DTIC ADA469307: Speaker Adaptation of Language Models for Automatic Dialog Act Segmentation of Meetings

0.07 MB·English

by  Defense Technical Information Center

Speaker Adaptation of Language Models
for Automatic Dialog Act Segmentation of Meetings
Ja´chymKola´rˇ1,YangLiu2,ElizabethShriberg3,4
1UniversityofWestBohemia,DepartmentofCybernetics,Pilsen,CzechRepublic
2UniversityofTexasatDallas,DepartmentofComputer Science,Richardson,TX,USA
3SRIInternational, USA 4International ComputerScienceInstitute, USA
[email protected], [email protected], [email protected]
Abstract capturespeakers’idiosyncraticlexicalpatternsassociatedwith
DAboundaries.
Dialog act (DA) segmentation in meeting speech is important
GeneralLMadaptationhasbeenstudiedratherextensively
for meeting understanding. In this paper, we explore speaker
inspeechrecognitionandotherlanguageprocessingtasks,us-
adaptationofhiddeneventlanguagemodels(LMs)forDAseg-
ingbothsupervised[10,11]andunsupervised[12,13,14]ap-
mentationusingtheICSIMeetingCorpus. Speakeradaptation
proaches.AusefulsurveyofLMadaptationtechniquesisgiven
isperformedusingalinearcombinationofthegenericspeaker-
in[15]. Thetypicalapproachistousethetestwordstodeter-
independentLMandanLMtrainedononlythedatafromin-
mine the topic for the test document, and then use the topic-
dividual speakers. We test the method on 20 frequent speak-
specific LM or its combination with the generic LM. For the
ers,onbothreferencewordtranscriptsandtheoutputofauto-
sentencesegmentationtask,[16]successfullyusedbothacous-
maticspeechrecognition.Resultsindicateimprovementsfor17
ticandlexicaldatafromanotherdomain(withslightlydifferent
speakersonreferencetranscripts,andfor15speakersonauto-
definition for sentence boundaries) to aid automatic sentence
matictranscripts. Overall,thespeaker-adaptedLMyieldssta-
segmentationinmeetings.
tisticallysignificantimprovementoverthebaselineLMforboth
Although topic- and domain-based LM adaptation ap-
testconditions.
proaches have received significant attention in the literature,
IndexTerms: languagemodeladaptation,speakeradaptation,
muchlessisknownaboutLMadaptationforindividualtalkers.
meetings,dialogactsegmentation,sentencesegmentation
Akita and Kawahara [17] showed improved recognition per-
formanceusingLMspeakeradaptationbyscalingthen-gram
1. Introduction
probabilitieswiththeunigramprobabilitiesestimatedviaprob-
abilisticlatentsemanticanalysis. TurandStolcke[18]demon-
Segmentation of speech into sentence-like units is a crucial
stratedthatunsupervisedwithin-speakerLMadaptationsignif-
first step in spoken language processing, since many down-
icantlyreducedworderrorrateinmeetingrecognition. Unlike
streamlanguageprocessingtechniques(e.g.,parsing,automatic
previousworkinDAsegmentation(whichtypicallyfocuseson
summarization, information extraction and retrieval, machine
featuresormodelingapproachesinaspeaker-independentfash-
translation)aretypicallytrainedonwell-formattedinput(such
ion)orLMadaptation(mostlytopic-basedadaptationinthetask
as written text). Several different approaches have been em-
of speech recognition), in this study, we investigate whether
ployed for sentence segmentation, including hidden Markov
speakeradaptationofLMsmayhelpinautomaticDAsegmen-
models (HMMs), multilayer perceptrons, maximum entropy,
tationofmeetings.
conditional random fields, AdaBoosting, and support vector
The remainder of the paper is organized as follows. We
machines[1,2,3,4,5,6]. Manyoftheseapproachesrelyon
describeourapproachinSection2.Theexperimentalsetup,re-
bothacoustic(prosodic)andlexicalinformation.
