Note: All papers are available as PostScript or Adobe Acrobat Portable Document Format (PDF) files; click on the appropriate icon that follows the paper's title.

Somewhat over half of the papers are also viewable as Web-accessible HTML documents. Their names appear as hotlinks. If your computer does not have the capability to view either PDF or PostScript documents, click here for instructions on obtaining the necessary viewing software.

Abstracts

Author Index

Introduction and Background

Chair: David S. Pallett, NIST

• Overview of the 1997 DARPA Speech Recognition Workshop, David S. Pallett, NIST
  11 K 161 K
• Specification of the 1996 Hub 4 Broadcast News Evaluation, Richard Stern: CMU
  21 K 96 K
• The 1996 Broadcast News Speech and Language-Model Corpus, David Graff: Linguistic Data Consortium
  30 K 40 K
• Design and Preparation of the 1996 Hub-4 Broadcast News Benchmark Test Corpora, John S. Garofolo, Jonathan G. Fiscus, and William M. Fisher: National Institute of Standards and Technology
  85 K 664 K
• 1996 Preliminary Broadcast News Benchmark Tests, David S. Pallett, Jonathan G. Fiscus, and Mark A. Przybocki: National Institute of Standards and Technology
  2 M (complete) 610 K (paper only; for paper with figures, see the PDF version)

High-Level System Overview Session, Part 1

Chair: Phil Woodland, Cambridge University

• SRI System Overview, SRI [see SRI papers on Language Modeling and Acoustic Modeling].

• Development of the RU Hub-4 System, C. Che, D. Yuk, S. Chennoukh, and L. Flanagan: Rutgers University
  86 K 81 K
• NYU Language Modeling Experiments for the 1996 CSR Evaluation, Satoshi Sekine, Andrew Borthwick, and Ralph Grishman: New York University
  67 K 71 K
• Transcribing Broadcast News: The LIMSI Nov96 Hub4 System, J.L. Gauvain, G. Adda, L. Lamel, M. Adda-Decker: LIMSI
  156 K 1.6 M

High-Level System Overview Session, Part 2

Chair: Francis Kubala, BBN

• Transcription of Broadcast News Shows with the IBM Large Vocabulary Speech Recognition System, Raimo Bakis, Scott Chen, Ponani Gopalakrishnan, Ramesh Gopinath, Stephane Maes, Lazaros Polymenakos, and Martin Franz: IBM
  124 K 162 K
• The Development of the 1996 HTK Broadcast News Transcription System, P.C. Woodland, M.J.F. Gales, D. Pye, and S.J. Young: Cambridge University
  49 K 63 K
• Transcription of Broadcast Television and Radio News: The 1996 ABBOTT System, G.D. Cook, D.J. Kershaw, J.D.M. Christie, and A.J. Robinson: Cambridge University
  69 K 117 K
• The 1996 Hub-4 Sphinx-3 System, P. Placeway, S. Chen, M. Eskenazi, U. Jain, V. Parikh, B. Raj, M. Ravishankar, R. Rosenfeld, K. Seymore, M. Siegler, R. Stern, and E. Thayer: Carnegie Mellon University
  43 K 66 K
• The 1996 BBN Byblos Hub-4 Transcription System, Francis Kubala, Hubert Jin, Spyros Matsoukas, Long Nguyen, Rich Schwartz, John Makhoul: BBN
  116 K 106 K

Technical Session I: Acoustic Modeling

Chair: Richard Schwartz, BBN

• Acoustic Modeling in the CU-CON Broadcast News System, G.D. Cook and D.J. Kershaw: Cambridge University [see paper on The ABBOTT System]
• Acoustic Modeling in the 1996 HTK Broadcast News System, Cambridge University HTK [see paper on The HTK System]
• Acoustic Modeling in the LIMSI Nov96 Hub4 System, Jean-Luc Gauvain, Lori Lamel, Martine Adda-Decker: LIMSI [see paper on The LIMSI System]

Technical Session II: Miscellaneous Technical Papers

Chair: Salim Roukos, IBM

• Automatic Segmentation, Classification, and Clustering of Broadcast News Audio, Matthew A. Siegler, Uday Jain, Bhiksha Raj, and Richard M. Stern: Carnegie Mellon University
  26 K 233 K
• Efficient 2-Pass N-Best Decoder, Long Nguyen and Richard Schwartz: BBN
  142 K 123 K
• Some Results on Search Complexity vs Accuracy, Mosur K. Ravishankar: Carnegie Mellon University
  50 K 64 K
• Automatic Speaker Clustering, Hubert Jin, Francis Kubala, and Rich Schwartz: BBN
  122 K 119 K

