these events. The number of targets generated by the temporal decomposition process is taken to be the number of states used for the HMM, and the position, shape and length of the interpolation functions are used to provide initial estimates for the transition probabilities and observation probability densities of the HMM. The performance of such a system is assessed for a single-speaker environment.>

The authors report results obtained in the framework of phonetic feature extraction using TDNNs (time-delay neural networks). Experiments were conducted on a set of 5270 phonemes manually segmented from natural continuous speech uttered by one male speaker. Nearly all scores on binary phonetic features range between 90% and 99%. Comments on the visualization of the weights in several networks are also presented. In particular, it is noted that there exist quantitative similarities between patterns of weights within the TDNN and acoustical characteristics of the phonetic feature.>

The automatic recognition of continuous speech may use a symbolic representation of the acoustic signal in order to facilitate lexical access. The allophones of the language form a practical set of symbols. A major issue is a reliable localisation of these units in the speech stream and their identification. Localisation is obtained using a robust implementation of temporal decomposition, a technique originally proposed by Atal (1983), for speech coding. Speech is decomposed in terms of overlapping events characterized by both a spectral target and a time-limited interpolation function. An undershot target may be reestimated using neighbours and the associated functions. The possibility of undoing the effects of coarticulation is the major contribution of this work. The identification of these corrected targets is therefore possible with no further contextual rules. The recognition of spelled surnames (letters of the alphabet) is used for evaluation. 76% of correct phones allow 70% of correct letters.>

Localization is realized using a robust implementation of temporal decomposition, a technique originally proposed for speech coding. Speech is decomposed in terms of overlapping events characterized by spectral targets and time-limited interpolation functions. The spectral targets are evaluated iteratively. An undershot target may be reestimated using neighbors and the associated functions. The possibility of undoing the effects of coarticulation is also suggested. The identification of these corrected targets is therefore possible with no further contextual rules. Applications to speech recognition strategies are discussed. The recognition of spelled surnames (letters of the alphabet) is used for evaluation. The correct recognition of 76% of correct phones is the basis of 70% correct recognition letters.>

ATAL [1] introduced a technique for decomposing speech into phone-length temporal events in terms of overlapping and interacting articulatory gestures. This paper reports on simplifications of this technique with applications to acoustic-phonetic synthesis. Spectral evolution is represented by time-indexed trajectories in the p-dimensional space of Log-Area Ratios{y_{i}= \Ln ((1+k_{i})/(1-k_{i}))}where kiare the reflection coefficients obtained from short-time stationary LPC analysis. The vocal tract configuration (spectral vector) associated with each interpolation function belongs to a finite set of articulatory targets (vector quantization code book). A set of speech segments ("polysons") has been encoded using this technique. It includes diphones, demi-syllables, and other units that are difficult to segment. Temporal decomposition using target spectra can break the complex encoding of these segments. In particular, coarticulation effects are analyticaiy explained and modeled. It is demonstrated that these new tools provide an adequate environment in our search for better rules in acoustic speech synthesis.