International audience

I4U is a joint entry of nine research Institutes and Universities across 4 continents to NIST SRE 2012. It started with a brief discussion during the Odyssey 2012 workshop in Singapore. An online discussion group was soon set up, providing a discussion platform for different issues surrounding NIST SRE’12. Noisy test segments, uneven multi-session training, variable enrollment duration, and the issue of open-set identification were actively discussed leading to various solutions integrated to the I4U submission. The joint submission and several of its 17 sub-systems were among top-performing systems. We summarize the lessons learnt from this large-scale effort.

Footstep recognition is a relatively new biometric which aims to discriminate people using walking characteristics extracted from floor-based sensors. This paper reports for the first time a comparative assessment of the spatiotemporal information contained in the footstep signals for person recognition. Experiments are carried out on the largest footstep database collected to date, with almost 20,000 valid footstep signals and more than 120 people. Results show very similar performance for both spatial and temporal approaches (5 to 15 percent EER depending on the experimental setup), and a significant improvement is achieved for their fusion (2.5 to 10 percent EER). The assessment protocol is focused on the influence of the quantity of data used in the reference models, which serves to simulate conditions of different potential applications such as smart homes or security access scenarios.

Abstract Sound is an integral part of most movies and videos. In many situations, viewers of a video are unable to hear the sound track, for example, when watching it in a fast forward mode, viewing it by hearing‐impaired viewers or when the plot is given as a storyboard. In this paper, we present an automated visualization solution to such problems. The system first detects the common components (such as music, speech, rain, explosions, and so on) from a sound track, then maps them to a collection of programmable visual metaphors, and generates a composite visualization. This form of sound visualization, which is referred to as SoundRiver, can be also used to augment various forms of video abstraction and annotated key frames and to enhance graphical user interfaces for video handling software. The SoundRiver conveys more semantic information to the viewer than traditional graphical representations of sound illustration, such as phonoautographs, spectrograms or artistic audiovisual animations.

This paper describes a semi-automatic system to capture and label a reasonable size biometric database. In our case, the biometric to be assessed are footstep signals, but the system could be extendable to other biometrics. Extra biometric data such as the voice and video recordings of the face and the gait are used to assist the database labelling to minimise the error. Thus, audio identifier recordings are used to automatically label the database with a speaker recognition system achieving results of 0.15% of equal error rate (EER) of person verification using Gaussian mixture models (GMM). Also, a footstep detector system has been developed to reduce the presence of invalid signals from the database having a percentage of less than 1% of correct footsteps miss-classified using features from the ground reaction force (GRF) and using a support vector machine (SVM) classifier. To date, more than 20,000 footstep signals have been collected from more than 100 people, which is well beyond previously reported databases. The database is collected in different sessions which will allow us to study how different factors such as footwear, the person carrying a load or different walking speeds affect the recognition of persons using their footsteps.

The BioSecure Network of Excellence has collected a large multi- biometric publicly available database and organized the BioSecure Multimodal Evaluation Campaigns (BMEC) in 20072. This paper reports on the Talking Faces campaign. Open source reference systems were made available to participants and four laboratories submitted executable code to the organizer who performed tests on sequestered data. Several deliberate impostures were tested. It is demonstrated that forgeries are a real threat for such systems. A technological race is ongoing between deliberate impostors and system developers.

Embedded speaker recognition in mobile devices could involve several ergonomic constraints and a limited amount of com-puting resources. Even if they have proved their efficiency in more classical contexts, GMM/UBM based systems show their limits in such situations, with good accuracy demanding a rel-atively large quantity of speech data, but with negligible har-nessing of linguistic content. The proposed approach addresses these limitations and takes advantage from the linguistic nature of the speech material into the GMM/UBM framework by us-ing client-customised utterances. The GMM/UBM is then rein-forced with new temporal information. Experiments on the MyIdea database are performed when im-postors know the client-utterance and also when they do not, highlighting the potential of this new approach. A relative gain up to 45 % in terms of EER is achieved when impostors do not know the client utterance and performance is equivalent to the GMM/UBM baseline system in other configurations. 1.

