Publications by Steve Renals

We investigate the use of cross-lingual acoustic data to initialise deep neural network (DNN) acoustic models by means of unsupervised restricted Boltzmann machine (RBM) pretraining. DNNs for German are pretrained using one or all of German, Portuguese, Spanish and Swedish. The DNNs are used in a tandem configuration, where the network outputs are used as features for a hidden Markov model (HMM) whose emission densities are modeled by Gaussian mixture models (GMMs), as well as in a hybrid configuration, where the network outputs are used as the HMM state likelihoods. The experiments show that unsupervised pretraining is more crucial for the hybrid setups, particularly with limited amounts of transcribed training data. More importantly, unsupervised pretraining is shown to be language-independent.

We describe our work on developing a speech recognition system for multi-genre media archives. The high diversity of the data makes this a challenging recognition task, which may benefit from systems trained on a combination of in-domain and out-of-domain data. Working with tandem HMMs, we present Multi-level Adaptive Networks (MLAN), a novel technique for incorporating information from out-of-domain posterior features using deep neural networks. We show that it provides a substantial reduction in WER over other systems, with relative WER reductions of 15% over a PLP baseline, 9% over in-domain tandem features and 8% over the best out-of-domain tandem features.

Speech sound disorders (SSD) are the most common communication impairment in childhood, and can hamper social development and learning. Current speech therapy interventions rely predominantly on the auditory skills of the child, as little technology is available to assist in diagnosis and therapy of SSDs. Realtime visualisation of tongue movements has the potential to bring enormous benefit to speech therapy. Ultrasound scanning offers this possibility, although its display may be hard to interpret. Our ultimate goal is to exploit ultrasound to track tongue movement, while displaying a simplified, diagrammatic vocal tract that is easier for the user to interpret. In this paper, we outline a general approach to this problem, combining a latent space model with a dimensionality reducing model of vocal tract shapes. We assess the feasibility of this approach using magnetic resonance imaging (MRI) scans to train a model of vocal tract shapes, which is animated using electromagnetic articulography (EMA) data from the same speaker.

Keywords: Ultrasound, speech therapy, vocal tract visualisation

We implement two deep architectures for the acoustic-articulatory inversion mapping problem: a deep neural network and a deep trajectory mixture density network. We find that in both cases, deep architectures produce more accurate predictions than shallow architectures and that this is due to the higher expressive capability of a deep model and not a consequence of adding more adjustable parameters. We also find that a deep trajectory mixture density network is able to obtain better inversion accuracies than smoothing the results of a deep neural network. Our best model obtained an average root mean square error of 0.885 mm on the MNGU0 test dataset.

Keywords: Articulatory inversion, deep neural network, deep belief network, deep regression network, pretraining

This paper concerns cross-lingual acoustic modeling in the case when there are limited target language resources. We build on an approach in which a subspace Gaussian mixture model (SGMM) is adapted to the target language by reusing the globally shared parameters estimated from out-of-language training data. In current cross-lingual systems, these parameters are fixed when training the target system, which can give rise to a mismatch between the source and target systems. We investigate a maximum a posteriori (MAP) adaptation approach to alleviate the potential mismatch. In particular, we focus on the adaptation of phonetic subspace parameters using a matrix variate Gaussian prior distribution. Experiments on the GlobalPhone corpus using the MAP adaptation approach results in word error rate reductions, compared with the cross-lingual baseline systems and systems updated using maximum likelihood, for training conditions with 1 hour and 5 hours of target language data.

Keywords: Subspace Gaussian Mixture Model, Maximum a Posteriori Adaptation, Cross-lingual Speech Recognition

Common noise compensation techniques use vector Taylor series (VTS) to approximate the mismatch function. Recent work shows that the approximation accuracy may be improved by sampling. One such sampling technique is the unscented transform (UT), which draws samples deterministically from clean speech and noise model to derive the noise corrupted speech parameters. This paper applies UT to noise compensation of the subspace Gaussian mixture model (SGMM). Since UT requires relatively smaller number of samples for accurate estimation, it has significantly lower computational cost compared to other random sampling techniques. However, the number of surface Gaussians in an SGMM is typically very large, making the direct application of UT, for compensating individual Gaussian components, computationally impractical. In this paper, we avoid the computational burden by employing UT in the framework of joint uncertainty decoding (JUD), which groups all the Gaussian components into small number of classes, sharing the compensation parameters by class. We evaluate the JUD-UT technique for an SGMM system using the Aurora 4 corpus. Experimental results indicate that UT can lead to increased accuracy compared to VTS approximation if the JUD phase factor is untuned, and to similar accuracy if the phase factor is tuned empirically

Keywords: noise compensation, SGMM, JUD, UT

Joint uncertainty decoding (JUD) is an effective model-based noise compensation technique for conventional Gaussian mixture model (GMM) based speech recognition systems. In this paper, we apply JUD to subspace Gaussian mixture model (SGMM) based acoustic models. The total number of Gaussians in the SGMM acoustic model is usually much larger than for conventional GMMs, which limits the application of approaches which explicitly compensate each Gaussian, such as vector Taylor series (VTS). However, by clustering the Gaussian components into a number of regression classes, JUD-based noise compensation can be successfully applied to SGMM systems. We evaluate the JUD/SGMM technique using the Aurora 4 corpus, and the experimental results indicated that it is more accurate than conventional GMM-based systems using either VTS or JUD noise compensation.

Keywords: acoustic modelling, noise compensation, SGMM, JUD

In this work, we implement a deep belief network for the acoustic-articulatory inversion mapping problem. We find that adding up to 3 hidden-layers improves inversion accuracy. We also show that this improvement is due to the higher ex- pressive capability of a deep model and not a consequence of adding more adjustable parameters. Additionally, we show unsupervised pretraining of the sys- tem improves its performance in all cases, even for a 1 hidden-layer model. Our implementation obtained an average root mean square error of 0.95 mm on the MNGU0 test dataset, beating all previously published results.

A major factor which causes a deterioration in speech quality in HMM-based speech synthesis is the use of a simple delta pulse signal to generate the excitation of voiced speech. This paper sets out a new approach to using an acoustic glottal source model in HMM-based synthesisers instead of the traditional pulse signal. The goal is to improve speech quality and to better model and transform voice characteristics. We have found the new method decreases buzziness and also improves prosodic modelling. A perceptual evaluation has supported this finding by showing a 55.6 preference for the new system, as against the baseline. This improvement, while not being as significant as we had initially expected, does encourage us to work on developing the proposed speech synthesiser further.

Subspace Gaussian mixture models (SGMMs) provide a compact representation of the Gaussian parameters in an acoustic model, but may still suffer from over-fitting with insufficient training data. In this letter, the SGMM state parameters are estimated using a penalized maximum-likelihood objective, based on ₁ and ₂ regularization, as well as their combination, referred to as the elastic net, for robust model estimation. Experiments on the 5000-word Wall Street Journal transcription task show word error rate reduction and improved model robustness with regularization.

We present the Ambient Spotlight system for personal meeting capture based on a portable USB microphone array and a laptop. The system combined distant speech recognition and content linking with personal productivity tools, and enables recognised meeting recordings to be integrated with desktop search, calender, and email.

We investigate cross-lingual acoustic modelling for low resource languages using the subspace Gaussian mixture model (SGMM). We assume the presence of acoustic models trained on multiple source languages, and use the global subspace parameters from those models for improved modelling in a target language with limited amounts of transcribed speech. Experiments on the GlobalPhone corpus using Spanish, Portuguese, and Swedish as source languages and German as target language (with 1 hour and 5 hours of transcribed audio) show that multilingually trained SGMM shared parameters result in lower word error rates (WERs) than using those from a single source language. We also show that regularizing the estimation of the SGMM state vectors by penalizing their ₁-norm help to overcome numerical instabilities and lead to lower WER.

Control over voice quality, e.g. breathy and tense voice, is important for speech synthesis applications. For example, transformations can be used to modify aspects of the voice re- lated to speaker's identity and to improve expressiveness. How- ever, it is hard to modify voice characteristics of the synthetic speech, without degrading speech quality. State-of-the-art sta- tistical speech synthesisers, in particular, do not typically al- low control over parameters of the glottal source, which are strongly correlated with voice quality. Consequently, the con- trol of voice characteristics in these systems is limited. In con- trast, the HMM-based speech synthesiser proposed in this paper uses an acoustic glottal source model. The system passes the glottal signal through a whitening filter to obtain the excitation of voiced sounds. This technique, called glottal post-filtering, allows to transform voice characteristics of the synthetic speech by modifying the source model parameters. We evaluated the proposed synthesiser in a perceptual ex- periment, in terms of speech naturalness, intelligibility, and similarity to the original speaker's voice. The results show that it performed as well as a HMM-based synthesiser, which generates the speech signal with a commonly used high-quality speech vocoder.

Keywords: HMM-based speech synthesis, voice quality, glottal post-filter

Linear regression based speaker adaptation approaches can improve Automatic Speech Recognition (ASR) accuracy significantly for a target speaker. However, when the available adaptation data is limited to a few seconds, the accuracy of the speaker adapted models is often worse compared with speaker independent models. In this paper, we propose an approach to select a set of reference speakers acoustically close to the target speaker whose data can be used to augment the adaptation data. To determine the acoustic similarity of two speakers, we propose a distance metric based on transforming sample points in the acoustic space with the regression matrices of the two speakers. We show the validity of this approach through a speaker identification task. ASR results on SCOTUS and AMI corpora with limited adaptation data of 10 to 15 seconds augmented by data from selected reference speakers show a significant improvement in Word Error Rate over speaker independent and speaker adapted models.

In this paper, the design, implementation and testing of a digital microphone array is presented. The array uses digital MEMS microphones which integrate the microphone, amplifier and analogue to digital converter on a single chip in place of the analogue microphones and external audio interfaces currently used. The device has the potential to be smaller, cheaper and more flexible than typical analogue arrays, however the effect on speech recognition performance of using digital microphones is as yet unknown. In order to evaluate the effect, an analogue array and the new digital array are used to simultaneously record test data for a speech recognition experiment. Initial results employing no adaptation show that performance using the digital array is significantly worse (14% absolute WER) than the analogue device. Subsequent experiments using MLLR and CMLLR channel adaptation reduce this gap, and employing MLLR for both channel and speaker adaptation reduces the difference between the arrays to 4.5% absolute WER.

