Broadcast news

This paper presents an automatic sentence segmentation method for an automatic speech summarization system. The segmentation method is based on combining word-and class-based statistical language models to predict sentence and non-sentence boundaries. We study both the performance of the sentence segmentation system itself and the effect of the segmentation on the summarization accuracy. The sentence segmentation is done by modelling the probability of a sentence boundary given a certain word history with language models trained on transcriptions and texts from several sources. The resulting segmented data is used as the input to an existing automatic summarization system to determine the effect it has on the summarization process. We conduct all our experiments with two types of evaluation data: broadcast news and lecture transcriptions. The automatic summarizations are created with different sentence segmentations and different summarization ratios (30% and 40%) and evaluated by comparing them to human-made summaries. We show that a proper sentence segmentation is essential to achieve good performance with an automatic summarization system.

M ultimedia content continues to grow rapidly. Bridging the semantic gap is essential to exploiting this growing data. Toward this goal, recent research has focused on automatically tagging multimedia content to support end-user interactions such as searching, filtering, mining, content-based routing, personalization, and summarization. However, to date, there's been limited progress on standardizing the set of semantics for machine tagging. We describe a recent collaborative undertaking to develop the Large-Scale Concept Ontology for Multimedia. This effort is being led by IBM, Carnegie Mellon University, and Columbia University with participation from CyC corporation and various other research academic and industrial groups. (Some of the authors are leading this effort.) The Disruptive Technology Office (DTO) sponsored LSCOM as a series of workshops that brought together experts from multiple communities to create a taxonomy of 1,000 concepts for describing broadcast news video. The LSCOM taxonomy was designed to satisfy multiple criteria of utility, coverage, feasibility, and observability. Along with the taxonomy of 1,000 86

The NEMLAR project: Network for Euro-Mediterranean LAnguage Resource and human language technology development and support; (www.nemlar.org) is a project supported by the EC with partners from Europe and the Middle East; whose objective is to build a network of specialized partners to promote and support the development of Arabic Language Resources in the Mediterranean region. The project focused on identifying the state of the art of LRs in the region, assessing priority requirements through consultations with language industry and communication players, and establishing a protocol for developing and identifying a Basic Language Resource Kit (BLARK) for Arabic, and to assess first priority requirements. The BLARK is defined as the minimal set of language resources that is necessary to do any pre-competitive research and education, in addition to the development of crucial components for any future NLP industry. Following the identification of high priority resources the NEMLAR partners agreed to focus on, and produce three main resources, which are: 1) Annotated Arabic written corpus of about 500 K words, 2) Arabic speech corpus for TTS applications of 2x5 hours, and 3) Arabic broadcast news speech corpus of 40 hours Modern Standard Arabic. For each of the resources underlying linguistic models and assumptions of the corpus, technical specifications, methodologies for the collection and building of the resources, validation and verification mechanisms were put and applied for the three LRs.

The Linguistic Data Consortium at the University of Pennsylvania has recently been engaged in the creation of large-scale annotated corpora of broadcast news materials in support of the ongoing Topic Detection and Tracking (TDT) research project. The TDT corpora were designed to support three basic research tasks: segmentation, topic detection, and topic tracking in newswire, television and radio sources from English and Mandarin Chinese. The most recent TDT corpus, TDT3, added two tasks, story link and first story detection. Annotation of the TDT corpora involved a large staff of annotators who produced millions of human judgements. As with any large corpus creation effort, quality assurance and inter-annotator consistency were a major concern. This paper reports the quality control measures adopted by the LDC during the creation of the TDT corpora, presents techniques that were utilized to evaluate and improve the consistency of human annotators for all annotation tasks, and discu...

Effective human and automatic processing of speech requires recovery of more than just the words. It also involves recovering phenomena such as sentence boundaries, filler words, and disfluencies, referred to as structural metadata. We describe a metadata detection system that combines information from different types of textual knowledge sources with information from a prosodic classifier. We investigate maximum entropy and conditional random field models, as well as the predominant HMM approach, and find that discriminative models generally outperform generative models. We report system performance on both broadcast news and conversational telephone speech tasks, illustrating significant performance differences across tasks and as a function of recognizer performance. The results represent the state of the art, as assessed in the NIST RT-04F evaluation.

