Dynamic visual features for audio-visual speaker verification

The cascading appearance-based (CAB) feature extraction technique has established itself as the state-of-the-art in extracting dynamic visual speech features for speech recognition. In this paper, we will focus on investigating the effectiveness of this technique for the related speaker verification application. By investigating the speaker verification ability of each stage of the cascade we will demonstrate that the same steps taken to reduce static speaker and environmental information for the visual speech recognition application also provide similar improvements for visual speaker recognition. A further study is conducted comparing synchronous HMM (SHMM) based fusion of CAB visual features and traditional perceptual linear predictive (PLP) acoustic features to show that higher complexity inherit in the SHMM approach does not appear to provide any improvement in the final audio-visual speaker verification system over simpler utterance level score fusion.

Fast and accurate phonetic spoken term detection

For the first time in human history, large volumes of spoken audio are being broadcast, made available on the internet, archived, and monitored for surveillance every day. New technologies are urgently required to unlock these vast and powerful stores of information. Spoken Term Detection (STD) systems provide access to speech collections by detecting individual occurrences of specified search terms. The aim of this work is to develop improved STD solutions based on phonetic indexing. In particular, this work aims to develop phonetic STD systems for applications that require open-vocabulary search, fast indexing and search speeds, and accurate term detection. Within this scope, novel contributions are made within two research themes, that is, accommodating phone recognition errors and, secondly, modelling uncertainty with probabilistic scores. A state-of-the-art Dynamic Match Lattice Spotting (DMLS) system is used to address the problem of accommodating phone recognition errors with approximate phone sequence matching. Extensive experimentation on the use of DMLS is carried out and a number of novel enhancements are developed that provide for faster indexing, faster search, and improved accuracy. Firstly, a novel comparison of methods for deriving a phone error cost model is presented to improve STD accuracy, resulting in up to a 33% improvement in the Figure of Merit. A method is also presented for drastically increasing the speed of DMLS search by at least an order of magnitude with no loss in search accuracy. An investigation is then presented of the effects of increasing indexing speed for DMLS, by using simpler modelling during phone decoding, with results highlighting the trade-off between indexing speed, search speed and search accuracy. The Figure of Merit is further improved by up to 25% using a novel proposal to utilise word-level language modelling during DMLS indexing. Analysis shows that this use of language modelling can, however, be unhelpful or even disadvantageous for terms with a very low language model probability. The DMLS approach to STD involves generating an index of phone sequences using phone recognition. An alternative approach to phonetic STD is also investigated that instead indexes probabilistic acoustic scores in the form of a posterior-feature matrix. A state-of-the-art system is described and its use for STD is explored through several experiments on spontaneous conversational telephone speech. A novel technique and framework is proposed for discriminatively training such a system to directly maximise the Figure of Merit. This results in a 13% improvement in the Figure of Merit on held-out data. The framework is also found to be particularly useful for index compression in conjunction with the proposed optimisation technique, providing for a substantial index compression factor in addition to an overall gain in the Figure of Merit. These contributions significantly advance the state-of-the-art in phonetic STD, by improving the utility of such systems in a wide range of applications.

Discriminative optimisation of the figure of merit for phonetic spoken term detection

This work proposes to improve spoken term detection (STD) accuracy by optimising the Figure of Merit (FOM). In this article, the index takes the form of phonetic posterior-feature matrix. Accuracy is improved by formulating STD as a discriminative training problem and directly optimising the FOM, through its use as an objective function to train a transformation of the index. The outcome of indexing is then a matrix of enhanced posterior-features that are directly tailored for the STD task. The technique is shown to improve the FOM by up to 13% on held-out data. Additional analysis explores the effect of the technique on phone recognition accuracy, examines the actual values of the learned transform, and demonstrates that using an extended training data set results in further improvement in the FOM.