This paper is about interpreting human communication in meetings using audio, video and other signals. Automatic meeting recognition and understanding is extremely challenging, since communication in a meeting is spontaneous and conversational, and involves multiple speakers and multiple modalities. This leads to a number of significant research problems in signal processing, in speech recognition, and in discourse interpretation, taking account of both individual and group behaviours. Addressing these problems requires an interdisciplinary effort. In this paper, I discuss the capture and annotation of multimodal meeting recordings - resulting in the AMI meeting corpus - and how we have built on this to develop techniques and applications for the recognition and interpretation of meetings.

It has recently become possible to record any small meeting using a laptop equipped with a plug-and-play USB microphone array. We show the potential for such recordings in a personal aid that allows project managers to record their meetings and, when reviewing them afterwards through a standard calendar interface, to find relevant documents on their computer. This interface is intended to supplement or replace the textual searches that managers typically perform. The prototype, which relies on meeting speech recognition and topic segmentation, formulates and runs desktop search queries in order to present its results.

Traditional n-gram language models are widely used in state-of-the-art large vocabulary speech recognition systems. This simple model suffers from some limitations, such as overfitting of maximum-likelihood estimation and the lack of rich contextual knowledge sources. In this paper, we exploit a hierarchical Bayesian interpretation for language modeling, based on a nonparametric prior called the Pitman-Yor process. This offers a principled approach to language model smoothing, embedding the power-law distribution for natural language. Experiments on the recognition of conversational speech in multiparty meetings demonstrate that by using hierarchical Bayesian language models, we are able to achieve significant reductions in perplexity and word error rate.

Keywords: AMI corpus , conversational speech recognition , hierarchical Bayesian model , language model (LM) , meetings , smoothing

We present an approximation to the Bayesian hierarchical Pitman-Yor process language model which maintains the power law distribution over word tokens, while not requiring a computationally expensive approximate inference process. This approximation, which we term power law discounting, has a similar computational complexity to interpolated and modified Kneser-Ney smoothing. We performed experiments on meeting transcription using the NIST RT06s evaluation data and the AMI corpus, with a vocabulary of 50,000 words and a language model training set of up to 211 million words. Our results indicate that power law discounting results in statistically significant reductions in perplexity and word error rate compared to both interpolated and modified Kneser-Ney smoothing, while producing similar results to the hierarchical Pitman-Yor process language model.

Microtask platforms such as Amazon Mechanical Turk (AMT) are increasingly used to create speech and language resources. AMT in particular allows researchers to quickly recruit a large number of fairly demographically diverse participants. In this study, we investigated whether AMT can be used for comparing the intelligibility of speech synthesis systems. We conducted two experiments in the lab and via AMT, one comparing US English diphone to US English speaker-adaptive HTS synthesis and one comparing UK English unit selection to UK English speaker-dependent HTS synthesis. While AMT word error rates were worse than lab error rates, AMT results were more sensitive to relative differences between systems. This is mainly due to the larger number of listeners. Boxplots and multilevel modelling allowed us to identify listeners who performed particularly badly, while thresholding was sufficient to eliminate rogue workers. We conclude that AMT is a viable platform for synthetic speech intelligibility comparisons.

With ageing, human voices undergo several changes which are typically characterized by increased hoarseness and changes in articulation patterns. In this study, we have examined the effect on Automatic Speech Recognition (ASR) and found that the Word Error Rates (WER) on older voices is about 9% absolute higher compared to those of adult voices. Subsequently, we compared several voice source parameters including fundamental frequency, jitter, shimmer, harmonicity and cepstral peak prominence of adult and older males. Several of these parameters show statistically significant difference for the two groups. However, artificially increasing jitter and shimmer measures do not effect the ASR accuracies significantly. Artificially lowering the fundamental frequency degrades the ASR performance marginally but this drop in performance can be overcome to some extent using Vocal Tract Length Normalisation (VTLN). Overall, we observe that the changes in the voice source parameters do not have a significant impact on ASR performance. Comparison of the likelihood scores of all the phonemes for the two age groups show that there is a systematic mismatch in the acoustic space of the two age groups. Comparison of the phoneme recognition rates show that mid vowels, nasals and phonemes that depend on the ability to create constrictions with tongue tip for articulation are more affected by ageing than other phonemes.

The EPSRC‐funded Edinburgh Speech Production is built around two synchronized Carstens AG500 electromagnetic articulographs (EMAs) in order to capture articulatory∕acoustic data from spontaneous dialogue. An initial articulatory corpus was designed with two aims. The first was to elicit a range of speech styles∕registers from speakers, and therefore provide an alternative to fully scripted corpora. The second was to extend the corpus beyond monologue, by using tasks that promote natural discourse and interaction. A subsidiary driver was to use dialects from outwith North America: dialogues paired up a Scottish English and a Southern British English speaker. Tasks. Monologue: Story reading of “Comma Gets a Cure” [Honorof et al. (2000)], lexical sets [Wells (1982)], spontaneous story telling, diadochokinetic tasks. Dialogue: Map tasks [Anderson et al. (1991)], “Spot the Difference” picture tasks [Bradlow et al. (2007)], story‐recall. Shadowing of the spontaneous story telling by the second participant. Each dialogue session includes approximately 30 min of speech, and there are acoustics‐only baseline materials. We will introduce the corpus and highlight the role of articulatory production data in helping provide a fuller understanding of various spontaneous speech phenomena by presenting examples of naturally occurring covert speech errors, accent accommodation, turn taking negotiation, and shadowing.

The Ambient Spotlight is a prototype system based on personal meeting capture using a laptop and a portable microphone array. The system automatically recognises and structures the meeting content using automatic speech recognition, topic segmentation and extractive summarisation. The recognised speech in the meeting is used to construct queries to automatically link meeting segments to other relevant material, both multimodal and textual. The interface to the system is constructed around a standard calendar interface, and it is integrated with the laptop's standard indexing, search and retrieval.

When deciding whether to adapt relevant aspects of the system to the particular needs of older users, spoken dialogue systems often rely on automatic detection of chronological age. In this paper, we show that vocal ageing as measured by acoustic features is an unreliable indicator of the need for adaptation. Simple lexical features greatly improve the prediction of both relevant aspects of cognition and interactions style. Lexical features also boost age group prediction. We suggest that adaptation should be based on observed behaviour, not on chronological age, unless it is not feasible to build classifiers for relevant adaptation decisions.

The Hierarchical Pitman Yor Process Language Model (HPYLM) is a Bayesian language model based on a non-parametric prior, the Pitman-Yor Process. It has been demonstrated, both theoretically and practically, that the HPYLM can provide better smoothing for language modeling, compared with state-of-the-art approaches such as interpolated Kneser-Ney and modified Kneser-Ney smoothing. However, estimation of Bayesian language models is expensive in terms of both computation time and memory; the inference is approximate and requires a number of iterations to converge. In this paper, we present a parallel training algorithm for the HPYLM, which enables the approach to be applied in the context of automatic speech recognition, using large training corpora with large vocabularies. We demonstrate the effectiveness of the proposed algorithm by estimating language models from corpora for meeting transcription containing over 200 million words, and observe significant reductions in perplexity and word error rate.

A major cause of degradation of speech quality in HMM-based speech synthesis is the use of a simple delta pulse signal to generate the excitation of voiced speech. This paper describes a new approach to using an acoustic glottal source model in HMM-based synthesisers. The goal is to improve speech quality and parametric flexibility to better model and transform voice characteristics.

In this work we describe a large-scale extrinsic evaluation of automatic speech summarization technologies for meeting speech. The particular task is a decision audit, wherein a user must satisfy a complex information need, navigating several meetings in order to gain an understanding of how and why a given decision was made. We compare the usefulness of extractive and abstractive technologies in satisfying this information need, and assess the impact of automatic speech recognition (ASR) errors on user performance. We employ several evaluation methods for participant performance, including post-questionnaire data, human subjective and objective judgments, and a detailed analysis of participant browsing behavior. We find that while ASR errors affect user satisfaction on an information retrieval task, users can adapt their browsing behavior to complete the task satisfactorily. Results also indicate that users consider extractive summaries to be intuitive and useful tools for browsing multimodal meeting data. We discuss areas in which automatic summarization techniques can be improved in comparison with gold-standard meeting abstracts.

Although older people are an important user group for smart environments, there has been relatively little work on adapting natural language interfaces to their requirements. In this paper, we focus on a particularly thorny problem: processing speech input from older users. Our experiments on the MATCH corpus show clearly that we need age-specific adaptation in order to recognize older users' speech reliably. Language models need to cover typical interaction patterns of older people, and acoustic models need to accommodate older voices. Further research is needed into intelligent adaptation techniques that will allow existing large, robust systems to be adapted with relatively small amounts of in-domain, age appropriate data. In addition, older users need to be supported with adequate strategies for handling speech recognition errors.

We describe an evaluation of spoken dialogue strategies designed using hierarchical reinforcement learning agents. The dialogue strategies were learnt in a simulated environment and tested in a laboratory setting with 32 users. These dialogues were used to evaluate three types of machine dialogue behaviour: hand-coded, fully-learnt and semi-learnt. These experiments also served to evaluate the realism of simulated dialogues using two proposed metrics contrasted with ‘Precision-Recall’. The learnt dialogue behaviours used the Semi-Markov Decision Process (SMDP) model, and we report the first evaluation of this model in a realistic conversational environment. Experimental results in the travel planning domain provide evidence to support the following claims: (a) hierarchical semi-learnt dialogue agents are a better alternative (with higher overall performance) than deterministic or fully-learnt behaviour; (b) spoken dialogue strategies learnt with highly coherent user behaviour and conservative recognition error rates (keyword error rate of 20%) can outperform a reasonable hand-coded strategy; and (c) hierarchical reinforcement learning dialogue agents are feasible and promising for the (semi) automatic design of optimized dialogue behaviours in larger-scale systems.

This paper describes a speaker-adaptive HMM-based speech synthesis system. The new system, called “HTS-2007,” employs speaker adaptation (CSMAPLR+MAP), feature-space adaptive training, mixed-gender modeling, and full-covariance modeling using CSMAPLR transforms, in addition to several other techniques that have proved effective in our previous systems. Subjective evaluation results show that the new system generates significantly better quality synthetic speech than speaker-dependent approaches with realistic amounts of speech data, and that it bears comparison with speaker-dependent approaches even when large amounts of speech data are available. In addition, a comparison study with several speech synthesis techniques shows the new system is very robust: It is able to build voices from less-than-ideal speech data and synthesize good-quality speech even for out-of-domain sentences.