sults,anddiscussionareshowninSection3;Section4provides
Studiesonsentencesegmentationhavebeenconductedin
conclusions.
different domains, including broadcast news, conversational
telephone speech, and meetings. Our task in this paper is to
automatically segment meeting recordings into sentences, or 2. Method
dialog acts (DAs) for this domain. In many real-life meet-
2.1. Hidden-eventLanguageModel
ingapplications,thespeakersareoftenknownbeforehandand
recordedonaseparatechannel.Moreover,manymeetingshave Foragivenwordsequencew1w2...wi...wn,thetaskofDAseg-
recurring participants, presenting the opportunity for adapting mentation is to determine which inter-word boundaries corre-
models to the individual talkers. Speaker adaptation methods spond to a DA boundary. We label each inter-word boundary
were first successfully used in the cepstral domain for speech as either a within-unit boundary or a boundary between DAs.
recognition [7, 8]. In [9], we evaluated speaker-dependent Forexample,intheutterance“yesweshouldbedonebynoon”,
prosodymodelsforDAsegmentationinmeetings. However,it there are two dialog acts: “yes” (an answer), and “we should
wasleftunansweredwhetherspeaker-dependentlanguagemod- bedonebynoon”(astatement). Eachendsinasegmentation
els(LMs)wouldalsobenefitDAsegmentation.Inthispaperwe boundary.
aimtoaddressthisquestion;ourgoalistotrytoadapttheLMto We use a hidden event LM [19] to automatically detect
Report Documentation Page Form Approved
OMB No. 0704-0188
Public reporting burden for the collection of information is estimated to average 1 hour per response, including the time for reviewing instructions, searching existing data sources, gathering and
maintaining the data needed, and completing and reviewing the collection of information. Send comments regarding this burden estimate or any other aspect of this collection of information,
including suggestions for reducing this burden, to Washington Headquarters Services, Directorate for Information Operations and Reports, 1215 Jefferson Davis Highway, Suite 1204, Arlington
VA 22202-4302. Respondents should be aware that notwithstanding any other provision of law, no person shall be subject to a penalty for failing to comply with a collection of information if it

does not display a currently valid OMB control number.
1. REPORT DATE 3. DATES COVERED
2007 2. REPORT TYPE 00-00-2007 to 00-00-2007
4. TITLE AND SUBTITLE 5a. CONTRACT NUMBER
Speaker Adaptation of Language Models for Automatic Dialog Act
5b. GRANT NUMBER
Segmentation of Meetings
5c. PROGRAM ELEMENT NUMBER
6. AUTHOR(S) 5d. PROJECT NUMBER
5e. TASK NUMBER
5f. WORK UNIT NUMBER
7. PERFORMING ORGANIZATION NAME(S) AND ADDRESS(ES) 8. PERFORMING ORGANIZATION
SRI International,333 Ravenswood Avenue,Menlo Park,CA,94025-3493 REPORT NUMBER
9. SPONSORING/MONITORING AGENCY NAME(S) AND ADDRESS(ES) 10. SPONSOR/MONITOR’S ACRONYM(S)
11. SPONSOR/MONITOR’S REPORT
NUMBER(S)
12. DISTRIBUTION/AVAILABILITY STATEMENT
Approved for public release; distribution unlimited
13. SUPPLEMENTARY NOTES
http://wwww.speech.sri.com/papers/IS07-kolar-p1271.pdf
14. ABSTRACT

15. SUBJECT TERMS
16. SECURITY CLASSIFICATION OF: 17. LIMITATION OF 18. NUMBER 19a. NAME OF
ABSTRACT OF PAGES RESPONSIBLE PERSON
a. REPORT b. ABSTRACT c. THIS PAGE Same as 4
unclassified unclassified unclassified Report (SAR)
Standard Form 298 (Rev. 8-98)
Prescribed by ANSI Std Z39-18
DA boundaries in the unstructured word sequence. The hid- 38.2%(onthewholecorpus). Togeneratethe“reference”DA
deneventLMdescribesthejointdistributionofwordsandDA boundariesfortheSTTwords, wealignedthereferencesetup
boundaries,PLM(W,B).Themodelistrainedbyexplicitlyin- to the recognition output with the constraint that two aligned
cludingtheDAboundaryasatokeninthevocabularyinword- wordscouldnotoccurfurtherapartthanafixedtimethreshold.