This paper assesses 9 prominent objective quality measures for their potential in intelligibility estimation. Degradation considered include additive noises and those introduced by coding and enhancement schemes, totalling 78 types. This paper is believed to be the first to conduct an assessment on such a large combination of quality measures and degradations allowing side-byside analysis. Experimental results show that the sophisticated perceptual-based measures which are superior for quality estimation, do not necessarily correlate well with human intelligibility and, in fact, give poorer correlations when enhancement schemes are considered. Meanwhile, the weighted spectral slope (WSS) emerges to be the most promising approach among all measures considered, scoring the highest correlation in 5 out of the 6 test sets. Worth noting are the positive correlations obtained with WSS which range from 0.14 to 0.86, as opposed to those with PESQ from -0.58 to 0.74. Such findings put WSS, a relatively conventional measure, in a new light as a potential intelligibility assessor.

Embedded speaker recognition in mobile devices could involve several ergonomic constraints and a limited amount of computing resources. Even if they have proved their efficiency in more classical contexts, GMM/UBM based systems show their limits in such situations, with good accuracy demanding a relatively large quantity of speech data, but with negligible harnessing of linguistic content. The proposed approach addresses these limitations and takes advantage of the linguistic nature of the speech material into the GMM/UBM framework by using clientcustomised utterances. Furthermore, the acoustic structure is then reinforced with video information. Experiments on the MyIdea database are performed when impostors know the client utterance and also when they do not, highlighting the potential of this new approach. A relative gain up to 47% in terms of EER is achieved when impostors do not know the client utterance and performance is equivalent to the GMM/UBM baseline system in other configurations.

International audience

This paper illustrates an evolution in state-of-the-art speaker verification by highlighting the contribution from newly developed techniques. Starting from a baseline system based on Gaussian mixture models that reached state-of-the-art performances during the NIST'04 SRE, final systems with new intersession compensation techniques show a relative gain of around 50%. This work highlights that a key element in recent improvements is still the classical maximum a posteriori (MAP) adaptation, while the latest compensation methods have a crucial impact on overall performances. Nuisance attribute projection (NAP) and factor analysis (FA) are examined and shown to provide significant improvements. For FA, a new symmetrical scoring (SFA) approach is proposed. We also show further improvement with an original combination between a support vector machine and SFA. This work is undertaken through the open-source ALIZE toolkit.

As with many approaches to noise robust automatic speech recognition (ASR) the benefits of spectral subtraction tend to diminish as noise levels in the order of 0 dB are approached. Whilst the majority of related work focuses on reducing magnitude errors a number of new approaches addressing the often overlooked, additional sources of error have appeared in the literature in recent years. Relatively lacking in the literature, however, is an empirical assessment which compares the effects of each error when noisy speech is processed by spectral subtraction. Such studies are vital in order to appreciate the potential penalty in performance when sources of error are overlooked. The objective in this paper is to assess, through ASR, the performance penalty associated with each source of error when noisy speech is treated with spectral subtraction. Experimental evidence based on two standard European databases and ASR protocols illustrates that, perhaps contrary to popular belief, for noise levels in the order of 0 dB and below, these often overlooked sources of error can lead to non-negligible degradations in performance. Whilst not a new idea, here the original emphasis is a thorough assessment that empirically highlights both the fundamental limitations and potential benefit of including the full complement of errors in the spectral subtraction model