Elements pour adapter les systemes de recherche d'information aux dyslexiques [Elements to adapt information retrieval systems to dyslexic users]

RÉSUMÉ. La prise en compte des troubles de la communication dans l’utilisation des systèmes de recherche d’information tels qu’on peut en trouver sur le Web est généralement réalisée par des interfaces utilisant des modalités n’impliquant pas la lecture et l’écriture. Peu d’applications existent pour aider l’utilisateur en difficulté dans la modalité textuelle. Nous proposons la prise en compte de la conscience phonologique pour assister l’utilisateur en difficulté d’écriture de requêtes (dysorthographie) ou de lecture de documents (dyslexie). En premier lieu un système de réécriture et d’interprétation des requêtes entrées au clavier par l’utilisateur est proposé : en s’appuyant sur les causes de la dysorthographie et sur les exemples à notre disposition, il est apparu qu’un système combinant une approche éditoriale (type correcteur orthographique) et une approche orale (système de transcription automatique) était plus approprié. En second lieu une méthode d’apprentissage automatique utilise des critères spécifiques , tels que la cohésion grapho-phonémique, pour estimer la lisibilité d’une phrase, puis d’un texte. ABSTRACT. Most applications intend to help disabled users in the information retrieval process by proposing non-textual modalities. This paper introduces specific parameters linked to phonological awareness in the textual modality. This will enhance the ability of systems to deal with orthographic issues and with the adaptation of results to the reader when for example the reader is dyslexic. We propose a phonology based sentence level rewriting system that combines spelling correction, speech synthesis and automatic speech recognition. This has been evaluated on a corpus of questions we get from dyslexic children. We propose a specific sentence readability measure that involves phonetic parameters such as grapho-phonemic cohesion. This has been learned on a corpus of reading time of sentences read by dyslexic children.

An actor-network research frame for analysing complex socio-technical situations

This paper presents a “research frame” which we have found useful in analyzing complex socio- technical situations. The research frame is based on aspects of actor-network theory: “interressment”, “enrollment”, “points of passage” and the “trial of strength”. Each of these aspects are described in turn, making clear their purpose in the overall research frame. Having established the research frame it is used to analyse two examples. First, the use of speech recognition technology is examined in two different contexts, showing how to apply the frame to compare and contrast current situations. Next, a current medical consultation context is described and the research frame is used to consider how it could change with innovative technology. In both examples, the research frame shows that the use of an artefact or technology must be considered together with the context in which it is used.

Speaker diarization : "who spoke when"

Speaker diarization is the process of annotating an input audio with information that attributes temporal regions of the audio signal to their respective sources, which may include both speech and non-speech events. For speech regions, the diarization system also specifies the locations of speaker boundaries and assign relative speaker labels to each homogeneous segment of speech. In short, speaker diarization systems effectively answer the question of ‘who spoke when’. There are several important applications for speaker diarization technology, such as facilitating speaker indexing systems to allow users to directly access the relevant segments of interest within a given audio, and assisting with other downstream processes such as summarizing and parsing. When combined with automatic speech recognition (ASR) systems, the metadata extracted from a speaker diarization system can provide complementary information for ASR transcripts including the location of speaker turns and relative speaker segment labels, making the transcripts more readable. Speaker diarization output can also be used to localize the instances of specific speakers to pool data for model adaptation, which in turn boosts transcription accuracies. Speaker diarization therefore plays an important role as a preliminary step in automatic transcription of audio data. The aim of this work is to improve the usefulness and practicality of speaker diarization technology, through the reduction of diarization error rates. In particular, this research is focused on the segmentation and clustering stages within a diarization system. Although particular emphasis is placed on the broadcast news audio domain and systems developed throughout this work are also trained and tested on broadcast news data, the techniques proposed in this dissertation are also applicable to other domains including telephone conversations and meetings audio. Three main research themes were pursued: heuristic rules for speaker segmentation, modelling uncertainty in speaker model estimates, and modelling uncertainty in eigenvoice speaker modelling. The use of heuristic approaches for the speaker segmentation task was first investigated, with emphasis placed on minimizing missed boundary detections. A set of heuristic rules was proposed, to govern the detection and heuristic selection of candidate speaker segment boundaries. A second pass, using the same heuristic algorithm with a smaller window, was also proposed with the aim of improving detection of boundaries around short speaker segments. Compared to single threshold based methods, the proposed heuristic approach was shown to provide improved segmentation performance, leading to a reduction in the overall diarization error rate. Methods to model the uncertainty in speaker model estimates were developed, to address the difficulties associated with making segmentation and clustering decisions with limited data in the speaker segments. The Bayes factor, derived specifically for multivariate Gaussian speaker modelling, was introduced to account for the uncertainty of the speaker model estimates. The use of the Bayes factor also enabled the incorporation of prior information regarding the audio to aid segmentation and clustering decisions. The idea of modelling uncertainty in speaker model estimates was also extended to the eigenvoice speaker modelling framework for the speaker clustering task. Building on the application of Bayesian approaches to the speaker diarization problem, the proposed approach takes into account the uncertainty associated with the explicit estimation of the speaker factors. The proposed decision criteria, based on Bayesian theory, was shown to generally outperform their non- Bayesian counterparts.