Acoustic modeling based on hidden Markov models (HMMs) is employed by state-of-the-art stochastic speech recognition systems. Although HMMs are a natural choice to warp the time axis and model the temporal phenomena in the speech signal, their conditional independence properties limit their ability to model spectral phenomena well. In this paper, a new acoustic modeling paradigm based on augmented conditional random fields (ACRFs) is investigated and developed. This paradigm addresses some limitations of HMMs while maintaining many of the aspects which have made them successful. In particular, the acoustic modeling problem is reformulated in a data driven, sparse, augmented space to increase discrimination. Acoustic context modeling is explicitly integrated to handle the sequential phenomena of the speech signal. We present an efficient framework for estimating these models that ensures scalability and generality. In the TIMIT phone recognition task, a phone error rate of 23.0% was recorded on the full test set, a significant improvement over comparable HMM-based systems.

We continue our previous work on the modeling of topic and role information from multiparty meetings using a hierarchical Dirichlet process (HDP), in the context of language model adaptation. In this paper we focus on three problems: 1) an empirical analysis of the HDP as a nonparametric topic model; 2) the mismatch problem of vocabularies of the baseline n-gram model and the HDP; and 3) an automatic speech recognition experiment to further verify the effectiveness of our adaptation framework. Experiments on a large meeting corpus of more than 70 hours speech data show consistent and significant improvements in terms of word error rate for language model adaptation based on the topic and role information.

We present a method for predicting the midsagittal tongue contour from the locations of a few landmarks (metal pellets) on the tongue surface, as used in articulatory databases such as MOCHA and the Wisconsin XRDB. Our method learns a mapping using ground-truth tongue contours derived from ultrasound data and drastically improves over spline interpolation. We also determine the optimal locations of the landmarks, and the number of landmarks required to achieve a desired prediction error: 3-4 landmarks are enough to achieve 0.3-0.2 mm error per point on the tongue.

This paper presents a method to control the characteristics of synthetic speech flexibly by integrating articulatory features into a Hidden Markov Model (HMM)-based parametric speech synthesis system. In contrast to model adaptation and interpolation approaches for speaking style control, this method is driven by phonetic knowledge, and target speech samples are not required. The joint distribution of parallel acoustic and articulatory features considering cross-stream feature dependency is estimated. At synthesis time, acoustic and articulatory features are generated simultaneously based on the maximum-likelihood criterion. The synthetic speech can be controlled flexibly by modifying the generated articulatory features according to arbitrary phonetic rules in the parameter generation process. Our experiments show that the proposed method is effective in both changing the overall character of synthesized speech and in controlling the quality of a specific vowel.

In this paper we introduce our attempts to incorporate the participant role information in multiparty meetings for document modeling using the hierarchical Dirichlet process. The perplexity and automatic speech recognition results demonstrate that the participant role information is a promising prior knowledge source to be combined with language models for automatic speech recognition and interaction modeling for multiparty meetings.

The AMI and AMIDA projects are collaborative EU projects concerned with the automatic recognition and interpretation of multiparty meetings. This paper provides an overview of the advances we have made in these projects with a particular focus on the multimodal recording infrastructure, the publicly available AMI corpus of annotated meeting recordings, and the speech recognition framework that we have developed for this domain.

Keywords: AMI corpus; Meetings; evaluation; speech recognition

This paper presents the results of a longitudinal study of ASR performance on ageing voices. Experiments were conducted on the audio recordings of the proceedings of the Supreme Court Of The United States (SCOTUS). Results show that the Automatic Speech Recognition (ASR) Word Error Rates (WERs) for elderly voices are significantly higher than those of adult voices. The word error rate increases gradually as the age of the elderly speakers increase. Use of maximum likelihood linear regression (MLLR) based speaker adaptation on ageing voices improves the WER though the performance is still considerably lower compared to adult voices. Speaker adaptation however reduces the increase in WER with age during old age.

In this letter, we introduce an hidden Markov model (HMM)-based inversion system to recovery articulatory movements from speech acoustics. Trajectory HMMs are used as generative models for modelling articulatory data. Experiments on the MOCHA-TIMIT corpus indicate that the jointly trained acoustic-articulatory models are more accurate (lower RMS error) than the separately trained ones, and that trajectory HMM training results in greater accuracy compared with conventional maximum likelihood HMM training. Moreover, the system has the ability to synthesize articulatory movements directly from a textual representation.

In this work we describe a large-scale extrinsic evaluation of automatic speech summarization technologies for meeting speech. The particular task is a decision audit, wherein a user must satisfy a complex information need, navigating several meetings in order to gain an understanding of how and why a given decision was made. We compare the usefulness of extractive and abstractive technologies in satisfying this information need, and assess the impact of automatic speech recognition (ASR) errors on user performance. We employ several evaluation methods for participant performance, including post-questionnaire data, human subjective and objective judgments, and an analysis of participant browsing behaviour.

We present a method for detecting action items in spontaneous meeting speech. Using a supervised approach incorporating prosodic, lexical and structural features, we can classify such items with a high degree of accuracy. We also examine how well various feature subclasses can perform this task on their own.

In this paper we investigate the combination of complementary acoustic feature streams in large vocabulary continuous speech recognition (LVCSR). We have explored the use of acoustic features obtained using a pitch-synchronous analysis, STRAIGHT, in combination with conventional features such as mel frequency cepstral coefficients. Pitch-synchronous acoustic features are of particular interest when used with vocal tract length normalisation (VTLN) which is known to be affected by the fundamental frequency. We have combined these spectral representations directly at the acoustic feature level using heteroscedastic linear discriminant analysis (HLDA) and at the system level using ROVER. We evaluated this approach on three LVCSR tasks: dictated newspaper text (WSJCAM0), conversational telephone speech (CTS), and multiparty meeting transcription. The CTS and meeting transcription experiments were both evaluated using standard NIST test sets and evaluation protocols. Our results indicate that combining conventional and pitch-synchronous acoustic feature sets using HLDA results in a consistent, significant decrease in word error rate across all three tasks. Combining at the system level using ROVER resulted in a further significant decrease in word error rate.

This paper presents a fully automatic news skimming system which takes a broadcast news audio stream and provides the user with the segmented, structured and highlighted transcript. This constitutes a system with three different, cascading stages: converting the audio stream to text using an automatic speech recogniser, segmenting into utterances and stories and finally determining which utterance should be highlighted using a saliency score. Each stage must operate on the erroneous output from the previous stage in the system; an effect which is naturally amplified as the data progresses through the processing stages. We present a large corpus of transcribed broadcast news data enabling us to investigate to which degree information worth highlighting survives this cascading of processes. Both extrinsic and intrinsic experimental results indicate that mistakes in the story boundary detection has a strong impact on the quality of highlights, whereas erroneous utterance boundaries cause only minor problems. Further, the difference in transcription quality does not affect the overall performance greatly.

In this paper, we address the modeling of topic and role information in multiparty meetings, via a nonparametric Bayesian model called the hierarchical Dirichlet process. This model provides a powerful solution to topic modeling and a flexible framework for the incorporation of other cues such as speaker role information. We present our modeling framework for topic and role on the AMI Meeting Corpus, and illustrate the effectiveness of the approach in the context of adapting a baseline language model in a large-vocabulary automatic speech recognition system for multiparty meetings. The adapted LM produces significant improvements in terms of both perplexity and word error rate.

This paper is about the extractive summarization of meeting speech, using the ICSI and AMI corpora. In the first set of experiments we use prosodic, lexical, structural and speaker-related features to select the most informative dialogue acts from each meeting, with the hypothesis being that such a rich mixture of features will yield the best results. In the second part, we present an approach in which the identification of “meta-comments” is used to create more informative summaries that provide an increased level of abstraction. We find that the inclusion of these meta comments improves summarization performance according to several evaluation metrics.

We have investigated the use of a pitch adaptive spectral representation on large vocabulary speech recognition, in conjunction with speaker normalisation techniques. We have compared the effect of a smoothed spectrogram to the pitch adaptive spectral analysis by decoupling these two components of STRAIGHT. Experiments performed on a large vocabulary meeting speech recognition task highlight the importance of combining a pitch adaptive spectral representation with a conventional fixed window spectral analysis. We found evidence that STRAIGHT pitch adaptive features are more speaker independent than conventional MFCCs without pitch adaptation, thus they also provide better performances when combined using feature combination techniques such as Heteroscedastic Linear Discriminant Analysis.

This paper is concerned with the automatic recognition of dialogue acts (DAs) in multiparty conversational speech. We present a joint generative model for DA recognition in which segmentation and classification of DAs are carried out in parallel. Our approach to DA recognition is based on a switching dynamic Bayesian network (DBN) architecture. This generative approach models a set of features, related to lexical content and prosody, and incorporates a weighted interpolated factored language model. The switching DBN coordinates the recognition process by integrating the component models. The factored language model, which is estimated from multiple conversational data corpora, is used in conjunction with additional task-specific language models. In conjunction with this joint generative model, we have also investigated the use of a discriminative approach, based on conditional random fields, to perform a reclassification of the segmented DAs. We have carried out experiments on the AMI corpus of multimodal meeting recordings, using both manually transcribed speech, and the output of an automatic speech recognizer, and using different configurations of the generative model. Our results indicate that the system performs well both on reference and fully automatic transcriptions. A further significant improvement in recognition accuracy is obtained by the application of the discriminative reranking approach based on conditional random fields.

The AMI and AMIDA projects are concerned with the recognition and interpretation of multiparty (face-to-face and remote) meetings. Within these projects we have developed the following: (1) an infrastructure for recording meetings using multiple microphones and cameras; (2) a one hundred hour, manually annotated meeting corpus; (3) a number of techniques for indexing, and summarizing of meeting videos using automatic speech recognition and computer vision, and (4) a extensible framework for browsing, and searching of meeting videos. We give an overview of the various techniques developed in AMI (mainly involving face-to-face meetings), their integration into our meeting browser framework, and future plans for AMIDA (Augmented Multiparty Interaction with Distant Access), the follow-up project to AMI. Technical and business information related to these two projects can be found at www.amiproject.org, respectively on the Scientific and Business portals.