This paper presents a Korean large vocabulary continuous speech recognition system based on pseudomorpheme units. In Korean, an eojeol (word phrase) is a unit for spacing and a morpheme is the smallest unit with semantic meaning. If the eojeol is used as the dictionary and language modeling unit, the number of the unit becomes enormous. Instead we propose to use modified morpheme or pseudomorpheme as the basic recognition unit. We can recover the original eojeol by concatenating graphemes of pseudomorpheme components. We used a dictionary and language model with pseudomorpheme/part-ofspeech entries where each entry can have multiple pronunciations according to the morphology rule. With 32k-word vocabulary, the speaker-independent character, pseudomorpheme, and eojeol recognition accuracies on economy article database were 90.8%, 84.5%, and 81.3%, respectively.

This research explores the influence of mass media use and community context on civic engagement. The article presents a multilevel test of print, broadcast, and Internet effects on interpersonal trust and civic participation that acknowledges there are (a) micro-level differences in the motives underlying media use, (b) age-cohort differences in patterns of media use and levels of civic engagement, and (c) macro-level differences in community/communication context. Accordingly, the effects of individual differences in media use and aggregate differences in community context are analyzed within generational subsamples using a pooled data set developed from the 1998 and 1999 DDB Life Style Studies. The data suggest that informational uses of mass media are positively related to the production of social capital, whereas socialrecreational uses are negatively related to these civic indicators. Informational uses of mass media were also found to interact with community context to influence civic engagement. Analyses within subsamples find that among the youngest adult Americans, use of the Internet for information exchange more strongly influences trust in people and civic participation than do uses of traditional print and broadcast news media.

Most speech summarization research is conducted on broadcast news. In our viewpoint, spontaneous conversations are a more "typical" speech source that distinguishes speech summarization from text summarization, and hence a more appropriate domain for studying speech summarization. For example, spontaneous conversations contain more spoken-language characteristics, e.g. disfluencies and false starts. They are also more vulnerable to ASR errors. Previous research has studied some aspects of this type of data, but this paper addresses the problem further in several important respects. First, we summarize spontaneous conversations with features of a wide variety that have not been explored before. Second, we examine the role of disfluencies in summarization, which in all previous work was either not explicitly handled or removed as noise. Third, we breakdown and analyze the impact of WER on the individual features for summarization.

This article examines the structured panel discussion as a new form of broadcast news interaction. This involves live conversation among the anchorperson and news journalists on political news stories. The article draws upon the conversation analytic literature on news interviews, as well as detailed discourse analysis of journalistic discourse. By analysing data from Greek commercial prime-time news, it is argued that both the sequential organization and intra-turn design of journalists' talk help construct their professional role as that of an authoritative expert (analyst and opinionated commentator) on political current affairs. The rhetoric of expertise legitimizes the journalists' attribution of accountability, as well as their formulation of personal points of view. Given the absence of political actors from such extended exchanges, journalists are enabled to 'impose' their preferred readings of political actions and events on the audience. The structured panel discussion is a unique inter-journalistic conversational format, which exists alongside the more standard news interview, and is consequential for the representation of politics and political actors by the broadcast media.

This thesis considers the relationship between television foreign correspondents and their locally-hired 'fixers' in order to ascertain the centrality and significance of this relationship in facilitating international news production. The design of this research project, with its emphasis on news production practice, was guided by scholars in the sociology of news tradition, such as Jeremy Tunstall, Michael Schudson, Stephen Reese, Pamela Shoemaker and Simon Cottle. The main research question asks to what extent, and how, is the relationship DECLARATION This is to certify that i. the thesis comprises only my original work towards the PhD, ii. due acknowledgement has been made in the text to all other material used, iii. the thesis is less than 100,000 words in length, exclusive of tables, bibliographies and appendices.

This paper discusses a previously undescribed phenomenon in broadcast news inter- views, namely the practice of interviewees laughing in response to an interviewer's question prior to providing a substantive response. Specifically, it does so through an investigation of Hillary Rodham Clinton's (HRC) use of laughter in news interviews during her 2007 campaign for the Democratic nomination for President of the

Although satellite channels still have a central importance for the circulation of news in the Middle East and North Africa, the attention of many scholars has recently shifted towards the development of social media and their political use in the region. This article aims to focus on the convergence between the two media fields, specifically looking at how social networks’ contents have been selected and amplified by Al Jazeera English during the Tunisian and Egyptian 2011 revolutions. These issues are illustrated through the analysis of a selection of videos and posts published online by Al Jazeera English from 4 January to 11 February 2011.
Even though other media brands previously experimented with user-generated contents, the article argues that Al Jazeera English promoted a new configuration of ‘big’ and ‘small’ media in the MENA (Middle East and North African countries) region, based on the already well-established local media convergence.
The article also intends to explore the profile of the international public interacting with the channel’s website during the Egyptian Revolution. To do so, it analyses a selection of the comments left by the readers from 28 January to 11 February 2011, showing how during the revolts the channel became a reference point for a large international audience.