This paper investigates the accuracy of automatic speech recognition (ASR) and 6 other well-reported objective quality measures for the task of estimating speech intelligibility. It is believed to be the first assessment of such a range of measures side-by-side and in the context of intelligibility. A total of 39 degradation conditions including those from a newly proposed low bit rate (0.3 to 1.5 kbps) codec and a noise suppression system are considered. They provide real and varied scenarios to assess the measures. The objective scores are compared to subjective listening scores, and their correlation used to assess the approach. All tests are conducted on the European standard Aurora 2 corpus. Experiments show that ASR and perceptual estimation of speech quality (PESQ) are potentially reliable estimators of intelligibility with subjective correlation as high as 0.99 and 0.96 respectively. Furthermore, ASR gives a trend corresponding to that of subjective intelligibility assessment for the different configurations of the new codec, while most others fail

This paper investigates the potential applicability of automatic speech recognition (ASR) and 6 well-reported objective quality measures for the task of ranking intelligibility of speech degraded by different real life background noises. In a recent investigation ASR has been reported to give high subjective correlation with human assessment when tested with various system degradations. This paper extends this investigation in two directions. First, the usefulness of the measures in the context of different real-life noises is considered. Second, the direct correspondence between statistics computed by an ASR system and human perceived intelligibility is assessed. Subjective listening tests are carried out to provide ground truth. Results show that ASR and WSS (weighted spectral slope) are the only two measures out of the seven considered to give good correlation with human opinion. Specially noted is performance of ASR with correlations ranging from 0.77 to 0.90. Index Terms: speech intelligibility, objective quality measures, ASR

This paper begins by introducing biometrics and their underlying performance factors. Biometrics are sometimes classed as either behavioural or physiological. Difficulties with these classes are discussed in terms of the importance of dynamics, highlighting the key point that definitions are clarified if the biometric information-bearing signal itself is considered. Emphasis is then given to visual speech in the form of lip profiles The case is made that these are a special case in that they provide a vehicle for a twin biometric: both behavioural and physiological It is argued that lips might well be unique in providing a practical twin biometric. Illustration is presented in the form of practical experiments based around visual speech and lip profiles. Experimental results using short, test and training segments from video recordings give recognition error rates as: physiological lip-profiles 2% and behavioural lip-profiles 15%.

In the context of mobile telephony there is a need for low resource, computationally efficient noise compensation and speech enhancement approaches. This paper assesses the performance of efficient quantile-based noise estimation integrated into a nonlinear spectral subtraction framework. The approach has been implemented in real-time with minimal latency on a 500Mhz processor and is well within the processing capabilities. Experiments are reported on the AURORA 2 and AURORA 3 corpa. Results show an average relative improvement of 15% on the clean and multicondition training sets of the AURORA 2 database and an overall average relative improvement of 20% across the four AURORA 3 databases. It is acknowledged that these are not state-of-the-art results and further optimisation is anticipated.

Applying speech technology in appliances available around the world cannot restrict the functionality to a certain language. However, most of today's text-independent verification systems based on Gaussian mixture models, GMMs, use an adaptive approach for training the speaker model. This assumes that the world model incorporates the same language as that of the target speaker. We investigate language mismatches between the target speaker and the world model in a GMM speaker verification system. Experiments performed with different world model languages showed major degradations, in particular for Mandarin and Vietnamese when the target speakers spoke American English. Experiments with world models trained on data pooled from different languages revealed only minor performance degradations.

The problem addressed here is the classification of videos at the highest level into pre-defined genre. The approach adopted is based on the dynamic content of short sequences (/spl sim/30 secs). This paper presents two methods of extracting motion from a video sequence: foreground object motion and background camera motion. These dynamics are extracted, processed and applied to classify 3 broad classes: sports, cartoons and news. Experimental results for this 3 class problem give error rates of 17%, 8% and 6% for camera motion, object motion and both combined respectively, on /spl sim/30 second sequences.