In this paper we investigate the application of a novel technique for language modeling - a hierarchical Bayesian language model (LM) based on the Pitman-Yor process - on automatic speech recognition (ASR) for multiparty meetings. The hierarchical Pitman-Yor language model (HPYLM), which was originally proposed in the machine learning field, provides a Bayesian interpretation to language modeling. An approximation to the HPYLM recovers the exact formulation of the interpolated Kneser-Ney smoothing method in n-gram models. This paper focuses on the application and scalability of HPYLM on a practical large vocabulary ASR system. Experimental results on NIST RT06s evaluation meeting data verify that HPYLM is a competitive and promising language modeling technique, which consistently performs better than interpolated Kneser-Ney and modified Kneser-Ney n-gram LMs in terms of both perplexity (PPL) and word error rate (WER).

For constructing a speech synthesis system which can achieve diverse voices, we have been developing a speaker independent approach of HMM-based speech synthesis in which statistical average voice models are adapted to a target speaker using a small amount of speech data. In this paper, we incorporate a high-quality speech vocoding method STRAIGHT and a parameter generation algorithm with global variance into the system for improving quality of synthetic speech. Furthermore, we introduce a feature-space speaker adaptive training algorithm and a gender mixed modeling technique for conducting further normalization of the average voice model. We build an English text-to-speech system using these techniques and show the performance of the system.

This paper addresses the problem of dialogue optimization on large search spaces. For such a purpose, in this paper we propose to learn dialogue strategies using multiple Semi-Markov Decision Processes and hierarchical reinforcement learning. This approach factorizes state variables and actions in order to learn a hierarchy of policies. Our experiments are based on a simulated flight booking dialogue system and compare flat versus hierarchical reinforcement learning. Experimental results show that the proposed approach produced a dramatic search space reduction (99.36%), and converged four orders of magnitude faster than flat reinforcement learning with a very small loss in optimality (on average 0.3 system turns). Results also report that the learnt policies outperformed a hand-crafted one under three different conditions of ASR confidence levels. This approach is appealing to dialogue optimization due to faster learning, reusable subsolutions, and scalability to larger problems.

Joint Dialogue Act segmentation and classification of the new AMI meeting corpus has been performed through an integrated framework based on a switching dynamic Bayesian network and a set of continuous features and language models. The recognition process is based on a dictionary of 15 DA classes tailored for group decision-making. Experimental results show that a novel interpolated Factored Language Model results in a low error rate on the automatic segmentation task, and thus good recognition results can be achieved on AMI multiparty conversational speech.

We propose a joint segmentation and classification approach for the dialogue act recognition task on natural multi-party meetings (ICSI Meeting Corpus). Five broad DA categories are automatically recognised using a generative Dynamic Bayesian Network based infrastructure. Prosodic features and a switching graphical model are used to estimate DA boundaries, in conjunction with a factored language model which is used to relate words and DA categories. This easily generalizable and extensible system promotes a rational approach to the joint DA segmentation and recognition task, and is capable of good recognition performance.

Prosody has been actively studied as an important knowledge source for speech recognition and understanding. In this paper, we are concerned with the question of exploiting prosody for language models to aid automatic speech recognition in the context of meetings. Using an automatic syllable detection algorithm, the syllable-based prosodic features are extracted to form the prosodic representation for each word. Two modeling approaches are then investigated. One is based on a factored language model, which directly uses the prosodic representation and treats it as a `word'. Instead of direct association, the second approach provides a richer probabilistic structure within a hierarchical Bayesian framework by introducing an intermediate latent variable to represent similar prosodic patterns shared by groups of words. Four-fold cross-validation experiments on the ICSI Meeting Corpus show that exploiting prosody for language modeling can significantly reduce the perplexity, and also have marginal reductions in word error rate.

n this paper we give an overview of the AMI project. AMI developed the following: (1) an infrastructure for recording meetings using multiple microphones and cameras; (2) a one hundred hour, manually annotated meeting corpus; (3) a number of techniques for indexing, and summarizing of meeting videos using automatic speech recognition and computer vision, and (4) an extensible framework for browsing, and searching of meeting videos. We give an overview of the various techniques developed in AMI, their integration into our meeting browser framework, and future plans for AMIDA (Augmented Multiparty Interaction with Distant Access), the follow-up project to AMI.

The AMI and AMIDA projects are concerned with the recognition and interpretation of multiparty meetings. Within these projects we have: developed an infrastructure for recording meetings using multiple microphones and cameras; released a 100 hour annotated corpus of meetings; developed techniques for the recognition and interpretation of meetings based primarily on speech recognition and computer vision; and developed an evaluation framework at both component and system levels. In this paper we present an overview of these projects, with an emphasis on speech recognition and content extraction.

The majority of speech summarization research has focused on extracting the most informative dialogue acts from recorde d, archived data. However, a potential use case for speech sum- marization in the meetings domain is to facilitate a meeting in progress by providing the participants - whether they are at tend- ing in-person or remotely - with an indication of the most im- portant parts of the discussion so far. This requires being a ble to determine whether a dialogue act is extract-worthy befor e the global meeting context is available. This paper introduces a novel method for weighting dialogue acts using only very lim- ited local context, and shows that high summary precision is possible even when information about the meeting as a whole is lacking. A new evaluation framework consisting of weighted precision, recall and f-score is detailed, and the novel onl ine summarization method is shown to significantly increase recall and f-score compared with a method using no contextual infor- mation.

This paper explores the issue of term-weighting in the genre of spontaneous, multi-party spoken dialogues, with the intent of using such term-weights in the creation of extractive meeting summaries. The field of text information retrieval has yielded many term-weighting tech- niques to import for our purposes; this paper implements and compares several of these, namely tf.idf, Residual IDF and Gain. We propose that term-weighting for multi-party dialogues can exploit patterns in word us- age among participant speakers, and introduce the su.idf metric as one attempt to do so. Results for all metrics are reported on both manual and automatic speech recognition (ASR) transcripts, and on both the ICSI and AMI meeting corpora.

This paper proposes the use of the Liljencrants-Fant model (LF-model) to represent the glottal source signal in HMM-based speech synthesis systems. These systems generally use a pulse train to model the periodicity of the excitation signal of voiced speech. However, this model produces a strong and uniform harmonic structure throughout the spectrum of the excitation which makes the synthetic speech sound buzzy. The use of a mixed band excitation and phase manipulation reduces this effect but it can result in degradation of the speech quality if the noise component is not weighted carefully. In turn, the LF-waveform has a decaying spectrum at higher frequencies, which is more similar to the real glottal source excitation signal. We conducted a perceptual experiment to test the hypothesis that the LF-model can perform as well as or better than the pulse train in a HMM-based speech synthesizer. In the synthesis, we used the mean values of the LF-parameters, calculated by measurements of the recorded speech. The result of this study is important not only regarding the improvement in speech quality of these type of systems, but also because the LF-model can be used to model many characteristics of the glottal source, such as voice quality, which are important for voice transformation and generation of expressive speech.

Multiparty meetings are a ubiquitous feature of organizations, and there are considerable economic benefits that would arise from their automatic analysis and structuring. In this paper, we are concerned with the segmentation and structuring of meetings (recorded using multiple cameras and microphones) into sequences of group meeting actions such as monologue, discussion and presentation. We outline four families of multimodal features based on speaker turns, lexical transcription, prosody, and visual motion that are extracted from the raw audio and video recordings. We relate these low-level features to more complex group behaviors using a multistream modelling framework based on multistream dynamic Bayesian networks (DBNs). This results in an effective approach to the segmentation problem, resulting in an action error rate of 12.2%, compared with 43% using an approach based on hidden Markov models. Moreover, the multistream DBN developed here leaves scope for many further improvements and extensions.

In this paper we propose partially specified dialogue strategies for dialogue strategy optimization, where part of the strategy is specified deterministically and the rest optimized with Reinforcement Learning (RL). To do this we apply RL with Hierarchical Abstract Machines (HAMs). We also propose to build simulated users using HAMs, incorporating a combination of hierarchical deterministic and probabilistic behaviour. We performed experiments using a single-goal flight booking dialogue system, and compare two dialogue strategies (deterministic and optimized) using three types of simulated user (novice, experienced and expert). Our results show that HAMs are promising for both dialogue optimization and simulation, and provide evidence that indeed partially specified dialogue strategies can outperform deterministic ones (on average 4.7 fewer system turns) with faster learning than the traditional RL framework.

Learning dialogue strategies using the reinforcement learning framework is problematic due to its expensive computational cost. In this paper we propose an algorithm that reduces a state-action space to one which includes only valid state-actions. We performed experiments on full and reduced spaces using three systems (with 5, 9 and 20 slots) in the travel domain using a simulated environment. The task was to learn multi-goal dialogue strategies optimizing single and multiple confirmations. Average results using strategies learnt on reduced spaces reveal the following benefits against full spaces: 1) less computer memory (94% reduction), 2) faster learning (93% faster convergence) and better performance (8.4% less time steps and 7.7% higher reward).

The trajectory HMM has been shown to be useful for model-based speech synthesis where a smoothed trajectory is generated using temporal constraints imposed by dynamic features. To evaluate the performance of such model on an ASR task, we present a trajectory decoder based on tree search with delayed path merging. Experiment on a speaker-dependent phone recognition task using the MOCHA-TIMIT database shows that the MLE-trained trajectory model, while retaining attractive properties of being a proper generative model, tends to favour over-smoothed trajectory among competing hypothesises, and does not perform better than a conventional HMM. We use this to build an argument that models giving better fit on training data may suffer a reduction of discrimination by being too faithful to training data. This partially explains why alternative acoustic models that try to explicitly model temporal constraints do not achieve significant improvements in ASR.

This paper explores the usefulness of prosody in automatically compressing dialogue acts from meeting speech. Specifically, this work attempts to compress utterances by preserving the pitch contour of the original whole utterance. Two methods of doing this are described in detail and are evaluated subjectively using human annotators and objectively using edit distance with a human-authored gold-standard. Both metrics show that such a prosodic approach is much better than the random baseline approach and significantly better than a simple text compression method.