Nobel laureate Niko Tinbergen coined the term ‘supernormal stimulus’ after discovering that birds who lay small, pale blue eggs speckled with grey prefer to sit on larger, bright blue eggs with black polka-dots. He found that he could exaggerate the key qualities of any natural stimulus to produce a stronger reaction yet--be it nurturance, courtship, fear or aggression.
    Animals encounter supernormal stimuli mostly when experimenters build them.  Humans can generate their own and media are awash with them: friendly smiling faces more consistently welcoming than ones own friends, animated cartoon characters with wider eyes than any real child’s, seductive bodies exceeding any proportions found in nature, and a huge array of supernatural menaces.
    This chapter examines film, television, spectator sports, video games, and social websites as well as more traditional media such as print and musical performance in terms of the concept of supernormal stimuli.

This paper presents a real-time speech recognition system used to transcribe broadcast radio speech. The system is based on Abbot, the hybrid connectionist-HMM large vocabulary continuous speech recognition system developed at the Cambridge University Engineering Department 1]. Developments designed to make the system more robust to acoustic variability and to improve performance when decoding spontaneous speech are described. Modi cations necessary to increase the speed of the system so that it operates in real-time are also described. Recognition results and latency gures are presented for speech collected from broadcast news segments on BBC Radio 4.

This paper addresses the automatic transcription of audiovideo archives using a state-of-the-art broadcast news speech transcription system. A 9-hour corpus spanning the latter half of the 20th century has been transcribed and an analysis of the transcription quality carried out. In addition to the challenges of transcribing heterogenous broadcast news data, we are faced with changing properties of the archive over time, such as the audio quality, the speaking style, vocabulary items and manner of expression. After assessing the performance of the transcription system, several paths are explored in an attempt to reduce the mismatch between the acoustic and language models and the archived data.

This paper describes the development of an Arabic broadcast news transcription system. The presented system is a speaker-independent large vocabulary natural Arabic speech recognition system, and it is intended to be a test bed for further research into the open ended problem of achieving natural language man-machine conversation. The system addresses a number of challenging issues pertaining to the Arabic language, e.g. generation of fully vocalized transcription, and rule-based spelling dictionary. The developed Arabic speech recognition system is based on the Carnegie Mellon university Sphinx tools. The Cambridge HTK tools were also utilized at various testing stages. The system was trained on 7.0 hours of a 7.5 hours of Arabic broadcast news corpus and tested on the remaining half an hour. The corpus was made to focus on economics and sport news. At this experimental stage, the Arabic news transcription system uses five-state HMM for triphone acoustic models, with 8 and 16 Gaussian mixture distributions. The state distributions were tied to about 1680 senons. The language model uses both bi-grams and tri-grams. The test set consisted of 400 utterances containing 3585 words. The Word Error Rate (WER) came initially to 10.14 percent. After extensive testing and tuning of the recognition parameters the WER was reduced to about 8.61% for non-vocalized text transcription.

Since the nineteenth century, more kinds of news outlets and ways of presenting news grew along with telegraphic, telephonic, and digital communications, leading journalists, policymakers, and critics to assume that more events be-came available than ever before. Attentive audiences say in surveys that they feel overloaded with information, and journalists tend to agree. Although news seems to have become more focused on events, several studies analyzing U.S. news content for the past century and a half show that journalists have been including fewer events within their cover-age. In newspapers the events in stories declined over the twentieth century, and national newscasts decreased the share of event coverage since 1968 on television and since 1980 on public radio. Mainstream news websites continued the trend through the 2000s. Instead of providing access to more of the “what”, journalists moved from event centered to meaning-centered news, still claiming to give a factual account in their stories, built on a foundation of American realism. As journalists concentrated on fewer and bigger events to compete, audiences turned away from mainstream news to look for what seems like an abundance of events in digital media.