This paper is in two parts. The first part quantatively assesses an approach to skin segmentation. The second part describes the development and quantitative assessment of an approach to face detection (FD), with the application of content-based image retrieval in mind. Skin detection is introduced as a front-end to an earlier approach to FD by Huang (1994). The baseline approach searches grey scale images only, and is found to be susceptible to variations in lighting conditions and complex backgrounds. It is hypothesised that by integrating colour information into Huang's approach, the number of false faces can be reduced. A skin probability map (SPM) is generated from a large quantity of labeled data (530 images containing faces and 714 images that do not) and is used to pre-process colour test images. Image regions are then ranked in terms of their skin content, thus removing improbable face regions. The performance improvements are shown in terms of false acceptance (FA) and false rejection (FR) scores. As a front-end to Huang's approach, the benefits of skin segmentation can be seen by a reduction in the FA score from 79% to 15% with a negligible impact on FR.

This paper describes the development and quantitative assessment of an approach to face detection (FD), with the application of image classification in mind. The approach entitled 'crude automatic face extraction' (CAFE), is based on skin colour segmentation, multiple image representations and multi-resolution scanning. It is a direct extension of an approach by Huang (1994). Assessment is in terms of false acceptance (FA) and false rejection (FR) scores using a large quantity of unconstrained typical Internet images, some of which contain faces and some of which do not. Skin-based colour segmentation is introduced as a front-end, along with several additional low-level image representations, the benefits of which can be seen by a reduction in the FA score from 74% to 17%. It is also noted that a skin probability map can improve FD significantly.

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.

Automatic speech recognition performance tends to be degraded in noisy conditions. Spectral subtraction is a simple, popular approach of noise compensation. In conventional spectral subtraction [1, 2], noise statistics are updated during speech gaps and subtracted from a corrupt signal during speech intervals. Some means of explicit speech, non-speech detection is therefore essential. Recent proposals have avoided the problem of speech, non-speech detection [3, 4, 5, 6, 7] by continually updating noise estimates whether speech is present or not.

In this paper we propose an approach to high-level classification of video into genre: sport, cartoon, news, commercial and music. An important issue for automatic high-level classification systems is the amount of time needed to classify a video. Here we investigate classification performance as a function of the test sequence length. In addition we present performance against different orders and combinations of static and dynamic mel-frequency cepstral coefficients (MFCC). We find that static and delta MFCCs perform well for this classification task. A test sequence length of approximately 25 seconds for the 5 class problem gives approximately 80 % correct identification. 1.

Standard Gaussian mixture modelling does not possess time sequence information (TSI) other than that which might be embedded in the acoustic features. Dynamic time warping relates directly to TSI, time-warping two sequences of features into alignment. Here, a hybrid system embedding DTW into a GMM is presented. Improved automatic speaker verification performance is demonstrated. Testing 1000 speakers in a fully text independent, world-model-adapted mode shows an equal error improvement over a standard GMM from 4.1% to 3.8%.

In this paper we introduce the concept of a continuous speaker verification system in a mobile phone environment. The system verifies the speaker during the phone call. We discuss speaker-centric score normalisation and time pattern analysis which extends a speaker verification system to allow continuous verification. Experiments showed that speaker-centric score normalisation improved performance. Moreover, it standardises the target score distribution to allow the accurate prediction of miss probabilities. The performance evaluation of time pattern analysis revealed an impostor locking of 95% while the target speakers remained unlocked.

This paper assesses the merits of three different approaches to pixel-level human skin detection. The basis for the 3 approaches has been reported in the literature. The first two approaches use simple ratios and colour space transforms respectively, whereas the third is a numerically efficient approach based on a 3D RGB probability map, first implemented by Rehg-Jones (1999). The Bayesian probabilities are made possible to compute only with the availability of a large appropriately labeled database. Over 12000 images from the Compaq skin and non-skin databases are used to quantitatively assess the three approaches. Thresholds are determined empirically to detect 95% of all skin-associated pixels and assessment is then made in terms of the percentage of non-skin pixels incorrectly accepted. The lowest of these false acceptance rates is found to be about 20% given by the 3D probability map.