The research presented herein explores the usefulness of incorporating speaker and discourse features in an automatic speech summarization system applied to meeting recordings from the ICSI Meetings corpus. By analyzing speaker activity, turn-taking and discourse cues, it is hypothesized that a system can outperform solely text-based methods inherited from the field of text summarization. The summarization methods are described, two evaluation methods are applied and compared, and the results clearly show that utilizing such features is advantageous and efficient. Even simple methods relying on discourse cues and speaker activity can outperform text summarization approaches.

This paper investigates the usefulness of prosodic features in classifying rhetorical relations between utterances in meeting recordings. Five rhetorical relations of contrast, elaboration, summary, question and cause are explored. Three training methods - supervised, unsupervised, and combined - are compared, and classification is carried out using support vector machines. The results of this pilot study are encouraging but mixed, with pairwise classification achieving an average of 68% accuracy in discerning between relation pairs using only prosodic features, but multi-class classification performing only slightly better than chance.

We address the problem of segmentation and recognition of sequences of multimodal human interactions in meetings. These interactions can be seen as a rough structure of a meeting, and can be used either as input for a meeting browser or as a first step towards a higher semantic analysis of the meeting. A common lexicon of multimodal group meeting actions, a shared meeting data set, and a common evaluation procedure enable us to compare the different approaches. We compare three different multimodal feature sets and our modelling infrastructures: a higher semantic feature approach, multi-layer HMMs, a multistream DBN, as well as a multi-stream mixed-state DBN for disturbed data.

In this paper, we investigate the prob- lem of automatically predicting segment boundaries in spoken multiparty dialogue. We extend prior work in two ways. We first apply approaches that have been pro- posed for predicting top-level topic shifts to the problem of identifying subtopic boundaries. We then explore the impact on performance of using ASR output as opposed to human transcription. Exam- ination of the effect of features shows that predicting top-level and predicting subtopic boundaries are two distinct tasks: (1) for predicting subtopic boundaries, the lexical cohesion-based approach alone can achieve competitive results, (2) for predicting top-level boundaries, the ma- chine learning approach that combines lexical-cohesion and conversational fea- tures performs best, and (3) conversational cues, such as cue phrases and overlapping speech, are better indicators for the top- level prediction task. We also find that the transcription errors inevitable in ASR output have a negative impact on models that combine lexical-cohesion and conver- sational features, but do not change the general preference of approach for the two tasks.

This paper presents a probabilistic method to simulate task-oriented human-computer dialogues at the intention level, that may be used to improve or to evaluate the performance of spoken dialogue systems. Our method uses a network of Hidden Markov Models (HMMs) to predict system and user intentions, where a “language model” predicts sequences of goals and the component HMMs predict sequences of intentions. We compare standard HMMs, Input HMMs and Input-Output HMMs in an effort to better predict sequences of intentions. In addition, we propose a dialogue similarity measure to evaluate the realism of the simulated dialogues. We performed experiments using the DARPA Communicator corpora and report results with three different metrics: dialogue length, dialogue similarity and precision-recall.

Vocal Tract Length Normalisation (VTLN) is a commonly used technique to normalise for inter-speaker variability. It is based on the speaker-specific warping of the frequency axis, parameterised by a scalar warp factor. This factor is typically estimated using maximum likelihood. We discuss how VTLN may be applied to multiparty conversations, reporting a substantial decrease in word error rate in experiments using the ICSI meetings corpus. We investigate the behaviour of the VTLN warping factor and show that a stable estimate is not obtained. Instead it appears to be influenced by the context of the meeting, in particular the current conversational partner. These results are consistent with predictions made by the psycholinguistic interactive alignment account of dialogue, when applied at the acoustic and phonological levels.

Several approaches to automatic speech summarization are discussed below, using the ICSI Meetings corpus. We contrast feature-based approaches using prosodic and lexical features with maximal marginal relevance and latent semantic analysis approaches to summarization. While the latter two techniques are borrowed directly from the field of text summarization, feature-based approaches using prosodic information are able to utilize characteristics unique to speech data. We also investigate how the summarization results might deteriorate when carried out on ASR output as opposed to manual transcripts. All of the summaries are of an extractive variety, and are compared using the software ROUGE.

The research below explores schemes for evaluating automatic summaries of business meetings, using the ICSI Meeting Corpus. Both automatic and subjective evaluations were carried out, with a central interest being whether or not the two types of evaluations correlate with each other. The evaluation metrics were used to compare and contrast differing approaches to automatic summarization, the deterioration of summary quality on ASR output versus manual transcripts, and to determine whether manual extracts are rated significantly higher than automatic extracts.

This paper presents an automatic system for structuring and preparing a news broadcast for applications such as speech summarization, browsing, archiving and information retrieval. This process comprises transcribing the audio using an automatic speech recognizer and subsequently segmenting the text into utterances and topics. A maximum entropy approach is used to build statistical models for both utterance and topic segmentation. The experimental work addresses the effect on performance of the topic boundary detector of three factors: the information sources used, the quality of the ASR transcripts, and the quality of the utterance boundary detector. The results show that the topic segmentation is not affected severely by transcripts errors, whereas errors in the utterance segmentation are more devastating.

The automatic processing of speech collected in conference style meetings has attracted considerable interest with several large scale projects devoted to this area. In this paper we explore the use of various meeting corpora for the purpose of automatic speech recognition. In particular we investigate the similarity of these resources and how to efficiently use them in the construction of a meeting transcription system. The analysis shows distinctive features for each resource. However the benefit in pooling data and hence the similarity seems sufficient to speak of a generic conference meeting domain . In this context this paper also presents work on development for the AMI meeting transcription system, a joint effort by seven sites working on the AMI (augmented multi-party interaction) project.

In this paper we describe the 2005 AMI system for the transcription of speech in meetings used in the 2005 NIST RT evaluations. The system was designed for participation in the speech to text part of the evaluations, in particular for transcription of speech recorded with multiple distant microphones and independent headset microphones. System performance was tested on both conference room and lecture style meetings. Although input sources are processed using different frontends, the recognition process is based on a unified system architecture. The system operates in multiple passes and makes use of state of the art technologies such as discriminative training, vocal tract length normalisation, heteroscedastic linear discriminant analysis, speaker adaptation with maximum likelihood linear regression and minimum word error rate decoding. In this paper we describe the system performance on the official development and test sets for the NIST RT05s evaluations. The system was jointly developed in less than 10 months by a multi-site team and was shown to achieve competitive performance.

The analysis of scenarios in which a number of microphones record the activity of speakers, such as in a roundtable meeting, presents a number of computational challenges. For example, if each participant wears a microphone, it can receive speech from both the microphone's wearer (local speech) and from other participants (crosstalk). The recorded audio can be broadly classified in four ways: local speech, crosstalk plus local speech, crosstalk alone and silence. We describe two experiments related to the automatic classification of audio into these four classes. The first experiment attempted to optimise a set of acoustic features for use with a Gaussian mixture model (GMM) classifier. A large set of potential acoustic features were considered, some of which have been employed in previous studies. The best-performing features were found to be kurtosis, fundamentalness and cross-correlation metrics. The second experiment used these features to train an ergodic hidden Markov model classifier. Tests performed on a large corpus of recorded meetings show classification accuracies of up to 96%, and automatic speech recognition performance close to that obtained using ground truth segmentation.

Spoken word audio collections cover many domains, including radio and television broadcasts, oral narratives, governmental proceedings, lectures, and telephone conversations. The collection, access and preservation of such data is stimulated by political, economic, cultural and educational needs. This paper outlines the major issues in the field, reviews the current state of technology, examines the rapidly changing policy issues relating to privacy and copyright, and presents issues relating to the collection and preservation of spoken audio content.

The aim of this work is to develop a practical framework, which extends the classical Hidden Markov Models (HMM) for continuous speech recognition based on the Maximum Entropy (MaxEnt) principle. The MaxEnt models can estimate the posterior probabilities directly as with Hybrid NN/HMM connectionist speech recognition systems. In particular, a new acoustic modelling based on discriminative MaxEnt models is formulated and is being developed to replace the generative Gaussian Mixture Models (GMM) commonly used to model acoustic variability. Initial experimental results using the TIMIT phone task are reported.

This paper investigates the automatic analysis and segmentation of meetings. A meeting is analysed in terms of individual behaviours and group interactions, in order to decompose each meeting in a sequence of relevant phases, named meeting actions. Three feature families are extracted from multimodal recordings: prosody from individual lapel microphone signals, speaker activity from microphone array data and lexical features from textual transcripts. A statistical approach is then used to relate low-level features with a set of abstract categories. In order to provide a flexible and powerful framework, we have employed a dynamic Bayesian network based model, characterized by multiple stream processing and flexible state duration modelling. Experimental results demonstrate the strength of this system, providing a meeting action error rate of 9%.

This paper presents a text-independent speaker verification system using support vector machines (SVMs) with score-space kernels. Score-space kernels, generalize Fisher kernels, and are based on an underlying generative model, such as a Gaussian mixture model (GMM). This approach provides direct discrimination between whole sequences, in contrast to the frame-level approaches at the heart of most current systems. The resultant SVMs have a very high dimensionality, since it is related to the number of parameters in the underlying generative model. To ameliorate problems that can arise in the resultant optimization, we introduce a technique called spherical normalization that preconditions the Hessian matrix. We have performed speaker verification experiments using the PolyVar database. The SVM system presented here reduces the relative error rates by 34% compared to a GMM likelihood ratio system.

This paper presents trainable methods for extracting principal content words from voicemail messages. The short text summaries generated are suitable for mobile messaging applications. The system uses a set of classifiers to identify the summary words, with each word being identified by a vector of lexical and prosodic features. We use an ROC-based algorithm, Parcel, to select input features (and classifiers). We have performed a series of objective and subjective evaluations using unseen data from two different speech recognition systems, as well as human transcriptions of voicemail speech.

"How analysis, retrieval and delivery phases make spoken audio content more accessible"

The automatic processing of speech collected in conference style meetings has attracted considerable interest with several large scale projects devoted to this area. This paper describes the development of a baseline automatic speech transcription system for meetings in the context of the AMI (Augmented Multiparty Interaction) project. We present several techniques important to processing of this data and show the performance in terms of word error rates (WERs). An important aspect of transcription of this data is the necessary flexibility in terms of audio pre-processing. Real world systems have to deal with flexible input, for example by using microphone arrays or randomly placed microphones in a room. Automatic segmentation and microphone array processing techniques are described and the effect on WERs is discussed. The system and its components presented in this paper yield compettive performance and form a baseline for future research in this domain.