Properties of the Broadcast News Data and Future Tests

Chair: Patti Price, SRI

Technical Session III: Adaptation: Environment/Channel/Speaker

Chair: Richard Stern, CMU

• Modeling Those F-Conditions - Or Not, Richard Schwartz, Hubert Jin, Francis Kubala, and Spyros Matsoukas: BBN
  109 K 100 K
• Speaker Adaptation and Environmental Compensation for the 1996 Broadcast News Task, Vipul N. Parikh, Bhiksha Raj, and Richard M. Stern: Carnegie Mellon University
  24 K 86 K
• Speaker Adaptation by Correlation (ABC), Scott Shaobing Chen and Peter DeSouza: IBM
  119 K 147 K
• Acoustic Modeling for the SRI Hub-4 Partitioned Evaluation Continuous Speech Recognition System, Ananth Sankar, Larry Heck, and Andreas Stolcke: SRI International
  62 K 101 K
• Practical Implementations of Speaker-Adaptive Training, Spyros Matsoukas, Rich Schwartz, Hubert Jin, and Long Nguyen: BBN
  174 K 251 K

Technical Session IV: Language Modeling

Chair: Roni Rosenfeld, CMU

• Language and Pronunciation Modeling in the CMU 1996 Hub 4 Evaluation, Kristie Seymore, Stanley Chen, Maxine Eskenazi, and Ronald Rosenfeld: Carnegie Mellon University
  72 K 93 K
• Language Modeling in the 1996 HTK Broadcast News System, Cambridge University HTK [see paper on The HTK System]
• Language Modeling and Lexicon in the LIMSI Nov96 Hub4 System, Gilles Adda, Lori Lamel, Martine Adda-Decker, Jean-Luc Gauvain: LIMSI [see paper on The LIMSI System]
• Hub-4 Language Modeling Using Domain Interpolation and Data Clustering, Fuliang Weng, Andreas Stolcke, and Ananth Sankar: SRI International
  57 K 64 K

TREC Speech Track/Spoken Document Retrieval Task

Chair: Donna Harman, NIST

• The TREC-6 Spoken Document Retrieval Track, Ellen Voorhees, John Garofolo: National Institute of Standards and Technology; Karen Sparck Jones: Cambridge University
  135 K 186 K
• Speech Recognition and Information Retrieval: Experiments in Retrieving Spoken Documents, Michael Witbrock and Alexander G. Hauptmann: Carnegie Mellon University
  69 K 470 K
• Experimental Results in Audio Indexing, S. Dharanipragada and S. Roukos: IBM
  117 K 114 K
• Topic Indexing of Broadcast News, Richard Schwartz: BBN
  54 K 266 K

Other Relevant Research

Chair: Candace Culhane, NSA

• Toward Automatic Recognition of Japanese Broadcast News, Tatsuo Matsuoka, Yuichi Taguchi, Katsutoshi Ohtsuki, Sadaoki Furui, and Katsuhiko Shirai: NTT Human Interface Labs and Waseda University
  28 K 117 K
• Other Relevant Research--Session Summary, Candace S. Culhane: National Security Agency
  9 K 20 K