basedn-gramLM.Duringtesting,thehiddeneventLMusesthe ArobustestimationoftheinterpolationweightλinEq(1)
pairofwordandDAboundaryasthehiddenstateinHMM,and may be a problem because of data sparsity. In the jackknife
thewordsasobservations, andperformsaViterbiorforward- approach, one half of speaker’s test data is used to estimate
backwarddecodingtofindtheDAboundariesgiventheword λfortheotherhalf, andviceversa. Wetesttwomethods for
sequence. We used trigram LMs with modified Kneser-Ney estimating λs. First, λs are estimated individually for each
smoothing[20]intheSRILMtoolkit[21]. speaker. Second, we set λ as the average value of the inter-
polationweightsacrossallthespeakers. Notethatinthelatter
2.2. SpeakerAdaptationApproach approach,however,westilluseonlydatafromthefirsthalvesof
individualtestsetstoestimateλforthesecondhalves,andvice
To adapt the generic speaker-independent LM to a particular
versa.Thisapproacheliminateshavingtwosignificantlydiffer-
speaker,weuseaninterpolationapproach.Thespeaker-adapted
entvaluesofλforasinglespeaker,whichdidoccurforsomeof
modelSAisobtainedfromalinearcombinationofthespeaker-
the20analyzedspeakers. Itindicatedthatforthosespeakers,
independentmodelSI andaspeaker-dependentmodelSDas
there is a mismatch in the two halves of the test data used in
follows:
jackknife,andthustheweightswerenotoptimizedproperlyfor
PSA(ti|hi;λ)=λPSI(ti|hi)+(1−λ)PSD(ti|hi) (1) thetestset.
where ti denotes a token (word or DA boundary) and hi its
3.2. EvaluationMetrics
historyofn−1tokensinann-gramLM.λisaweightingfac-
torthatisempiricallyoptimizedonheld-outdata. Wecompare WemeasureDAsegmentationperformanceusinga“boundary
differentapproachestoestimateλs,asdescribedinSection3. errorrate”[1]:
NotethattheSD dataisalreadycontainedintheSI datafor I+M
BER= [%] (2)
LMtraining;therefore,thisinterpolationdoesnothelpreduce NW
out-of-vocabularyrate,itrathergivesalargerweightton-grams
whereI denotesthenumberoffalseDAboundaryinsertions,
observedinthedatacorrespondingtoaparticularspeakerand
M thenumberofmisses,andNW thenumberofwordsinthe
isexpectedtobebettersuitabletothisspeaker.
testset.Inaddition,wereportoverallresultsusinganothercom-
mon metric, the NIST error rate. For DA segmentation, it is
3. ResultsandDiscussion defined as the number of misclassified boundaries divided by
thetotalnumberofDAboundariesinthereference. Itcanbe
3.1. DataandExperimentalSetup
expressedas
TheICSImeetingcorpus[22]containsapproximately72hours I+M
NIST = [%] (3)
ofmultichannelconversationalspeechdataandassociatedhu- NDA
mantranscripts,manuallyannotatedforDAs[23].Weselected
whereNDA denotesthenumberofDAboundariesinthetest
the top 20 speakers in terms of total words. Each speaker’s
set.Notethatthiserrorratemaybeevenhigherthan100%.