A crucial step in processing speech audio data for information extraction, topic detection, or browsing/playback is to segment the input into sentence and topic units. Speech segmentation is challenging, since the cues typically present for segmenting text (headers, paragraphs, punctuation) are absent in spoken language. We investigate the use of prosody (information gleaned from the timing and melody of speech) for these tasks. Using decision tree and hidden Markov modeling techniques, we combine prosodic cues with word-based approaches, and evaluate performance on two speech corpora, Broadcast News and Switchboard. Results show that the prosodic model alone performs on par with, or better than, word-based statistical language models-for both true and automatically recognized words in news speech. The prosodic model achieves comparable performance with significantly less training data, and requires no hand-labeling of prosodic events. Across tasks and corpora, we obtain a significant improvement over word-only models using a probabilistic combination of prosodic and lexical information. Inspection reveals that the prosodic models capture language-independent boundary indicators described in the literature. Finally, cue usage is task and corpus dependent. For example, pause and pitch features are highly informative for segmenting news speech, whereas pause, duration and word-based cues dominate for natural conversation.

In this project, we have developed a sign language tutor that lets users learn isolated signs by watching recorded videos and by trying the same signs. The system records the user's video and analyses it. If the sign is recognized, both verbal and animated feedback is given to the user. The system is able to recognize complex signs that involve both hand gestures and head movements and expressions. Our performance tests yield a 99% recognition rate on signs involving only manual gestures and 85% recognition rate on signs that involve both manual and non manual components, such as head movement and facial expressions.

Within the context of international benchmarks and collection specific projects, much work on spoken document retrieval has been done in recent years. In 2000 the issue of automatic speech recognition for spoken document retrieval was declared 'solved' for the broadcast news domain. Many collections, however, are not in this domain and automatic speech recognition for these collections may contain specific new challenges. This requires a method to evaluate automatic speech recognition optimization schemes for these application areas. Traditional measures such as word error rate and story word error rate are not ideal for this. In this paper, three new metrics are proposed. Their behaviour is investigated on a cultural heritage collection and performance is compared to traditional measurements on TREC broadcast news data.

The Arabic language presents a number of challenges for speech recognition, arising in part from the significant differences in the spoken and written forms, in particular the conventional form of texts being non-vowelized. Being a highly inflected language, the Arabic language has a very large lexical variety and with typically several possible (generally semantically linked) vowelizations for each written form. This paper summarizes research carried out over the last few years on speech-to-text transcription of broadcast data in Arabic. The initial research was oriented towards processing of broadcast news data in Modern Standard Arabic, and has since been extended to address a larger variety of broadcast data, which as a consequence results in the need to also be able to handle dialectal speech. While standard techniques in speech recognition have been shown to apply well to the Arabic language, taking into account language specificities help to significantly improve system performance. The "word error" rate is commonly used to measure speech recognition performance. It takes into account three types of errors: substitutions (a reference word is replaced by another word), insertions (a word is hypothesized that was not in the reference) and deletions (a word in the reference is missed). The word error rate is defined as #subs+#ins+#del # reference words , and is typically computed after a dynamic programming alignment of the reference and hypothesized transcriptions. Note that the word error can be over 100%.

We report on a series of experiments addressing the fact that German is less suited than English for word-based n-gram language models. Several systems were trained at different vocabulary sizes using various sets of lexical units. They were evaluated against a newly created corpus of German and Austrian broadcast news.

In this paper we describe a system we have developed for automatic broadcast-quality video indexing that successfully combines results from the fields of speaker verification, acoustic analysis, very large vocabulary speech recognition, content based sampling of video, information retrieval, natural language processing, dialogue systems, and MPEG2 delivery over IP. Our audio classification and anchorperson detection (in the case of news material) classifies video into news versus commercials using acoustic features and can reach 97% accuracy on our test data set. The processing includes very large vocabulary speech recognition (over 230K-word vocabulary) for synchronizing the closed caption stream with the audio stream. Broadcast news corpora are used to generate language models and acoustic models for speaker identification. Compared with conventional discourse segmentation algorithms based on only text information, our integrated method operates more efficiently with more accurate...