• Specification of the 1996 Hub 4 Broadcast News Evaluation, Richard Stern: CMU
  This paper reviews and discusses the specification and implementation of the 1996 DARPA Hub 4 Continuous Speech Recognition Evaluation, using speech obtained from a variety of television and radio broadcast news programs. The benchmark test consisted of required partitioned evaluation (PE) and optional unpartitioned evaluation (UE) components. In addition, a variety of acoustical focus conditions were identified and separately annotated in the data used for the evaluation. Nine sites participated in the evaluation, with the best system reporting a word error rate of 27.1%. Differences between error rates for the UE and PE components were very small for two out of the three sites reporting results for these two conditions.
  21 K 96 K
• The 1996 Broadcast News Speech and Language-Model Corpus, David Graff: Linguistic Data Consortium
  The Linguistic Data Consortium handled the recording, digitization, and transcription of 130 hours of radio and television news broadcasts. Of this material, 50 hours' worth was designated and published as a baseline training set, and 20 hours were prepared and distributed by NIST for use as development and evaluation test data, for the 1996 ARPA CSR Benchmark Tests. The remaining 60 hours were held in reserve for use as additional training and evaluation test data in the 1997 benchmarks. The LDC also acquired, conditioned and published a five-year archive of commercially produced broadcast transcripts for use in constructing language models for the broadcast news domain. These tasks posed a broad range of novel challenges for the LDC, as well as for those at NIST and elsewhere who were involved in defining, clarifying and applying the standards and requirements for this corpus. This paper summarizes the content of the corpus, reviews the established specifications for the transcription format, and briefly describes the tools and methods used in the transcription effort.
  30 K 40 K
• Design and Preparation of the 1996 Hub-4 Broadcast News Benchmark Test Corpora, John S. Garofolo, Jonathan G. Fiscus, and William M. Fisher: National Institute of Standards and Technology
  This paper describes the procedures used in the preparation of the 1996 DARPA CSR Hub-4 Broadcast News Benchmark Test corpora and some analyses of that data. A new annotation/transcription process was designed and implemented to ensure that the transcripts were practically error-free and to negate the need to hold a post-test Adjudication as in years past. This paper focuses on this new annotation paradigm as well as an analysis of the properties of the test material.
  85 K 664 K
• 1996 Preliminary Broadcast News Benchmark Tests, David S. Pallett, Jonathan G. Fiscus, and Mark A. Przybocki: National Institute of Standards and Technology
  This paper documents use of Broadcast News test materials in DARPA-sponsored Automatic Speech Recognition (ASR) Benchmark Tests conducted late in 1996. In this year's tests, the source materials were broadened to incorporate both television and radio news broadcasts. A form of "partitioned evaluation" (PE) testing was implemented for the first time. At three sites, an additional testing protocol - similar to that used in last year's "Dry Run" tests - was used, now termed an "Unpartitioned Evaluation" (UE). Participants in these tests included nine groups at eight sites: BBN Systems and Technologies, Cambridge University (two groups), Carnegie Mellon University, IBM, LIMSI, New York University, Rutgers University, and SRI International.
  2 M (complete) 610 K (paper only; for paper with figures, see the PDF version)
• Development of the RU Hub-4 System, C. Che, D. Yuk, S. Chennoukh, and L. Flanagan: Rutgers University
  This paper describes preliminary development of a broadcast news transcribing system for this year's Hub4 evaluation. The recognition system uses CROWNS (developed at RU for the 1995 Hub3 tasks) with several modifications to handle the news programming task. Features such as model adaptation have been added to quickly provide acoustic models thought appropriate for the new task, even though the environment-dependent data are limited. The architecture of decoding is changed from one pass to multi-pass that can handle higher order language models more efficiently. Due to the short development period before evaluation, the preliminary system for this year's Hub4 test has produced a higher error rate than expected. In fact, its performance is found to be worse than our previous system when compared on the baselin broadcast speech. We have continued investigation since the test and performed diagnostic experiments. Results and error analysis are given in this report.
  86 K 81 K
• NYU Language Modeling Experiments for the 1996 CSR Evaluation, Satoshi Sekine, Andrew Borthwick, and Ralph Grishman: New York University
  This paper describes NYU's effort toward improving recognition accuracy for the 1996 ARPA Large Vocabulary Continuous Speech Recognition evaluation. We are trying to develop different kinds of language models including longer-range models and a linguistically motivated model. For the system described here, we used as a starting point the scores produced by SRI's acoustic and language models. These are linearly combined with the scores produced by the NYU language models. This paper also describes some experiments we tried which were not used in the official experiment, including experiments with perplexity minimization, Maximum Entropy modeling and parsing.