datawassplitintoatrainingset(∼70%ofdata)andatestset
(∼30%),withthecaveatthataspeaker’srecordinginanypar-
3.3. ResultsforIndividualSpeakers
ticular meeting appeared in only one of the sets. Because of
datasparsity, especiallyforthelessfrequentspeakers, wedid Table1showsacomparisonofDAsegmentationperformance
not use a separate development set, but rather jackknifed the forthebaselinespeaker-independentLMandspeaker-adapted
testsetinourexperiments. LMs for individual speakers, using reference transcripts. The
Thetotaltrainingsetforspeaker-independentmodels(com- speakers displayed in the table are sorted according to the to-
prisingthetrainingportionsofthe20analyzedspeakers,aswell talnumbersofwordstheyhaveinthecorpus. Thenumbersof
asalldatafrom32otherless-frequentspeakers)contains567k wordsinadaptationandtestsetsarealsolisted. Theresultsin-
words. DatasetsizesforindividualspeakersareshowninTa- dicatethatfor17of20speakers,performanceimprovedusing
bles 1 and 2; size of training sets used for speaker adaptation bothindividualandglobalweights,andtwootherspeakersim-
(referred to as “adaptation sets”) ranges from 5.2k to 115.2k provedonlyforoneofthetwointerpolationmethods.However,
words. We use the official corpus speaker IDs. The first let- the degree of the improvement varies across particular speak-
ter of the ID denotes the sex of the speaker (“f” or “m”); the ers. For8talkers,theimprovementwasstatisticallysignificant
secondletterindicateswhetherthespeakerisanative(“e”)or atp ≤ 0.05usingaSigntestforbothmethods. For4othersit
nonnative(“n”)speakerofEnglish. wassignificantforonlyoneofthemethods.
Weusetwodifferenttestconditions: referencetranscripts Table2reportsthecorrespondingresultsfortheSTTcondi-
(REF) and speech recognition output (Speech-To-Text, STT). tions. Notethatthetestsetsizesdifferfromtheprevioustable,
Recognition results were obtained using the state-of-the-art since the number of words in STT outputs is usually smaller
SRI CTS system [24], which was trained using no acous- thanthatinthecorrespondingreference. Theresultsshowthat
tic data or transcripts from the analyzed meeting corpus. To 15speakersimprovedusingbothinterpolationmethods,while
represent a fully automatic system, we also used automatic 4otherspeakersimprovedjustforoneofthemethods. Again,
speech/nonspeechsegmentation. Worderrorratesforthisdif- for8talkers,theimprovementwassignificantatp ≤ 0.05for
ficult data are still quite high; the STT system performed at bothmethods,andfor4otherstheimprovementwassignificant
Table1: Boundaryerrorrates(BER)[%]inREFerenceconditionsforindividualspeakers. ID=SpeakerID,#Adaptand#Testdenote
thenumberofwordsintheadaptationandtestsetsforeachspeaker,Baselinespeaker-independentmodelperformance,AdInWadapta-
tionwithindividualweights,andAdGlWadaptationwithglobalweights.IDsofspeakerswhoimprovedusingbothmethodsareshown
inboldface,*and**indicatethattheimprovementissignificantbyaSigntestatp<=0.05foroneorbothmethods,respectively.