Robust automatic speaker linking and attribution

This research makes a major contribution which enables efficient searching and indexing of large archives of spoken audio based on speaker identity. It introduces a novel technique dubbed as “speaker attribution” which is the task of automatically determining ‘who spoke when?’ in recordings and then automatically linking the unique speaker identities within each recording across multiple recordings. The outcome of the research will also have significant impact in improving the performance of automatic speech recognition systems through the extracted speaker identities.

Visual front-end wars : Viola-Jones face detector vs Fourier Lucas-Kanade

The performance of visual speech recognition (VSR) systems are significantly influenced by the accuracy of the visual front-end. The current state-of-the-art VSR systems use off-the-shelf face detectors such as Viola- Jones (VJ) which has limited reliability for changes in illumination and head poses. For a VSR system to perform well under these conditions, an accurate visual front end is required. This is an important problem to be solved in many practical implementations of audio visual speech recognition systems, for example in automotive environments for an efficient human-vehicle computer interface. In this paper, we re-examine the current state-of-the-art VSR by comparing off-the-shelf face detectors with the recently developed Fourier Lucas-Kanade (FLK) image alignment technique. A variety of image alignment and visual speech recognition experiments are performed on a clean dataset as well as with a challenging automotive audio-visual speech dataset. Our results indicate that the FLK image alignment technique can significantly outperform off-the shelf face detectors, but requires frequent fine-tuning.

Acoustic adaptation in cross database audio visual SHMM training for phonetic spoken term detection

Visual information in the form of lip movements of the speaker has been shown to improve the performance of speech recognition and search applications. In our previous work, we proposed cross database training of synchronous hidden Markov models (SHMMs) to make use of external large and publicly available audio databases in addition to the relatively small given audio visual database. In this work, the cross database training approach is improved by performing an additional audio adaptation step, which enables audio visual SHMMs to benefit from audio observations of the external audio models before adding visual modality to them. The proposed approach outperforms the baseline cross database training approach in clean and noisy environments in terms of phone recognition accuracy as well as spoken term detection (STD) accuracy.

Incorporating visual information for spoken term detection

Spoken term detection (STD) is the task of looking up a spoken term in a large volume of speech segments. In order to provide fast search, speech segments are first indexed into an intermediate representation using speech recognition engines which provide multiple hypotheses for each speech segment. Approximate matching techniques are usually applied at the search stage to compensate the poor performance of automatic speech recognition engines during indexing. Recently, using visual information in addition to audio information has been shown to improve phone recognition performance, particularly in noisy environments. In this paper, we will make use of visual information in the form of lip movements of the speaker in indexing stage and will investigate its effect on STD performance. Particularly, we will investigate if gains in phone recognition accuracy will carry through the approximate matching stage to provide similar gains in the final audio-visual STD system over a traditional audio only approach. We will also investigate the effect of using visual information on STD performance in different noise environments.