In this paper we report on a series of experiments investigating the path from text-summarisation to style-specific summarisation of spoken news stories. We show that the portability of traditional text summarisation features to broadcast news is dependent on the diffusiveness of the information in the broadcast news story. An analysis of two categories of news stories (containing only read speech or some spontaneous speech) demonstrates the importance of the style and the quality of the transcript, when extracting the summary-worthy information content. Further experiments indicate the advantages of doing style-specific summarisation of broadcast news.

This paper is about the automatic structuring of multiparty meetings using audio information. We have used a corpus of 53 meetings, recorded using a microphone array and lapel microphones for each participant. The task was to segment meetings into a sequence of meeting actions, or phases. We have adopted a statistical approach using dynamic Bayesian networks (DBNs). Two DBN architectures were investigated: a two-level hidden Markov model (HMM) in which the acoustic observations were concatenated; and a multistream DBN in which two separate observation sequences were modelled. Additionally we have also explored the use of counter variables to constrain the number of action transitions. Experimental results indicate that the DBN architectures are an improvement over a simple baseline HMM, with the multistream DBN with counter constraints producing an action error rate of 6%.

This paper investigates the automatic segmentation of meetings into a sequence of group actions or phases. Our work is based on a corpus of multiparty meetings collected in a meeting room instrumented with video cameras, lapel microphones and a microphone array. We have extracted a set of feature streams, in this case extracted from the audio data, based on speaker turns, prosody and a transcript of what was spoken. We have related these signals to the higher level semantic categories via a multistream statistical model based on dynamic Bayesian networks (DBNs). We report on a set of experiments in which different DBN architectures are compared, together with the different feature streams. The resultant system has an action error rate of 9%.

In this paper we propose a discriminative approach to acoustic space dimensionality selection based on maximum entropy modelling. We form a set of constraints by composing the acoustic space with the space of phone classes, and use a continuous feature formulation of maximum entropy modelling to select an optimal feature set. The suggested approach has two steps: (1) the selection of the best acoustic space that efficiently and economically represents the acoustic data and its variability; (2) the combination of selected acoustic features in the maximum entropy framework to estimate the posterior probabilities over the phonetic labels given the acoustic input. Specific contributions of this paper include a parameter estimation algorithm (generalized improved iterative scaling) that enables the use of negative features, the parameterization of constraint functions using Gaussian mixture models, and experimental results using the TIMIT database.

This is a preprint of a tutorial on statistical language modelling, based on Yoshi Gotoh's course at the ELSNET-2000 Summer School on Text and Speech Triggered Information Access.

This paper is about the evaluation of a system that generates short text summaries of voicemail messages, suitable for transmission as text messages. Our approach to summarization is based on a speech-recognized transcript of the voicemail message, from which a set of summary words is extracted. The system uses a classifier to identify the summary words, with each word being identified by a vector of lexical and prosodic features. The features are selected using Parcel, an ROC-based algorithm. Our evaluations of the system, using a slot error rate metric, have compared manual and automatic summarization, and manual and automatic recognition (using two different recognizers). We also report on two subjective evaluations using mean opinion score of summaries, and a set of comprehension tests. The main results from these experiments were that the perceived difference in quality of summarization was affected more by errors resulting from automatic transcription, than by the automatic summarization process.

We investigate approaches to accessing information from the streams of audio data that result from multi-channel recordings of meetings. The methods investigated use word-level transcriptions, and information derived from models of speaker activity and speaker turn patterns. Our experiments include spoken document retrieval for meetings, automatic structuring of meetings based on self-similarity matrices of speaker turn patterns and a simple model of speaker activity. Meeting recordings are rich in both lexical and non-lexical information; our results illustrate some novel kinds of analysis made possible by a transcribed corpus of natural meetings.

In this paper we seek to explore the interaction between the style of a broadcast news story and its summarization technique. We report the performance of three different summarization techniques on broadcast news stories, which are split into planned speech and spontaneous speech. The initial results indicate that some summarization techniques work better for the documents with spontaneous speech than for those with planned speech. Even for human beings some documents are inherently dif cult to summarize. We observe this correlation between degree of dif culty in summarizing and performance of the three automatic summarizers. Given the high frequency of named entities in broadcast news and even greater number of references to these named entities, we also gauge the effect of named entity and coreference resolution in a news story, on the performance of these summarizers.

This paper is about a system that extracts principal content words from speech-recognized transcripts of voicemail messages and classifies them into proper names, telephone numbers, dates/times and `other'. The short text summaries generated are suitable for mobile messaging applications. The system uses a set of classifiers to identify the summary words, with each word being identified by a vector of lexical and prosodic features. The features are selected using Parcel, an ROC-based algorithm. We visually compare the role of a large number of individual features and discuss effective ways to combine them. We finally evaluate their performance on manual and automatic transcriptions derived from two different speech recognition systems.

Support vector machines with the Fisher and score-space kernels are used for text independent speaker verification to provide direct q discrimination between complete utterances. This is unlike approaches such as discriminatively trained Gaussian mixture models or other discriminative classifiers that discriminate at the frame-level only. Using the sequence-level discrimination approach we are able to achieve error-rates that are significantly better than the current state-of-the-art on the PolyVar database.

In this paper we report on a series of experiments which compare the effect of individual features on both text and speech summarisation, the effect of basing the speech summaries on automatic speech recognition transcripts with varying word error rates, and the effect of summarisation approach and transcript source on summary quality. We show that classical text summarisation features (based on stylistic and content information) are portable to broadcast news. However, the quality of the speech transcripts as well as the difference in information structure between broadcast and newspaper news affect the usability of the individual features.

An extension to the conventional speech / nonspeech classification framework is presented for a scenario in which a number of microphones record the activity of speakers present at a meeting (one microphone per speaker). Since each microphone can receive speech from both the participant wearing the microphone (local speech) and other participants (crosstalk), the recorded audio can be broadly classified in four ways: local speech, crosstalk plus local speech, crosstalk alone and silence. We describe a classifier in which a Gaussian mixture model (GMM) is used to model each class. A large set of potential acoustic features are considered, some of which have been employed in previous speech / nonspeech classifiers. A combination of two feature selection algorithms is used to identify the optimal feature set for each class. Results from the GMM classifier using the selected features are superior to those of a previously published approach.

This paper describes connectionist techniques for recognition of Broadcast News. The fundamental difference between connectionist systems and more conventional mixture-of-Gaussian systems is that connectionist models directly estimate posterior probabilities as opposed to likelihoods. Access to posterior probabilities has enabled us to develop a number of novel approaches to confidence estimation, pronunciation modelling and search. In addition we have investigated a new feature extraction technique based on the modulation-filtered spectrogram, and methods for combining multiple information sources. We have incorporated all of these techniques into a system for the transcription of Broadcast News, and we present results on the 1998 DARPA Hub-4E Broadcast News evaluation data.

Though the field of spoken dialogue systems has developed quickly in the last decade, rapid design of dialogue strategies remains uneasy. Several approaches to the problem of automatic strategy learning have been proposed and the use of Reinforcement Learning introduced by Levin and Pieraccini is becoming part of the state of the art in this area. However, the quality of the strategy learned by the system depends on the definition of the optimization criterion and on the accuracy of the environment model. In this paper, we propose to bring a model of an ASR system in the simulated environment in order to enhance the learned strategy. To do so, we introduced recognition error rates and confidence levels produced by ASR systems in the optimization criterion.

Support vector machines are evaluated on speaker verification and speaker identification tasks. We compare the polynomial kernel, the Fisher kernel, a likelihood ratio kernel and the pair hidden Markov model kernel with baseline systems based on a discriminative polynomial classifier and generative Gaussian mixture model classifiers. Simulations were carried out on the YOHO database and some promising results were obtained.

This paper presents a novel data-driven approach to summarizing spoken audio transcripts utilizing lexical and prosodic features. The former are obtained from a speech recognizer and the latter are extracted automatically from speech waveforms. We employ a feature subset selection algorithm, based on ROC curves, which examines different combinations of features at different target operating conditions. The approach is evaluated on the IBM Voicemail corpus, demonstrating that it is possible and desirable to avoid complete commitment to a single best classifier or feature set.

This paper describes an alternative architecture for voicemail data retrieval on the move. It is comprised of three distinct components: a speech recognizer, a text summarizer and a WAP push service initiator, enabling mobile users to receive a text summary of their voicemail in realtime without an explicit request. Our approach overcomes the cost and usability limitations of the conventional voicemail retrieval paradigm which requires a connection establishment in order to listen to spoken messages. We report performance results on all different components of the system which has been trained on a database containing 1843 North American English messages as well as on the duration of the corresponding data path. The proposed architecture can be further customized to meet the requirements of a complete voicemail value-added service.

This paper describes our participation in the TREC-9 Spoken Document Retrieval (SDR) track. The THISL SDR system consists of a realtime version of a hybrid connectionist/HMM large vocabulary speech recognition system and a probabilistic text retrieval system. This paper describes the configuration of the speech recognition and text retrieval systems, including segmentation and query expansion. We report our results for development tests using the TREC-8 queries, and for the TREC-9 evaluation.

This paper is about the development of statistical models of prosodic features to generate linguistic meta-data for spoken language. In particular, we are concerned with automatically punctuating the output of a broadcast news speech recogniser. We present a statistical finite state model that combines prosodic, linguistic and punctuation class features. Experimental results are presented using the Hub-4 Broadcast News corpus, and in the light of our results we discuss the issue of a suitable method of evaluating the present task.

When a speaker leaves a voicemail message there are prosodic cues that emphasize the important points in the message, in addition to lexical content. In this paper we compare and visualize the relative contribution of these two types of features within a voicemail summarization system. We describe the system's ability to generate summaries of two test sets, having trained and validated using 700 messages from the IBM Voicemail corpus. Results measuring the quality of summary artifacts show that combined lexical and prosodic features are at least as robust as combined lexical features alone across all operating conditions.

This paper discusses the development of trainable statistical models for extracting content from television and radio news broadcasts. In particular we concentrate on statistical finite state models for identifying proper names and other named entities in broadcast speech. Two models are presented: the first models name class information as a word attribute; the second explicitly models both word-word and class-class transitions. A common n-gram based formulation is used for both models. The task of named entity identification is characterized by relatively sparse training data and issues related to smoothing are discussed. Experiments are reported using the DARPA/NIST Hub-4E evaluation for North American Broadcast News.