This paper gives an insight into biometrics used for speaker recognition. Three different biometrics are presented, based on: acoustic, geometric lip, and holistic facial features. Experiments are carried out using a corpus of the DAVID audio-visual database. Recognition accuracy is found to be similar in the 2 domains. The geometric visual feature is based on a method of signature coding of the contour of the lips and the holistic feature is based on a mean dynamic signature, a method of capturing the motions of the face during a spoken utterance. Physical biometrics (static measurements) demand only small model sizes, perhaps just a single template, and therefore require less training data. Conversely behavioral biometrics contain more variation and demand more training data.

Given that the amount of speaker specific training data is always limited, for a given amount of data a speaker model has an optimum number of components. Here, this is investigated with regard to Gaussian mixture models (GMM) with and without world model adaption. Test results show that maximising the number of components in a speaker model can improve speaker recognition results. Comparisons with vector quantisation (VQ) indicate that sensible use of out-of-class data is essential for optimising a recognition system.

This paper evaluates lip features for person recognition, and compares the performance with that of the acoustic signal. Recognition accuracy is found to be equivalent in the two domains, agreeing with the findings of Chibelushi (1997). The optimum dynamic window length for both acoustic and visual modalities is found to be about 100 ms. Recognition performance of the upper lip is considerably better than the lower lip, achieving 15% and 35% identification error rates respectively, using a single digit test and training token.

The integration of multiple classifiers promises higher classification accuracy and robustness than can be obtained with a single classifier. This paper proposes a new adaptive technique for classifier integration based on a linear combination model. The proposed technique is shown to exhibit robustness to a mismatch between test and training conditions. It often outperforms the most accurate of the fused information sources. A comparison between adaptive linear combination and non-adaptive Bayesian fusion shows that, under mismatched test and training conditions, the former is superior to the latter in terms of identification accuracy and insensitivity to information source distortion.

The objective of this work is a computationally efficient method for inferring vocal tract shape trajectories from acoustic speech signals. We use an multilayer perceptron (MLP) to model the vocal tract shape-to-acoustics mapping, then in an analysis-by-synthesis approach, optimise an objective function that includes both the accuracy of the spectrum approximation and the credibility of the vocal tract dynamics. This optimisation carries out gradient descent using backpropagation of derivatives through the MLP. Employing a series of MLPs of increasing order avoids getting trapped in local optima caused by the many-to-one mapping between vocal tract shapes and acoustics. We obtain two orders of magnitude speed increase compared with our previous methods using codebooks and direct optimisation of a synthesiser.

Recent work on ASR by [1] [2] shows that band splitting gives recognition accuracy comparable with the conventional fullband. Sub-bands have different bandwidth spaced on a mel scale. Interestingly in the contex of speaker recognition improved accuracy has been reported in the case of a full-band approach using a linear scale. We demonstrate that both of these scales are likely to be suboptimum in the context of band splitting. We then describe, how sub-band error profiles can lead to a new scale, which is between a linear and a mel spacing, giving both an equalised sub-band error profile and an improved overall recognition accuracy. 1. INTRODUCTION This paper is concerned with splitting the conventional acoustic representation into sub-band units and processing these separately, with recombination at the decision stage. This idea has been investigated recently in the context of speech recognition [1] [2]; here our task is the complementary one of speaker recognition. Potential bene...

The vast majority of work in continuous speech recognition uses phoneme-like units as the basic recognition component. The work presented here investigates the practicability of syllable-like units as the building blocks for recognition. A phonetically annotated telephony database is analysed at the syllable level, and a set of syllable-based HMMs are built. Refinements including the introduction of syllable-level bigram probabilities, word- and syllable-level insertion penalties, and the investigation of different model topologies are found to improve recogniser performance. It is found that the syllablebased recogniser gives recognition accuracies of over 60%, which compares with 35% as the baseline accuracy for monophone recognition. It is envisaged that practical applications of syllable recognition could be in a hybrid system, where the most common syllable HMMs would be used in conjunction with whole-word and phoneme models. 1. INTRODUCTION Continuous word recognition using gram...