ID #Adapt #Test Baseline AdInW AdGlW ID #Adapt #Test Baseline AdInW AdGlW
me013** 115.2k 51.2k 6.75 6.55 6.52 mn052 10.7k 3.8k 7.33 7.28 7.28
me011* 50.6k 24.8k 7.40 7.38 7.25 mn021** 9.6k 4.1k 6.68 5.41 5.65
fe008** 50.6k 22.6k 7.51 7.12 7.16 me003 9.3k 3.6k 8.78 8.45 8.56
fe016* 32.0k 15.4k 7.35 7.22 7.18 mn005** 7.7k 3.1k 7.83 7.01 6.92
mn015** 31.9k 14.7k 8.05 7.75 7.80 me045 8.1k 2.4k 8.90 8.94 8.90
me018* 31.8k 14.7k 6.64 6.43 6.45 me025 7.7k 2.4k 8.06 8.02 7.85
me010** 26.1k 12.6k 7.24 6.96 6.84 me006 6.9k 1.5k 9.53 10.32 9.47
mn007* 21.0k 10.1k 7.59 7.36 7.31 me026 5.2k 2.5k 5.80 5.76 5.80
mn017** 21.0k 7.1k 7.02 6.44 6.44 me012** 5.3k 2.1k 6.85 6.29 6.29
mn082 13.3k 4.2k 6.33 6.28 6.21 fn002 5.9k 1.5k 10.92 10.79 11.33
Table2: Boundaryerrorrates(BER)[%]inSpeech-To-Textconditionsforindividualspeakers. ID=SpeakerID,#Adaptand#Test
denotethenumberofwordsintheadaptationandtestsetsforeachspeaker,Baselinespeaker-independentmodelperformance,AdInW
adaptationwithindividualweights,andAdGlWadaptationwithglobalweights. IDsofspeakerswhoimprovedusingbothmethods
areshowninboldface, *and**indicatethattheimprovementissignificantbyaSigntestatp <= 0.05foroneorbothmethods,
respectively.
ID #Adapt #Test Baseline AdInW AdGlW ID #Adapt #Test Baseline AdInW AdGlW
me013** 115.2k 43.4k 8.29 8.16 8.18 mn052* 10.7k 3.5k 10.87 10.49 10.17
me011** 50.6k 22.9k 8.81 8.59 8.51 mn021* 9.6k 4.1k 8.20 7.83 7.73
fe008* 50.6k 19.5k 9.19 9.05 8.89 me003 9.3k 3.2k 9.36 9.36 9.33
fe016 32.0k 13.9k 8.42 8.40 8.31 mn005** 7.7k 3.0k 11.47 8.94 10.42
mn015** 31.9k 13.7k 10.16 9.90 9.84 me045 8.1k 2.1k 11.08 11.42 11.27
me018** 31.8k 13.3k 8.12 7.83 7.90 me025 7.7k 1.6k 14.36 14.23 14.36
me010** 26.1k 11.3k 8.39 7.96 7.91 me006* 6.9k 1.3k 10.94 10.01 10.40
mn007** 21.0k 8.4k 11.28 10.73 10.78 me026 5.2k 2.3k 7.35 7.01 6.88
mn017** 21.0k 6.0k 8.92 8.01 7.84 me012 5.3k 1.9k 8.68 8.15 8.31
mn082 13.3k 3.7k 10.37 10.45 10.10 fn002 5.9k 1.4k 13.40 13.40 12.89
foronemethod.Aninterestingobservationisthatforbothtest- withaspeaker-dependentmodeltrainedontherecognizerout-
ingconditions,therelativeerrorreductionachievedbyspeaker put. The idea was to allow the model also to adapt for error
adaptationisnotcorrelatedwiththeamountofadaptationdata. patternstypicalforanindividualtalker. However,thisadapta-
Thisfindingsuggeststhatspeakersdifferinherentlyinhowsim- tionperformedlesswellthanusingreferencetranscriptionsas
ilartheyaretothegenericspeaker-independentLM.Sometalk- thetrainingdata,whichindicatesthat,atleastwiththeamount
ers probably differ more and thus show more gain, even with of data available for our experiments, it is preferable to adapt
lessdata. LMs using clean data. In consequence it also suggests that
prospectiveunsupervisedapproachestoLMspeakeradaptation
3.4. OverallResults willperformlesswellthanthesupervisedapproach.
An overall comparison of performance of baseline speaker-
independentandspeaker-adaptedLMsispresentedinTable3.