This paper describes a spoken document retrieval (SDR) system for British and North American Broadcast News. The system is based on a connectionist large vocabulary speech recognizer and a probabilistic information retrieval system. We discuss the development of a realtime Broadcast News speech recognizer, and its integration into an SDR system. Two advances were made for this task: automatic segmentation and statistical query expansion using a secondary corpus. Precision and recall results using the Text Retrieval Conference (TREC) SDR evaluation infrastructure are reported throughout the paper, and we discuss the application of these developments to a large scale SDR task based on an archive of British English broadcast news.

The class of finite mixtures of multivariate Bernoulli distributions is known to be nonidentifiable, i.e., different values of the mixture parameters can correspond to exactly the same probability distribution. In principle, this would mean that sample estimates using this model would give rise to different interpretations. We give empirical support to the fact that estimation of this class of mixtures can still produce meaningful results in practice, thus lessening the importance of the identifiability problem. We also show that the EM algorithm is guaranteed to converge to a proper maximum likelihood estimate, owing to a property of the log-likelihood surface. Experiments with synthetic data sets show that an original generating distribution can be estimated from a sample. Experiments with an electropalatography (EPG) data set show important structure in the data.

This paper describes the development of a system to transcribe and summarize voicemail messages. The results of the research presented in this paper are two-fold. First, a hybrid connectionist approach to the Voicemail transcription task shows that competitive performance can be achieved using a context-independent system with fewer parameters than those based on mixtures of Gaussian likelihoods. Second, an effective and robust combination of statistical with prior knowledge sources for term weighting is used to extract information from the decoders output in order to deliver summaries to the message recipients via a GSM Short Message Service (SMS) gateway.

The rate of occurrence of words is not uniform but varies from document to document. Despite this observation, parameters for conventional n-gram language models are usually derived using the assumption of a constant word rate. In this paper we investigate the use of variable word rate assumption, modelled by a Poisson distribution or a continuous mixture of Poissons. We present an approach to estimating the relative frequencies of words or n-grams taking prior information of their occurrences into account. Discounting and smoothing schemes are also considered. Using the Broadcast News task, the approach demonstrates a reduction of perplexity up to 10%.

This paper presents an approach to identifying sentence boundaries in broadcast speech transcripts. We describe finite state models that extract sentence boundary information statistically from text and audio sources. An n-gram language model is constructed from a collection of British English news broadcasts and scripts. An alternative model is estimated from pause duration information in speech recogniser outputs aligned with their programme script counterparts. Experimental results show that the pause duration model alone outperforms the language modelling approach and that, by combining these two models, it can be improved further and precision and recall scores of over 70% were attained for the task.

This paper describes the participation of the THISL group at the TREC-8 Spoken Document Retrieval (SDR) track. The THISL SDR system consists of the realtime version of the Abbot large vocabulary speech recognition system and the thislIR text retrieval system. The TREC-8 evaluation assessed SDR performance on a corpus of 500 hours of broadcast news material collected over a five month period. The main test condition involved retrieval of stories defined by manual segmentation of the corpus in which non-news material, such as commercials, were excluded. An optional test condition required required retrieval of the same stories from the unsegmented audio stream. The THISL SDR system participated at both test conditions. The results show that a system such as THISL can produce respectable information retrieval performance on a realistically-sized corpus of unsegmented audio material.

This paper describes the SPRACH system developed for the 1998 Hub-4E broadcast news evaluation. The system is based on the connectionist-HMM framework and uses both recurrent neural network and multi-layer perceptron acoustic models. We describe both a system designed for the primary transcription hub, and a system for the less-than 10 times real-time spoke. We then describe recent developments to CHRONOS, a time-first stack decoder. We show how these developments have simplified the evaluation system, and led to significant reductions in the error rate of the 10x real-time system. We also present a system designed to operate in real-time with negligible search error.

This paper described the THISL spoken document retrieval system for British and North American Broadcast News. The system is based on the Abbot large vocabulary speech recognizer and a probabilistic text retrieval system. We discuss the development of a realtime British English Broadcast News system, and its integration into a spoken document retrieval system. Detailed evaluation is performed using a similar North American Broadcast News system, to take advantage of the TREC SDR evaluation methodology. We report results on this evaluation, with particular reference to the effect of query expansion and of automatic segmentation algorithms.

In this paper we describe stochastic finite state model for named entity (NE) identification, based on explicit word-level n-gram relations. NE categories are incorporated in the model as word attributes. We present an overview of the approach, describing how the extensible vocabulary model may be used for NE identification. We report development and evaluation results on a North American Broadcast News task. This approach resulted in average precision and recall scores of around 83% on hand transcribed data, and 73% on the SPRACH recogniser output. We also present an error analysis and a comparison of our approach with an alternative statistical approach.

We present a latent variable approach to the acoustic-to-articulatory mapping problem, where different vocal tract configurations can give rise to the same acoustics. In latent variable modelling, the combined acoustic and articulatory data are assumed to have been generated by an underlying low-dimensional process. A parametric probabilistic model is estimated and mappings are derived from the respective conditional distributions. This has the advantage over other methods, such as articulatory codebooks or neural networks, of directly addressing the nonuniqueness problem. We demonstrate our approach with electropalatographic and acoustic data from the ACCOR database.

In this paper, we present a novel, efficient search strategy for large vocabulary continuous speech recognition. The search algorithm, based on a stack decoder framework, utilizes phone-level posterior probability estimates (produced by a connectionist/HMM acoustic model) as a basis for phone deactivation pruning - a highly efficient method of reducing the required computation. The single-pass algorithm is naturally factored into the time-asynchronous processing of the word sequence and the time-synchronous processing of the HMM state sequence. This enables the search to be decoupled from the language model while still maintaining the computational benefits of time-synchronous processing. The incorporation of the language model in the search is discussed and computationally cheap approximations to the full language model are introduced. Experiments were performed on the North American Business News task using a 60,000 word vocabulary and a trigram language model. Results indicate that the computational cost of the search may be reduced by more than a factor of 40 with a relative search error of less than 2% using the techniques discussed in the paper.

We introduce Named Entity (NE) Language Modelling, a stochastic finite state machine approach to identifying both words and NE categories from a stream of spoken data. We provide an overview of our approach to NE tagged language model (LM) generation together with results of the application of such a LM to the task of out-of-vocabulary (OOV) word reduction in large vocabulary speech recognition. Using the Wall Street Journal and Broadcast News corpora, it is shown that the tagged LM was able to reduce the overall word error rate by 14%, detecting up to 70% of previously OOV words. We also describe an example of the direct tagging of spoken data with NE categories.

This paper presents an approach to integrating functions for both transcription and named entity (NE) identification into a large vocabulary continuous speech recognition system. It builds on NE tagged language modelling approach, which was recently applied for development of the statistical NE annotation system. We also present results for proper name identification experiment using the Hub-4E open evaluation data.

In this paper we introduce a set of related confidence measures for large vocabulary continuous speech recognition (LVCSR) based on local phone posterior probability estimates output by an acceptor HMM acoustic model. In addition to their computational efficiency, these confidence measures are attractive as they may be applied at the state-, phone-, word- or utterance-levels, potentially enabling discrimination between different causes of low confidence recognizer output, such as unclear acoustics or mismatched pronunciation models. We have evaluated these confidence measures for utterance verification using a number of different metrics. Experiments reveal several trends in `profitability of rejection', as measured by the unconditional error rate of a hypothesis test. These trends suggest that crude pronunciation models can mask the relatively subtle reductions in confidence caused by out-of-vocabulary (OOV) words and disfluencies, but not the gross model mismatches elicited by non-speech sounds. The observation that a purely acoustic confidence measure can provide improved performance over a measure based upon both acoustic and language model information for data drawn from the Broadcast News corpus, but not for data drawn from the North American Business News corpus suggests that the quality of model fit offered by a trigram language model is reduced for Broadcast News data. We also argue that acoustic confidence measures may be used to inform the search for improved pronunciation models.

This paper describes an approach for constructing a mixture of language models based on simple statistical notions of semantics using probabilistic models developed for information retrieval. The approach encapsulates corpus-derived semantic information and is able to model varying styles of text. Using such information, the corpus texts are clustered in an unsupervised manner and a mixture of topic-specific language models is automatically created. The principal contribution of this work is to characterise the document space resulting from information retrieval techniques and to demonstrate the approach for mixture language modelling. A comparison is made between manual and automatic clustering in order to elucidate how the global content information is expressed in the space. We also compare (in terms of association with manual clustering and language modelling accuracy) alternative term-weighting schemes and the effect of singular valued decomposition dimension reduction (latent semantic analysis). Test set perplexity results using the British National Corpus indicate that the approach can improve the potential of statistical language modelling. Using an adaptive procedure, the conventional model may be tuned to track text data with a slight increase in computational cost.

This paper describes the THISL news retrieval system which maintains an archive of BBC radio and television news recordings. The system uses the Abbot large vocabulary continuous speech recognition system to transcribe news broadcasts, and the thislIR text retrieval system to index and access the transcripts. Decoding and indexing is performed automatically, and the archive is updated with three hours of new material every day. A web-based interface to the retrieval system has been devised to facilitate access to the archive.

This paper described the THISL spoken document retrieval system for British and North American Broadcast News. The system is based on the Abbot large vocabulary speech recognizer, using a recurrent network acoustic model, and a probabilistic text retrieval system. We discuss the development of a realtime British English Broadcast News system, and its integration into a spoken document retrieval system. Detailed evaluation is performed using a similar North American Broadcast News system, to take advantage of the TREC SDR evaluation methodology. We report results on this evaluation, with particular reference to the effect of query expansion and of automatic segmentation algorithms.

We have developed two conceptually different systems that are able to identify named entities from spoken audio. One (referred to as SPRACH-S) has a stochastic finite state machine structure for use with an acoustic model that identifies both words and named entities from speech data. The other (referred to as SPRACH-R) is a rule-based system which uses matching against stored name lists, part-of-speech tagging, and light phrasal parsing with specialised named entity grammars. We provide an overview of the two approaches and present results on the Hub-4E IE-NE evaluation task.