This paper describes a cross-validation method to determine the appropriate weight with which dynamic constraints should be applied when estimating vocal tract shapes from speech. This data-dependent method can estimate the weighting without the need for separate prior knowledge of the source and noise statistics. The principles are first demonstrated on a simple onedimensional system analogous to speech production. As the data here is synthetic, the statistics are known, and so the success of the method can be objectively assessed. Next, the same principles are extended to real speech to improve the estimation of vocal tract shape trajectories. 1 INTRODUCTION For speech coding, recognition, and synthesis, it is believed that an articulatory description may be preferable to conventional acoustic features because physical limitations of the vocal tract imply slowly changing parameters. Also, the close relationship between the articulatory and phonetic domains suggests that an articul...

Compared with the usual acoustic representations, articulatory models offer potential benefits in giving a compact, slowly changing representation, having a closer relationship with the phonetic domain and allowing a straightforward treatment of coarticulation and transitional effects. A long-standing difficulty preventing the realisation of these benefits is the estimation of appropriate parameters from the speech signal. This paper reviews recent progress in this area and describes some initial experiments that attempt to estimate the parameters of the recently proposed Distinctive Regions Model (DRM) [1], using a dynamic programming search of an articulatory codebook. 1 INTRODUCTION An articulatory description 1 may be preferable to an acoustic feature set because physical dynamic limitations of the vocal tract imply slowly changing parameters. Also, the close relationship between the articulatory and phonetic domains suggests that an articulatory parameter set could be a more ...

This paper evaluates continuous density hidden Markov models (CDHMM), dynamic time warping (DTW) and distortion-based vector quantisation (VQ) for speaker recognition, across incremental amounts of training data. In comparing VQ and CDHMMs for text-independent (TI) speaker recognition, it is shown that VQ performs better than an equivalent CDHMM with one training version, but is outperformed by the CDHMM when trained with ten training versions. In text-dependent (TD) experiments, a comparison of DTW, VQ and CDHMMs shows that DTW outperforms VQ and CDHMMs for sparse amounts of training data, but with more data, the performance of each model is indistinguishable. Further analysis shows TD to be superior to TI architecture for speaker recognition, and TD digit performance illustrates zero, 1 and 9 to be good discriminators. 1. INTRODUCTION Several analytical approaches have been applied to the task of speaker recognition, many of which originate in speech recognition. Dynamic time warpin...

The aim of the study described in this paper is to provide a thorough and quantitative assessment of the relative speaker discriminating properties of phonemes. A VQ codebook based approach to speaker modeling is used in conjunction with a phonetically hand-labelled database to produce phoneme rankings based on speaker verification scores. In broad groupings the nasals and vowels are found to provide the best speaker recognition performance, followed by the fricatives, affricates and approximants, with the stops providing the worst performance of all. A comparison at the individual phoneme level produces a more detailed ranking and of particular interest is the surprisingly good performance of the unvoiced fricative /s/. The ranking of phonemes is found to be largely unaffected by changes in experimental parameters such as the model size, the feature type and the speaker population.>

This paper examines the effects of additive Gaussian noise on the short-term cepstral analysis of speech. We identify three distinct modifications to the long-term statistics of the cepstrum that cause a gross mismatch after the addition of noise, namely: a mean shift, a change of variance and a distribution distorted from normal, with distinct bimodal characteristics. We assess the importance of each of these, and demonstrate the limitations of simple cepstral mappings. We then consider noise-masking, through the addition of a constant offset to the linear spectral estimates, which provides a feature space far more stable to changes in noise statistics. This leads to performance equivalent to that achieved by explicit modelling.>

Presents a face segmentation architecture using fuzzy inference. The head of a talker and two key structural features of his face (eyes and mouth) are located based on temporal and spatial information extracted from a head-and-shoulder image sequence. The architecture is modular and the segmentation uses a coarse-to-fine fuzzy reasoning strategy implemented across a three-level multi-resolution image pyramid. Results illustrating the performance of the system are given.>

this paper, the influences of such non-linearity and network topology on network performances are studied (see Figures 1 & 2). The three common classes of speech degradation, namely non-linear degradation introduced by CELP, additive noise including GWN (Gaussian white noise) and two real-world noises, and linear system degradation are considered. In addition, we also investigate whether or not the enhancer characteristics depend on the class of degradation.