  67 K 71 K
• Transcribing Broadcast News: The LIMSI Nov96 Hub4 System, J.L. Gauvain, G. Adda, L. Lamel, M. Adda-Decker: LIMSI
  In this paper we report on the LIMSI Nov96 Hub4 system for transcription of broadcast news shows. We describe the development work in moving from laboratory read speech data to real-world speech data in order to build a system for the ARPA Nov96 evaluation. Two main problems were addressed to deal with the continuous flow of inhomogenous data. These concern the varied acoustic nature of the signal (signal quality, environmental and transmission noise, music) and different linguistic styles (prepared and spontaneous speech on a wide range of topics, spoken by a large variety of speakers). The speech recognizer makes use of continuous density HMMs with Gaussian mixture for acoustic modeling and n-gram statistics estimated on large text corpora. The base acoustic models were trained on the WSJ0/WSJ1 corpus, and adapted using MAP estimation with 35 hours of transcribed task-specific training data. The 65k language models are trained on 160 million words of newspaper texts and 132 million words of broadcast news transcriptions. The problem of segmenting the continuous stream of data was investigated using 10 Market Place shows. The overall word transcription error of the Nov96 partitioned evaluation test data was 27.1%.
  156 K 1.6 M
• Transcription of Broadcast News Shows with the IBM Large Vocabulary Speech Recognition System, Raimo Bakis, Scott Chen, Ponani Gopalakrishnan, Ramesh Gopinath, Stephane Maes, Lazaros Polymenakos, and Martin Franz: IBM
  This paper gives an overview of the IBM Large Vocabulary Continuous Speech Recognition system used in the 1996 Hub4 evaluation. It describes the acoustic and language models and adaptation techniques in the partitioned and unpartitioned evaluations. Evaluation results, analysis and further experiments are reported.
  124 K 162 K
• The Development of the 1996 HTK Broadcast News Transcription System, P.C. Woodland, M.J.F. Gales, D. Pye, and S.J. Young: Cambridge University
  This paper describes our efforts in extending a large vocabulary speech recognition system to handle broadcast news transcription. Results using the 1995 DARPA H4 evaluation data set are presented for different front-end analyses and for the use of unsupervised model adaptation using maximum likelihood linear regression (MLLR). The HTK system for the 1996 H4 evaluation is then described. It includes a number of new features compared to previous HTK large vocabulary systems including decoder-guided segmentation, segment clustering, cache-based language modelling, and combined MAP and MLLR adaptation. The system makes multiple passes through the data and the detailed results of each pass are given. The overall word error rate obtained by the 1996 evaluation system was 27.5%, and a bug-fixed version reduced this to 26.6%.
  49 K 63 K
• Transcription of Broadcast Television and Radio News: The 1996 ABBOTT System, G.D. Cook, D.J. Kershaw, J.D.M. Christie, and A.J. Robinson: Cambridge University
  Abbot is a hybrid connectionist-HMM large vocabulary continuous speech recognition system developed at the Cambridge University Engineering Department. This uses a recurrent neural network acoustic model to map acoustic features into posterior phone probabilities. These posterior probabilities are then converted to scaled likelihoods and used as observation likelihoods for phone HMMs. This paper describes the development of the CU-CON system which participated in the 1996 ARPA Hub 4 Evaluations. The system is based on Abbot. The Hub 4 Evaluation task involves the transcription of broadcast television and radio news programmes. This is an extremely demanding task for state-of-the-art speech recognition systems. Typical programmes include a wide variety of speaking styles and acoustic conditions. These range from read speech recorded in the studio to extemporaneous speech recorded over telephone channels. Results are presented for the system at various stages of development, as well as for the final evaluation system.
  69 K 117 K
• The 1996 Hub-4 Sphinx-3 System, P. Placeway, S. Chen, M. Eskenazi, U. Jain, V. Parikh, B. Raj, M. Ravishankar, R. Rosenfeld, K. Seymore, M. Siegler, R. Stern, and E. Thayer: Carnegie Mellon University
  This paper describes the CMU Sphinx-3 system, and the configuration we used for the 1996 DARPA (Hub-4) evaluation. The model structure, acoustic modeling, language modeling, lexical modeling, and system structure are summarized. We also discuss the experimental results obtained with this system on the most recent DARPA evaluation, and some subsequent results are also discussed.
  43 K 66 K
• The 1996 BBN Byblos Hub-4 Transcription System, Francis Kubala, Hubert Jin, Spyros Matsoukas, Long Nguyen, Rich Schwartz, John Makhoul: BBN