The use of exemplar-based methods, such as support vector machines (SVMs), k-nearest neighbors (kNNs) and sparse representations (SRs), in speech recognition has thus far been limited. Exemplar-based techniques utilize information about individual training examples and are computationally expensive, making it particularly difficult to investigate these methods on large-vocabulary continuous speech recognition (LVCSR) tasks. While research in LVCSR provides a good testbed to tackle real-world speech recognition problems, research in this area suffers from two main drawbacks. First, the overall complexity of an LVCSR system makes error analysis quite difficult. Second, exploring new research ideas on LVCSR tasks involves training and testing state-of-the-art LVCSR systems, which can render a large turnaround time. This makes a small vocabulary task such as TIMIT more appealing. TIMIT provides a phonetically rich and hand-labeled corpus that allows easy insight into new algorithms. However, research ideas explored for small vocabulary tasks do not always provide gains on LVCSR systems. In this paper, we combine the advantages of using both small and large vocabulary tasks by taking well-established techniques used in LVCSR systems and applying them on TIMIT to establish a new baseline. We then utilize these existing LVCSR techniques in creating a novel set of exemplar-based sparse representation (SR) features. Using these existing LVCSR techniques, we achieve a phonetic error rate (PER) of 19.4% on the TIMIT task. The additional use of SR features reduce the PER to 18.6%. We then explore applying the SR features to a large vocabulary Broadcast News task, where we achieve a 0.3% absolute reduction in word error rate (WER).

The objective of this study is to automatically extract annotated sign data from the broadcast news recordings for the hearing impaired. These recordings present an excellent source for automatically generating annotated data: In news for the hearing impaired, the speaker also signs with the hands as she talks. On top of this, there is also corresponding sliding text superimposed on the video. The video of the signer can be segmented via the help of either the speech or both the speech and the text, generating segmented, and annotated sign videos. We call this application as Signiary, and aim to use it as a sign dictionary where the users enter a word as text and retrieves sign videos of the related sign. This application can also be used to automatically create annotated sign databases that can be used for training recognizers.

In this paper two aspects of generating and using phonetic Arabic dictionaries are described. First, the use of single pronunciation acoustic models in the context of Arabic large vocabulary Automatic Speech Recognition (ASR) is investigated. These have been found to be useful for English ASR systems, when combined with standard multiple pronunciation systems. The second area examined is automatically deriving phonetic "pronunciations" for words that standard approaches, such as the Buckwalter Morphological Analyzer, cannot handle. Without pronunciations for these words the OOV rates for various Arabic tasks significantly increase. Here, pronunciations are automatically found by first deriving grapheme-to-phone rules, and associated rule probabilities. These are then used to produce the most likely pronunciation, or pronunciations, for any word. These approaches are evaluated on a large vocabulary Arabic Broadcast News and Broadcast Conversation transcription task. Both schemes are found to yield gains with a multi-pass/combination framework.

The immense prosodic variation of natural conversational speech makes it challenging to predict which words are prosodically prominent in this genre. In this paper, we examine a new feature, accent ratio, which captures how likely it is that a word will be realized as prominent or not. We compare this feature with traditional accentprediction features (based on part of speech and N -grams) as well as with several linguistically motivated and manually labeled information structure features, such as whether a word is given, new, or contrastive. Our results show that the linguistic features do not lead to significant improvements, while accent ratio alone can yield prediction performance almost as good as the combination of any other subset of features. Moreover, this feature is useful even across genres; an accent-ratio classifier trained only on conversational speech predicts prominence with high accuracy in broadcast news. Our results suggest that carefully chosen lexicalized featur...

The concept of marine protected areas is currently under scrutiny by members of the U.S. Congress, coastal natural resource management agencies, special interest stakeholder groups, and the public. As a surrogate for public opinion polls, a media content analysis was performed for news stories related to marine protected areas. The analysis examined more than 25,000 on-line newspaper, news wire, and radio and television broadcast news stories from 1995 to 2001 for expressions of four broad categories of uses and values associated with marine protected areas, expressions of favorable and unfavorable attitudes associated with marine protected areas, and the stakeholder groups involved in the discussion. Ecological uses and values were expressed more often than social, recreational, and commercial uses and values. Favorable attitudes were expressed more frequently than unfavorable attitudes. Over the period of analysis, 1995 to 2001, a gradual upward trend was found in frequencies of e...

In this paper I combine an overview of the goals and major approaches in cross-language information retrieval with some observations of current trends and with a report on a CLIR project that differs in many respects from most research activities in the fast growing area. In the overview, I will start from a generic model of an information retrieval system. Then the necessary extensions will be introduced that are needed for allowing queries in a language different from the document language. Several options for adding translation technology will be contrasted. I will then report on the research strategy followed in the EU-funded international project Mulinex. In this project a complete modular CLIR system was developed and integrated as the core software for a number of applications and as a plattform for research and technology development.