This paper 1 discusses cepstral feature combinations via linear discriminant analysis (LDA) in the context of automatic speaker identification (ASI). Two static cepstral features are considered, namely standard MFCC and a sub-band filtered form derived via linear prediction known as RASTA-PLP. These two are compared along with their first order dynamic forms as both single and combined feature sets. LDA is shown to provide a useful means of combining (dissimilar) feature sets and permitting a direct trade-off between the number of coefficients and ASI performance, particularly when testing under noisy conditions. Also, the importance of pre-normalisation is demonstrated. It is shown that in the case of individual features, the two static forms give the best performance in clean conditions, with a cross-over to the two dynamic forms being better in the region of SNR=15dB. The LDA combination of the two static forms gives the best overall results under both clean and noisy conditions...

WITHINCLASSOPTIMIZATIONOFCEPSTRAFORSPEAKERRECOGNITIONJ.Thompson&J.S.MasonDepartmentofElectrical&ElectronicEngineering,UniversityCollegeofSwansea,SWANSEA,SA28PP,UKABSTRACTByidentifyinghighlysp eakersp eci casp ectsofcep-stralfeatures,thetaskofsp eakerrecognitioncanb ep otentiallysimpli ed.Theidenti cationpro cesscanb ep erformedb othphoneticallyandinthecepstraldomain.Thephoneticanalysisaimstodeterminethetemp oralasp ectsofutteranceswhichexhibitthehighestdegreeofsp eakersp eci city,whilecepstralanalysisexaminesindividualcepstrawithinthesetemp oraldivisions.Thispap eraimstocom-plementworkthathasalreadyb eenconductedonaphoneticbasis,byp erforminganalysisup onthein-dividualcepstralco ecientswithinutterances.Keywords:Sp eakerRecognition,Sp eci- city,Inter-sp eakerVariation,tra-sp eakaria-tion.1INTRODUCTIONInvestigationsintothephoneticasp ectsofsp eechsp eakerrecognitionarerep ortedbyEato ck[1 ],vandenHeuvel[2 ]andBonastre[3].Intheseinvestiga-tions,ithasb eenfoundthatsp eakersp eci cityvariesnoticeablyb etweendi erentphoneticsubgroups,butremainsmoreconstantwithinthem.Thesephonemesubgroupscanb ecrudelyrankedinthefollowingor-der:longvowels,nasals,shortfricatives,andplosives,wherelongvowelsexhibitthehighestdegreeofsp eakersp eci city,andplosivesthelowest.Thispap erexaminesthecepstralrepresentationsofutterances,inparticularthosesubgroupswhichex-hibitahighsp eakersp eci citysuchaslongvowelsandnasals.Moresp eci callytheaimistoidentifyindividualcepstralco ecientswhichcontainthema-jorityofthesp eakersp eci cinformation,andshowbyanalysisofcepstraldistributions,whsomesp eak-ersp erformsigni cantlyworsethanothers.Severalb ene tsfromidentifyingthemoste ectiveco ecientsexist.Firstlytherecognitionsystemcanb ereducedincomplexitybusingcepstrathatcon-tainahighdegreeofsp eakersp eci city.Secondly,p erhapsb etteroverallp erformancecanb eachievedby`liftering'thefeaturevectorspriortoclassi ca-tion[4],wherethecepstraweightingsarederiveddi-rectlyfromtheaverageoftheirsp eakersp eci cityforagroupofsp eakers.Exp erimentssimilarnaturearerep ortedonfrequencydomainfeaturesusingavariancemeasure[5 ]andmultiplefeaturesusingdy-namicprogramming[6 ][7].2CEPSTRALFEATURESThecepstralfeaturesconsideredherearegeneratedfromasp eechdatabaseusing24channelDFT-simulatedmel lter-bank[8 ],witha95%overlapb e-tweensuccessiveframes.This lterbankiswidelyusedandleadstothestandardformofmelcepstra.Thesp eechdatabaseconsistsof20sp eakers(10male,10female),utteringthelettersofalphab et(atoz)threetimes.Thesamplingrateofthedatais10KHz.Ofthe24p ossiblestandardcepstralfeaturesonlyMFCCs1-14areusedinordertoremovethema jor-ityofpitchinformation(whicmaexhibitamis-leadinglyhighsp eakersp eci citybutisprobablyanunreliableparameterinpracticalsp eakerrecognitionsystems[9 ]).Alsothisreduceddimensionfeaturesetisinverseariance,zeromeanweighted,onaglobalp o oledbasis.