  In this paper, we describe the BBN Byblos system used for the 1996 Hub-4 Partitioned Evaluation (PE) and Unpartitioned Evaluation (UE) tests. For the PE, we chose to ignore the segment feature labels that were given to the system as side-information so that our approach would generalize trivially to the UE. Moreover, we chose not to model specific channel conditions in the training because the observed gains were too small to warrant the additional system complexity required to support them. In the end, we estimated a single set of acoustic models from only 40 hours of broadcast news data. For the UE, the data was automatically segmented with a simple dual-gender phoneme recognizer that efficiently located pauses and changes in speakers' gender. After this preliminary stage of segmentation and gender-classification, our UE and PE systems were identical. We achieved a 30.2% word error rate on the PE test and 31.8% on the UE test - only a 5% relative degradation from our PE result.
  116 K 106 K
• Automatic Segmentation, Classification, and Clustering of Broadcast News Audio, Matthew A. Siegler, Uday Jain, Bhiksha Raj, and Richard M. Stern: Carnegie Mellon University
  Automatic recognition of broadcast feeds from radio and television sources has been gaining importance recently, especially with the success of systems such as the CMU Informedia system. In this work we describe the problems faced in adapting a system built to recognize one utterance at a time to a task that requires recognition of an entire half hour show. We break the problem into three components: segmentation, classification, and clustering. We show that a priori knowledge of acoustic conditions and speakers in the broadcast data is not required for segmentation. The system is able to detect changes in acoustics, recognize previously observed conditions, and use this to pool adaptation data. We also describe a novel application of the Symmetric Kullback-Leibler distance metric that is used as a single solution to both the segmentation and clustering problems. The three components are evaluated through comparisons between the Partitioned and Unpartitioned components of the 1996 ARPA Hub 4 evaluation test set.
  26 K 233 K
• Efficient 2-Pass N-Best Decoder, Long Nguyen and Richard Schwartz: BBN
  In this paper, we describe the new BBN BYBLOS efficient 2-Pass N-Best decoder used for the 1996 Hub-4 Benchmark Tests. The decoder uses a quick fastmatch to determine the likely word endings. Then in the second pass, it performs a time-synchronous beam search using a detailed continuous-density HMM and a trigram language model to decide the word starting positions. From these word starts, the decoder, without looking at the input speech, constructs a trigram word lattice, and generates the top N likely hypotheses. This new 2-pass N-Best decoder maintains comparable recognition performance as the old 4-pass N-Best decoder, while its search strategy is simpler and much more efficient.
  142 K 123 K
• Some Results on Search Complexity vs Accuracy, Mosur K. Ravishankar: Carnegie Mellon University
  This paper presents three different techniques applied in or developed during the 1996 Hub-4 broadcast news transcription task. First, an efficient shortest path graph search algorithm is applied to the word lattice created by Viterbi search, producing a globally optimum result. This reduces the word error rate by about 3-10% (relative), depending on the test set. The execution time is at or close to real time for most utterances. Second, a segmented N-best list generation algorithm is described for producing compact N-best lists for very long utterances. Finally, a temporal smoothing technique is compared to deleted interpolation. On one test set, temporal smoothing reduces the error rate by 3% for an 8% increase in search cost, while the latter improves by 6% for a 50% increase in search cost.
  50 K 64 K
• Automatic Speaker Clustering, Hubert Jin, Francis Kubala, and Rich Schwartz: BBN
  This paper presents a fully automatic speaker clustering algorithm, which consists of three components: building a distance matrix based on Gaussian models of the acoustic segments; performing hierarchical clustering on the distance matrix with the prior assumption that consecutive segments should be more likely to come from the same speaker; and selecting the best clustering solution automatically by minimizing the within-cluster dispersion with some penalty against too many clusters. We applied this automatic speaker clustering technique in 1996 Hub4 evaluation, and the results show that it contributed significantly to the word error rate (WER) reduction in unsupervised adaptation. From our experiments, the algorithm seldom misclassifies segments from the same speaker into different clusters. We used the same clustering procedure for both partitioned evaluation (PE) and unpartitioned evaluation (UE) tests. Experiments also show that this automatic speaker clustering algorithm improves unsupervised adaptation as much as the hand labeled ideal case where the clusters are generated based on true speaker, channel and background condition.
  122 K 119 K
• Modeling Those F-Conditions - Or Not, Richard Schwartz, Hubert Jin, Francis Kubala, and Spyros Matsoukas: BBN