The test set contains 203k words for REF and 180k words Table 3: Overall boundary error rates (BER) [%] and NIST
for STT conditions. These results show that for both condi- errorrates[%]inREFerenceandSTTconditions
tions,speaker-adaptedLMs—witheitherglobalinterpolation
Testconditions REF STT
weightsorindividualweights—outperformthebaseline. The
overallimprovementsbyLMspeakeradaptationforbothcon- Metric BER NIST BER NIST
ditionsarestatisticallysignificantatp < 10−15,usingaSign Baseline 7.30 48.56 9.06 68.65
test. Ofthetwoweightoptions,globalinterpolationresultsin Individualweights 7.02 46.74 8.79 66.61
better performance; however, the difference between the two Globalweights 6.99 46.56 8.76 66.38
approachesissignificantonlyatp<0.1. AdaptwithSTTdata N/A N/A 8.97 68.03
Inspeech-to-textconditionswealsotriedinterpolatingthe
speaker-independentmodeltrainedonreferencetranscriptions
4. Conclusions of Markov chains,” in IEEE Transaction on Speech and
AudioProcessing,vol.2,no.2.,pp.291–298,1994
Wehaveexploredspeakeradaptationofhiddeneventlanguage
modelsforautomaticDAsegmentationofmultipartymeetings. [8] Gales,M.J.F.:“MaximumLikelihoodLinearTransforma-
Thespeakeradaptationisbasedonalinearcombinationofthe tionsforHMM-basedSpeechRecognition,”inComputer
genericspeaker-independentandspeaker-dependentLMs. We SpeechandLanguage,vol.12.,pp.75–98,1998
evaluatedthemethodon20frequentspeakerswithawiderange [9] Kola´ˇr, J., Shriberg, E., Liu, Y.: “On Speaker-Specific
oftotalwordsavailableforLMadaptation. Weexaminedthe Prosodic Models for Automatic Dialog Act Segmenta-
approachusingbothreferencewordtranscriptsandrecognizer tionofMulti-PartyMeetings,” inProc.INTERSPEECH-
outputs. We found improvements for 17 speakers using ref- ICSLP2006,Pittsburgh,PA,USA,2006
erence transcripts, and for 15 speakers using automatic tran-
[10] Kneser, R., Peters, J., Klakow, D.: “Language Model
scripts.Overall,weachievedastatisticallysignificantimprove-
Adaptation Using Dynamic Marginals,” in Proc. EU-
mentoverthebaselineLMforbothtestingconditions.Theim-
ROSPEECH,Rhodes,Greece,1997
provementwasachievedevenforsomespeakerswhohadonly
arelativelysmallamountofdataavailableforadaptation. We [11] Seymore, K., Rosenfeld, R.: “Using Story Topics for
concludethatspeakeradaptationofLMsaidsDAsegmentation, LanguageModelAdaptation,”inProc.Eurospeech1997,
andthatfutureworkshouldinvestigatethepotentialofspeaker- Rhodes,Greece,1997
specificmodelingforothertasks. Otherimportantareasforfu-
[12] Gretter,R.,Riccardi,G.: “On-lineLearningofLanguage
tureextensionsincludetheintegrationoflexicalwithprosodic
Models with Word Error Probability Distributions,” in
or even multimodal information, and exploration of unsuper-
Proc.ICASSP,SaltLakeCity,UT,2001
visedapproaches.
[13] Niesler, T., Willett, D.: “UnsupervisedLanguageModel
Adaptation for Lecture Speech Transcription,” in Proc.