The performance of dynamic features in automatic speaker recognition is examined. Second- and third-order regression analysis examining the performance of the associated feature sets independently, in combination, and in the presence of noise is included. It is shown that each regression order has a clear optimum. These are independent of the analysis order of the static feature from which the dynamic features are derived, and insensitive to low-level noise added to the test speech. It is also demonstrated that while the static feature gives the best individual performance, multiple linear combinations of feature sets based on regression analysis can reduce error rates.>

A speaker recognition system, using a modified form of feedforward neural network based on radial basis functions (RBFs), is presented. Each person to be recognized has his/her own neural model which is trained to recognise spectral feature vectors representative of his/her speech. Experimental results on a 40-speaker database indicate that the modified neural approach significantly outperforms both a standard multilayer perceptron and a vector quantization based system. The best performance for 4 digit test utterances is obtained from an RBF network with 384 RBF nodes in the hidden layer, given an 8% true talker rejection rate for a fixed 1% imposter acceptance rate. Additional advantages include a substantial reduction in training time over an MLP approach, and the ability to readily interpret the resulting model.>

An approach to speaker recognition based on feedforward neural models is investigated. Each person known to the system has a personalized neural net that is trained to be active for only that person's speech. By including speech from many people in the training data of each net this approach can directly model differences in people's speech. The chosen architecture and amount of training performed is shown to strongly affect the recognition performance. Large models with two hidden layers are shown to be inferior to models with only a single hidden layer and fewer weights. Recognition performance is shown to be comparable to that of a vector quantization approach based on personalized codebooks. The neural approach outperforms the codebook system for small model sizes but does slightly less well for larger models. A multitransputer implementation used in the training phase is described. Near linear speedup is obtained by splitting the training data to given independent subtasks, and a dynamic allocation scheme is used to assign these tasks to processors.>

Results of an experimental study and the optimization of features for a conventional vector-quantization codebook-based automatic speaker identification (ASI) system are presented. Standard LPC (linear predictive coding) and a perceptually weighted feature termed PLP (perceptually based linear prediction) are compared using appropriate distance measures, namely, the log-likelihood, and three cepstral variants: constant weighting, the robot-power-sum, and the inverse variance. PLP features combined with a weighted cepstral measure are found to be consistently the best in a number of different digit-independent ASI experiments. Results support the hypothesis that the higher orders of PLP (>5) contain significant speaker-specific information, with ASI performance improving rapidly up to order 8, and then far more slowly yet consistently up to order 16. A similar pattern is seen for codebook size, with fast improvements up to size 64, with more gradual gains thereafter.>

The idea of the semi-hidden Markov model (SHMM) is described, central to which is a modified training process. A modification to the conventional Baum-Welch algorithm is the kernel of the SHMM, where states are classified into types, reflecting fundamentally different speech signal characteristics. A preset supervisory function is introduced to the Baum-Welch algorithm and biases the training by reflecting the fitness of local signal characteristics to different state types. A simple example is described, using transient and quasi-stationary state types, which is found to be successful in E-set recognition.>