Since the Second World War, radio has been playing a significant role in the dissemination of political campaign information and influencing public opinion. This study investigated the role Yobe Broadcasting Corporation (YBC) Damaturu radio plays in reaching campaign information to rural areas in Yobe State, Nigeria. This study specifically aimed to find out the perception of voters in rural Yobe toward political campaign information broadcast on YBC radio and whether they were influenced by the information they heard on the medium to make voting decisions from 1999 (when democratic governance returned in the country after over 16 years of military rule) to 2006. Using cluster sampling method, 90 participants of voting ages were selected and interviewed at Babbangida and Maisandari rural communities. It was found that although majority of the electorate accessed political campaign information on YBC radio, only about a half of them had been influenced by the information to vote. Hence, this study concludes that political information broadcast on YBC is perceived with little creditability and that YBC stands the risk of losing its audience and patronage by prospective advertisers. This study recommends that YBC should be given sufficient autonomy and the present reach of the station should be widened.

This paper presents a statistical method for fingerprinting text. In a large collection of independently written documents each text is associated with a fingerprint which should be different from all the others. If fingerprints are too close then it is suspected that passages of copied or similar text occur in two documents. Our method exploits the characteristic distribution of word trigrams, and measures to determine similarity are based on set theoretic principles. The system was developed using a corpus of broadcast news reports and has been successfully used to detect plagiarism in students' work. It can find small sections that are similar as well as those that are identical. The method is very simple and effective, but seems not to have been used before

With the advent of mobile and portable devices such as cell phones and PDAs, wireless content distribution has become a major means of communications and entertainment. In such applications, a central authority needs to deliver encrypted data to a large number of recipients in such a way that only a privileged subset of users can decrypt it. A broadcasting news channel may face this problem, for example, when a large number of people subscribe to a daily exclusive news feature. This is exactly the kind of problem that broadcast encryption attempts to efficiently solve. On top of this, especially in the current digital era, junk content or spam is a major turn off in almost every Internet application. If all the users who subscribe to the news feed receive meaningless noise or any unwanted content, then the broadcaster is going to lose them. This results in the additional requirement that subscribers have source authentication with respect to their broadcaster. Broadcast signcryption, which enables the broadcaster to simultaneously encrypt and sign the content meant for a specific set of users in a single logical step, provides the most efficient solution to the dual problem of confidentiality and authentication. Efficiency is a major concern, because mobile devices have limited memory and computational power and wireless bandwidth is an extremely costly resource. While several alternatives exist in implementing broadcast signcryption schemes, identity-based (ID-based) schemes are arguably the best suited because of the unique advantage that they provide -any unique, publicly available parameter of a user can be his public key, which eliminates the need for a complex public key infrastructure. In ASIAN 2004, Mu et al. [25] propose what they call an ID-based authenticated broadcast encryption scheme, which is also a broadcast signcryption scheme, as the security goals are the same. They claim that their scheme provides message authentication and confidentiality and formally prove that the broadcaster's secret is not compromised, but in this paper, we demonstrate that even without knowing the broadcaster's secret, it is possible for a legal user to impersonate the broadcaster. We demonstrate this by mounting a universal forgeability attack -any valid user, on receiving and decrypting a valid ciphertext from a broadcaster, can generate a valid ciphertext on any message on behalf of that broadcaster for the same set of legal receivers to which the broadcaster signcrypted the earlier message, without knowing any secrets. Following this, we propose a new ID-based broadcast signcryption (IBBSC) scheme, and 2 Authors Suppressed Due to Excessive Length formally prove its security under the strongest existing security models for broadcast signcryption (IND-CCA2 and EUF-CMA2).

Although keyword-based queries are now a familiar part of any user's experience with the World Wide Web, they are of limited direct applicability to the vast and growing quantity of multimedia information becoming available in materials such as broadcast news, video teleconferences, reconnaissance data, and audio-visual recordings of corporate meetings and classroom lectures. Content-based indexing, archiving and retrieval would facilitate access to large databases of such materials.

Pronunciation modeling for large vocabulary speech recognition attempts to improve recognition accuracy by identifying and modeling pronunciations that are not in the ASR systems pronunciation lexicon. Pronunciation variability in spontaneous Mandarin is studied using the newly created CASS corpus of phonetically annotated spontaneous speech. Pronunciation modeling techniques developed for English are applied to this corpus to train pronunciation models which are then used for Mandarin Broadcast News transcription.