5. Acknowledgments
ICSLP,Denver,CO,2002
This material is based upon work supported by the Ministry [14] Nanjo, H., Kawahara, T.: “Unsupervised Language
ofEducationoftheCzechRepublic,projectsNo.1M0567and Model Adaptation for Lecture Speech Recognition,” in
ME909,andbytheU.S.DefenseAdvancedResearchProjects Proc.ISCAandIEEESSPR,Tokyo,Japan,2003
Agency (DARPA) under Contract No. NBCHD030010. Any
[15] Bellegarda, J.R.: “Statistical Language Model Adapta-
opinions, findings and conclusions or recommendations ex-
tion: Review and Perspectives,” in Speech Communica-
pressedinthismaterialarethoseoftheauthorsanddonotnec-
tion,vol.42,pp.93–108,2004
essarilyreflecttheviewsoftheDARPAortheDepartmentof
Interior-NationalBusinessCenter(DOI-NBC). [16] Cuendet, S., Hakkani-Tu¨r, D., Tur, G.: “Model Adapta-
tion for Sentence Segmentation from Speech,” in Proc.
6. References IEEE/ACL Workshop on Spoken Language Technology
(SLT),Aruba,2006
[1] Shriberg, E. et al.: “Prosody-based Automatic Segmen-
[17] Akita, Y., Kawahara, T.: “Language Model Adapta-
tation of Speech into Sentences and Topics,” in Speech
tion based on PLSA of Topics and Speakers,” in Proc.
Communication,vol.32,no.1–2,pp.127–154,2000
INTERSPEECH-ICSLP2004,Jeju,Korea,2004
[2] Warnke, V. et al.: “Integrated Dialog Act Segmenta-
[18] Tur, G., Stolcke, A.: “Unsupervised Language Model
tionandClassificationUsingProsodicFeaturesandLan-
Adaptation for Meeting Recognition,” in Proc. IEEE
guageModels”inProc.EUROSPEECH97,pp.207–210,
ICASSP2007,Honolulu,HI,2007
Rhodes,Greece,1997
[19] Stolcke, A. et al.: “Automatic Detection of Sen-
[3] Huang, J., Zweig, G.: “Maximum Entropy Model for tenceBoundariesandDisfluenciesBasedonRecognized
Punctuation Annotation from Speech,” in Proc. ICSLP Words,”inProc.ICSLP,pp.2247–2250,Sydney,1998
2002,Denver,CO,2002
[20] Chen,S.F.,Goodman,J.:“AnEmpiricalStudyofSmooth-
[4] Liu,Y.etal.:“UsingConditionalRandomFieldsforSen- ingTechniquesforLanguageModeling,” Tech.Rep.TR-
tenceBoundaryDetectioninSpeech,”inProc.ACL,Ann 10-98,CSGroup,HarvardUniversity,1998
Arbor,2005
[21] Stolcke,A.:“SRILM-AnExtensibleLanguageModeling
[5] Kola´ˇr, J., Shriberg, E., Liu, Y.: “Using Prosody for Au- Toolkit,”inProc.ICSLP,Denver,CO,2002
tomaticSentenceSegmentationofMulti-PartyMeetings,”
[22] Janin, A. et al.: “The ICSI Meeting Corpus,” in Proc.
inText,SpeechandDialogue(TSD)2006,LectureNotes
ICASSP-2003,HongKong,2003
inArtificialIntelligence(LNAI),vol.4188,pp.629–636,
Springer-Verlag,BerlinHeidelberg [23] Dhillon,R.etal.:“MeetingRecorderProject:DialogAct
LabelingGuide,”ICSITechnicalReportTR-04-02,Inter-
[6] Akita,Y.etal.: “SentenceBoundaryDetectionofSpon-
nationalComputerScienceInstitute,Berkeley,CA,2004
taneous Japanese Using Statistical Language Model and
Support Vector Machines,” in Proc. Interspeech-ICSLP, [24] Zhu, Q. et al.: “Using MLP Features in SRI’s Conver-
Pittsburgh,PA,2006 sational Speech Recognition System,” in Proc. INTER-
SPEECH2005,pp.2141–2144,Lisboa,2005
[7] Gauvain, J.L., Lee, C.H.: “Maximum A Posteriori Es-
timation for Multivariate Gaussian Mixture Obsevation