  After several disappointing preliminary attempts to make condition-specific models, we decided that it would be more advantageous to spend our time just trying to improve the core recognition system and use general adaptation techniques to deal with variations. This simplified the system immensely and freed up people and also made the transition from the PE system to the UE system much easier. We describe our attempts at condition-specific modeling/adaptation/training. We show that the benefit for channel-specific modeling is smaller than that for general adaption procedures and we argue that the cost is too high.
  109 K 100 K
• Speaker Adaptation and Environmental Compensation for the 1996 Broadcast News Task, Vipul N. Parikh, Bhiksha Raj, and Richard M. Stern: Carnegie Mellon University
  In this paper we present and discuss the performance of the Maximum Likelihood Linear Regression (MLLR) adaptation algorithm for various environmental conditions presented in the 1996 ARPA Hub 4 Broadcast News task. We also present a comparison of the effects of cepstral compensation using Codeword Dependent Cepstral Normalization (CDCN) and MLLR on the likelihoods of speech data and on the corresponding recognition error rates. Finally we describe a possible extension of MLLR using a quadratic regression equation.
  24 K 86 K
• Speaker Adaptation by Correlation (ABC), Scott Shaobing Chen and Peter DeSouza: IBM
  This paper describes a new rapid speaker adaptation algorithm using a small amount of adaptation data. This algorithm, termed adaptation by correlation (ABC), exploits the intrinsic correlation among speech units to update the speech models. The algorithm updates the means of each Gaussian based on its correlation with means of the Gaussians which are observed in the adaptation data; the updating formula is derived from the theory of least squares. Our experiments on the ARPA NAB-94 evaluation and the ARPA Hub4-96 tasks indicate that ABC seems more stable than MLLR when the amount of data for adaptation is very small (~ 5 seconds), and that ABC seems to enhance MLLR when they are combined.
  119 K 147 K
• Acoustic Modeling for the SRI Hub-4 Partitioned Evaluation Continuous Speech Recognition System, Ananth Sankar, Larry Heck, and Andreas Stolcke: SRI International
  We describe the development of the SRI system evaluated in the 1996 DARPA continuous speech recognition (CSR) Hub4 partitioned evaluation (PE). The task for the Hub4 evaluation was to recognize speech from broadcast television and radio shows. Recognizing such speech by machines poses many challenges. First, the segments to be recognized could be very long. This introduces a problem in training and recognition because of the consequent increased system memory requirement. A simple segmentation technique is used to break long segments into shorter, more manageable lengths. The speech from broadcast news sources exhibits a variety of difficult acoustic conditions, such as spontaneous speech, band-limited speech, and speech in the presence of noise, music, or background speakers. Such background conditions lead to significant degradation in performance. We describe techniques, based on acoustic adaptation, that adapt recognition models to the different acoustic background conditions, so as to improve recognition performance. We also present a novel algorithm that clusters the test data segments so that the resulting clusters are homogeneous with respect to speakers. This is followed by acoustic adaptation to the individual clusters, resulting in a significant performance improvement. Finally, we briefly describe our studies in language modeling for the Hub4 evaluation which is detailed further in another paper in these proceedings.
  62 K 101 K
• Practical Implementations of Speaker-Adaptive Training, Spyros Matsoukas, Rich Schwartz, Hubert Jin, and Long Nguyen: BBN
  Speaker Adaptive Training (SAT) has been shown to achieve significant word error reductions relative to the common Speaker Independent (SI) training paradigm, but its high requirements in disk I/O and space make it impractical for training on more than a couple hundred speakers. In the 1996 Hub-4 evaluation, the 38 hours of broadcast news training data consist of approximately 2000 speakers, half of them having less than 20 seconds of speech. In this paper we propose three implementations of SAT that are practical for training sets with a few thousands of speakers. First we present a two-pass SAT procedure that is mathematically equivalent to the original SAT method, with significantly reduced requirements in disk space, but essentially double the training time. Then we describe the Inverse Transform SAT (ITSAT) and the Least Squares SAT (LSSAT), two approximations to the SAT parameter estimation with time and space requirements that match those of common SI training. We show that the ITSAT method suffers only 1% degradation relative to the original SAT method.
  174 K 251 K
• Language and Pronunciation Modeling in the CMU 1996 Hub 4 Evaluation, Kristie Seymore, Stanley Chen, Maxine Eskenazi, and Ronald Rosenfeld: Carnegie Mellon University