Improving spoken term detection using complementary information

This research has made contributions to the area of spoken term detection (STD), defined as the process of finding all occurrences of a specified search term in a large collection of speech segments. The use of visual information in the form of lip movements of the speaker in addition to audio and the use of topic of the speech segments, and the expected frequency of words in the target speech domain, are proposed. By using these complementary information, improvement in the performance of STD has been achieved which enables efficient search of key words in large collection of multimedia documents.

A comparative study of speaker adaptation methods

For the problem of speaker adaptation in speech recognition, the performance depends on the availability of adaptation data. In this paper, we have compared several existing speaker adaptation methods, viz. maximum likelihood linear regression (MLLR), eigenvoice (EV), eigenspace-based MLLR (EMLLR), segmental eigenvoice (SEV) and hierarchical eigenvoice (HEV) based methods. We also develop a new method by modifying the existing HEV method for achieving further performance improvement in a limited available data scenario. In the sense of availability of adaptation data, the new modified HEV (MHEV) method is shown to perform better than all the existing methods throughout the range of operation except the case of MLLR at the availability of more adaptation data.

A Savitzky-Golay Filtering Perspective of Dynamic Feature Computation

We address the classical problem of delta feature computation, and interpret the operation involved in terms of Savitzky- Golay (SG) filtering. Features such as themel-frequency cepstral coefficients (MFCCs), obtained based on short-time spectra of the speech signal, are commonly used in speech recognition tasks. In order to incorporate the dynamics of speech, auxiliary delta and delta-delta features, which are computed as temporal derivatives of the original features, are used. Typically, the delta features are computed in a smooth fashion using local least-squares (LS) polynomial fitting on each feature vector component trajectory. In the light of the original work of Savitzky and Golay, and a recent article by Schafer in IEEE Signal Processing Magazine, we interpret the dynamic feature vector computation for arbitrary derivative orders as SG filtering with a fixed impulse response. This filtering equivalence brings in significantly lower latency with no loss in accuracy, as validated by results on a TIMIT phoneme recognition task. The SG filters involved in dynamic parameter computation can be viewed as modulation filters, proposed by Hermansky.

Language model combination and adaptation using weighted finite state transducers

In speech recognition systems language model (LMs) are often constructed by training and combining multiple n-gram models. They can be either used to represent different genres or tasks found in diverse text sources, or capture stochastic properties of different linguistic symbol sequences, for example, syllables and words. Unsupervised LM adaptation may also be used to further improve robustness to varying styles or tasks. When using these techniques, extensive software changes are often required. In this paper an alternative and more general approach based on weighted finite state transducers (WFSTs) is investigated for LM combination and adaptation. As it is entirely based on well-defined WFST operations, minimum change to decoding tools is needed. A wide range of LM combination configurations can be flexibly supported. An efficient on-the-fly WFST decoding algorithm is also proposed. Significant error rate gains of 7.3% relative were obtained on a state-of-the-art broadcast audio recognition task using a history dependently adapted multi-level LM modelling both syllable and word sequences. ©2010 IEEE.

Language model cross adaptation for LVCSR system combination

State-of-the-art large vocabulary continuous speech recognition (LVCSR) systems often combine outputs from multiple subsystems developed at different sites. Cross system adaptation can be used as an alternative to direct hypothesis level combination schemes such as ROVER. In normal cross adaptation it is assumed that useful diversity among systems exists only at acoustic level. However, complimentary features among complex LVCSR systems also manifest themselves in other layers of modelling hierarchy, e.g., subword and word level. It is thus interesting to also cross adapt language models (LM) to capture them. In this paper cross adaptation of multi-level LMs modelling both syllable and word sequences was investigated to improve LVCSR system combination. Significant error rate gains up to 6.7% rel. were obtained over ROVER and acoustic model only cross adaptation when combining 13 Chinese LVCSR subsystems used in the 2010 DARPA GALE evaluation. © 2010 ISCA.