This paper describes the THISL system that participated in the TREC-7 evaluation, Spoken Document Retrieval (SDR) Track, and presents the results obtained, together with some analysis. The THISL system is based on the Abbot speech recognition system and the thislIR text retrieval system. In this evaluation we were concerned with investigating the suitability for SDR of a recognizer running at less than ten times realtime, the use of multiple transcriptions and word graphs, the effect of simple query expansion algorithms and the effect of varying standard IR parameters.

This paper describes a spoken document retrieval system, combining the Abbot large vocabulary continuous speech recognition (LVCSR) system developed by Cambridge University, Sheffield University and SoftSound, and the PRISE information retrieval engine developed by NIST. The system was constructed to enable us to participate in the TREC 6 Spoken Document Retrieval experimental evaluation. Our key aims in this work wer e to produce a complete system for the SDR task, to investigate the effect of a word error rate of 30-50% on retrieval performance and to investigate the integration of LVCSR and word spotting in a retrieval task.

THISL is an ESPRIT Long Term Research Project focused the development and construction of a system to items from an archive of television and radio news broadcasts. In this paper we outline our spoken document retrieval system based on the Abbot speech recognizer and a text retrieval system based on Okapi term-weighting . The system has been evaluated as part of the TREC-6 and TREC-7 spoken document retrieval evaluations and we report on the results of the TREC-7 evaluation based on a document collection of 100 hours of North American broadcast news.

We use electropalatographic (EPG) data as a test bed for dimensionality reduction methods based in latent variable modelling, in which an underlying lower dimension representation is inferred directly from the data. Several models (and mixtures of them) are investigated, including factor analysis and the generative topographic mapping (GTM). Experiments indicate that nonlinear latent variable modelling reveals a low-dimensional structure in the data inaccessible to the investigated linear models.

In this paper we define an acoustic confidence measure based on the estimates of local posterior probabilities produced by a HMM/ANN large vocabulary continuous speech recognition system. We use this measure to segment continuous audio into regions where it is and is not appropriate to expend recognition effort. The segmentation is computationally inexpensive and provides reductions in both overall word error rate and decoding time. The technique is evaluated using material from the Broadcast News corpus.

The THISL spoken document retrieval system is based on the Abbot Large Vocabulary Continuous Speech Recognition (LVCSR) system developed by Cambridge University, Sheffield University and SoftSound, and uses PRISE (NIST) for indexing and retrieval. We participated in full SDR mode. Our approach was to transcribe the spoken documents at the word level using Abbot, indexing the resulting text transcriptions using PRISE. The LVCSR system uses a recurrent network-based acoustic model (with no adaptation to different conditions) trained on the 50 hour Broadcast News training set, a 65,000 word vocabulary and a trigram language model derived from Broadcast News text. Words in queries which were out-of-vocabulary (OOV) were word spotted at query time (utilizing the posterior phone probabilities output by the acoustic model), added to the transcriptions of the relevant documents and the collection was then re-indexed. We generated pronunciations at run-time for OOV words using the Festival TTS system (University of Edinburgh).

In this paper we define a number of confidence measures derived from an acceptor HMM and evaluate their performance for the task of utterance verification using the North American Business News (NAB) and Broadcast News (BN) corpora. Results are presented for decodings made at both the word and phone level which show the relative profitability of rejection provided by the diverse set of confidence measures. The results indicate that language model dependent confidence measures have reduced performance on BN data relative to that for the more grammatically constrained NAB data. An explanation linking the observations that rejection is more profitable for noisy acoustics, for a reduced vocabulary and at the phone level is also given.

We consider the problem of obtaining a reduced dimension representation of electropalatographic (EPG) data. An unsupervised learning approach based on latent variable modelling is adopted, in which an underlying lower dimension representation is inferred directly from the data. Several latent variable models are investigated, including factor analysis and the generative topographic mapping (GTM). Experiments were carried out using a subset of the EUR-ACCOR database, and the results indicate that these automatic methods capture important, adaptive structure in the EPG data. Nonlinear latent variable modelling clearly outperforms the investigated linear models in terms of log-likelihood and reconstruction error and suggests a substantially smaller intrinsic dimensionality for the EPG data than that claimed by previous studies. A two-dimensional representation is produced with applications to speech therapy, language learning and articulatory dynamics.

In this paper, we investigate the use of confidence measures for the evaluation of pronunciation models and the employment of these evaluations in an automatic baseform learning process. The confidence measures and pronunciation models are obtained from the Abbot hybrid Hidden Markov Model/Artificial Neural Network Large Vocabulary Continuous Speech Recognition system. Experiments were carried out for a number of baseform learning schemes using the ARPA North American Business News and the Broadcast News corpora from which it was found that a confidence measure based scheme provided the largest reduction in Word Error Rate.

The results of our research presented in this paper are two-fold. First, an estimation of global posteriors[5 5 is formalized in the framework of hybrid HMM/ANN systems. It is shown that hybrid HMM/ANN systems, in which the ANN part estimates local posteriors can be used to model global posteriors. This formalization provides us with a clear theory in which both REMAP and “classical” Viterbi trained hybrid systems are unified. Second, a new forward-backward training of hybrid HMM/ANN systems is derived from the previous formulation. Comparisons of performance between Viterbi and forward-backward hybrid systems are presented and discussed.

In this paper we introduce four acoustic confidence measures which are derived from the output of a hybrid HMM/ANN large vocabulary continuous speech recognition system. These confidence measures, based on local posterior probability estimates computed by an ANN, are evaluated at both phone and word levels, using the North American Business News corpus.

In this paper, an approach for constructing mixture language models (LMs) based on some notion of semantics is discussed. To this end, a technique known as latent semantic analysis (LSA) is used. The approach encapsulates corpus-derived semantic information and is able to model the varying style of the text. Using such information, the corpus texts are clustered in an unsupervised manner and mixture LMs are automatically created. This work builds on previous work in the field of information retrieval which was recently applied by Bellegarda et. al. to the problem of clustering words by semantic categories. The principal contribution of this work is to characterize the document space resulting from the LSA modeling and to demonstrate the approach for mixture LM application. Comparison is made between manual and automatic clustering in order to elucidate how the semantic information is expressed in the space. It is shown that, using semantic information, mixture LMs performs better than a conventional single LM with slight increase of computational cost.

In this letter we introduce a new pruning strategy for large vocabulary continuous speech recognition based on direct estimates of local posterior phone probabilities. This approach is well suited to hybrid connectionist/hidden Markov model systems. Experiments on the Wall Street Journal task using a 20,000 word vocabulary and a trigram language model have demonstrated that phone deactivation pruning can increase the speed of recognition-time search by up to a factor of 10, with a relative increase in error rate of less than 2%.

This chapter describes a use of recurrent neural networks (ie, feedback is incorporated in the computation) as an acoustic model for continuous speech recognition. The form of the recurrent neural network is described, along with an appropriate parameter estimation procedure. For each frame of acoustic data, the recurrent network generates an estimate of the posterior probability of the possible phones given the observed acoustic signal. The posteriors are then converted into scaled likelihoods and used as the observation probabilities within a conventional decoding paradigm (eg, Viterbi decoding). The advantages of the using recurrent networks are that they require a small number of parameters and provide a fast decoding capability (relative to conventional large vocabulary HMM systems).

This work further develops and analyses the large vocabulary continuous speech recognition search strategy reported at ICASSP-95. In particular, the posterior-based phone deactivation pruning approach has been extended to include phone-dependent thresholds and an improved estimate of the least upper bound on the utterance log-probability has been developed. Analysis of the pruning procedures and of the search's interaction with the language model has also been performed. Experiments were carried out using the ARPA North American Business News task with a 20,000 word vocabulary and a trigram language model. As a result of these improvements and analyses, the computational cost of the recognition process performed by the Noway decoder has been substantially reduced.

In this paper we present a novel, efficient search strategy for large vocabulary continuous speech recognition (LVCSR). The search algorithm, based on stack decoding, uses posterior phone probability estimates to substantially increase its efficiency with minimal effect on accuracy. In particular, the search space is dramatically reduced by phone deactivation pruning where phones with a small local posterior probability are deactivated. This approach is particularly well-suited to hybrid connectionist/hidden Markov model systems because posterior phone probabilities are directly computed by the acoustic model. On large vocabulary tasks, using a trigram language model, this increased the search speed by an order of magnitude, with 2% or less relative search error. Results from a hybrid system are presented using the Wall Street Journal LVCSR database for a 20,000 word task using a backed-off trigram language model. For this task, our single-pass decoder took around 15 times realtime on an HP735 workstation. At the cost of 7% relative search error, decoding time can be speeded up to approximately realtime.

It is well known that recognition performance degrades significantly when moving from a speaker- dependent to a speaker-independent system. Traditional hidden Markov model (HMM) systems have successfully applied speaker-adaptation approaches to reduce this degradation. In this paper we present and evaluate some techniques for speaker-adaptation of a hybrid HMM-artificial neural network (ANN) continuous speech recognition system. These techniques are applied to a well trained, speaker-independent, hybrid HMM-ANN system and the recognizer parameters are adapted to a new speaker through off-line procedures. The techniques are evaluated on the DARPA RM corpus using varying amounts of adaptation material and different ANN architectures. The results show that speaker-adaptation within the hybrid framework can substantially improve system performance.

ABBOT is the hybrid connectionist-hidden Markov model (HMM) large-vocabulary continuous speech recognition (CSR) system developed at Cambridge University. This system uses a recurrent network to estimate the acoustic observation probabilities within an HMM framework. A major advantage of this approach is that good performance is achieved using context-independent acoustic models and requiring many fewer parameters than comparable HMM systems. This paper presents substantial performance improvements gained from new approaches to connectionist model combination and phone-duration modeling. Additional capability has also been achieved by extending the decoder to handle larger vocabulary tasks (20,000 words and greater) with a trigram language model. This paper describes the recent modifications to the system and experimental results are reported for various test and development sets from the November 1992, 1993, and 1994 ARPA evaluations of spoken language systems.

We are concerned with integrating connectionist networks into a hidden Markov model (HMM) speech recognition system. This is achieved through a statistical interpretation of connectionist networks as probability estimators. We review the basis of HMM speech recognition and point out the possible benefits of incorporating connectionist networks. Issues necessary to the construction of a connectionist HMM recognition system are discussed, including choice of connectionist probability estimator. We describe the performance of such a system, using a multi-layer perceptron probability estimator, evaluated on the speaker-independent DARPA Resource Management database. In conclusion, we show that a connectionist component improves a state-of-the-art HMM system.