  We describe several language and pronunciation modeling techniques that were applied to the 1996 Hub 4 Broadcast News transcription task. These include topic adaptation, the use of remote corpora, vocabulary size optimization, n-gram cutoff optimization, modeling of spontaneous speech, handling of unknown linguistic boundaries, higher order n-grams, weight optimization in rescoring, and lexical modeling of phrases and acronyms.
  72 K 93 K
• Hub-4 Language Modeling Using Domain Interpolation and Data Clustering, Fuliang Weng, Andreas Stolcke, and Ananth Sankar: SRI International
  In SRI's language modeling experiments for the Hub4 domain, three basic approaches were pursued: interpolating multiple models estimated from Hub4 and non-Hub4 training data, adapting the language model (LM) to the focus conditions, and adapting the LM to different topic types. In the first approach, we built separate LMs for the closely transcribed Hub4 material (acoustic training transcripts) and the loosely transcribed Hub4 material (LM training data), as well as the North-American Business News (NABN) and Switchboard training data, projected onto the Hub4 vocabulary. By interpolating the probabilities obtained from these models, we obtained a 20% reduction in perplexity and a 1.8% reduction in word error rate, compared to a baseline Hub4-only language model. Two adaptation approaches are also described: adapting language models to the speech styles correlated with different focus conditions, and building cluster-specific LM mixtures. These two approaches give some reduction in perplexity, but no significant reduction in word error. Finally, we identify the problems and future directions of our work.
  57 K 64 K
• The TREC-6 Spoken Document Retrieval Track, Ellen Voorhees, John Garofolo: National Institute of Standards and Technology; Karen Sparck Jones: Cambridge University
  The Text REtrieval Conference (TREC) workshops provide a forum for different groups to compare retrieval systems on common retrieval tasks. The 1997 TREC workshop will feature a Spoken Document Retrieval task for the first time. This paper motivates the task and describes the measures to be used to evaluate the effectiveness of the retrieval methodologies.
  135 K 186 K
• Speech Recognition and Information Retrieval: Experiments in Retrieving Spoken Documents, Michael Witbrock and Alexander G. Hauptmann: Carnegie Mellon University
  The Informedia Digital Video Library Project at Carnegie Mellon University is making large corpora of video and audio data available for full content retrieval by integrating natural language understanding, image processing, speech recognition and information retrieval. Information retrieval of from corpora of speech recognition output is critical to the projects success. In this paper, we outline how this output is combined information from other modalities to produce a successful interface. We then describe experiments that compare retrieval effectiveness on spoken and text documents and investigate the sources of retrieval errors on the former. Finally we investigate how improvements in speech recognizer accuracy may affect retrieval, and whether retrieval will still be effective when larger spoken corpora are indexed.
  69 K 470 K
• Experimental Results in Audio Indexing, S. Dharanipragada and S. Roukos: IBM
  In this paper we describe the IBM Audio-Indexing System and present some experimental results on the performance of the system on an audio indexing task.
  117 K 114 K
• Topic Indexing of Broadcast News, Richard Schwartz: BBN
  This presentation outlines a new model for the topic classification of broadcast news which is capable of classification among thousands of topics.
  54 K 266 K
• Toward Automatic Recognition of Japanese Broadcast News, Tatsuo Matsuoka, Yuichi Taguchi, Katsutoshi Ohtsuki, Sadaoki Furui, and Katsuhiko Shirai: NTT Human Interface Labs and Waseda University
  In this paper we report on automatic recognition of Japanese broadcast-news speech. We have been working on large-vocabulary continuous speech recognition (LVCSR) for Japanese newspaper speech transcription and achieved reasonably good performance. We have recently applied our LVCSR system to transcribing Japanese broadcast-news speech. We extended the vocabulary to 20k words and trained the language models using newspaper texts and broadcast-news manuscripts. These two language models were applied to our evaluation speech sets. The language model trained using broadcast-news manuscripts achieved better results for broadcast-news speech than the language model trained using newspaper texts which achieved better results for newspaper speech. In preliminary experiments on Japanese broadcast-news transcription, we achieved a word accuracy of 79.3% for anchor-speakers' speech by using a language model trained using broadcast-news manuscripts and newspaper texts.
  28 K 117 K
• Other Relevant Research--Session Summary, Candace S. Culhane: National Security Agency
  This session provided a forum to present other on-going research of interest to the speech recognition community currently working on the DARPA Hub 4 (Broadcast News) task. The two areas covered were automatic speech recognition of Japanese Broadcast News, and the Hub 5 task - automatic speech recognition of conversational speech over the telephone using multilingual corpora.
  9 K 20 K