Appendix I Evaluating Speech Recognizers - Home - …978-1-4615-3650-5/1.pdf · Appendix I...

Appendix IEvaluating Speech Recognizers

1.1. Perplexity

The perfonnance of a speech recognizer is a function of severalvariables:

• The quality of acoustic modeling.

• The quality of language modeling.

• The constraint imposed by the grammar (if any).

• The inherent confusability of the vocabulary.Throughout this thesis, we have been concerned with improved acousticmodeling. However, in order to compare different systems or differentlanguage models, the other three factors must be taken into account.Perplexity is a measure of the constraint imposed by the grammar, or thelevel of uncertainty given the grammar.

Before we define perplexity, let's first consider how a grammar reducesuncertainty during recognition. Without a grammar, the entire vocabularymust be considered at every decision point. With a grammar, it is possible toeliminate many candidates from consideration, or to assign higherprobabilities to some candidates than others. This constraint at a decisionpoint (j) can be measured by entropy (H), or the number of bits necessary tospecify the next word using an optimal encoding scheme:

vH(W IJ) =-L pew IJ) . log 2 [pew IJ)] (1)

w=l

The perplexity at the decision point i is defmed to be:Q(w Ii) = 2H(w IJ) (2)

If we have a finite state grammar with many states, or decision points, thenentropy is computed as:

H(L) =L 'ltV) H(W I j)j

(3)

where 'It(j) is the steady-state probability of being in state i. The per-wordperplexity of this language is:

146

Q(L) = 2H(L)

AUTOMATIC SPEECH RECOGNITION

(4)

(6)

The above method for computing entropy and perplexity are usefulwhen the true language model is a finite state grammar with or withoutprobabilities. But in some cases (such as IBM's natural language task), thetrue language model is very different from the trigram model used, and itsperplexity cannot be measured. In other words, the perplexity measuredfrom the language model does not reflect the uncertainty encountered duringrecognition. When this is the case, test-set perplexity [Jelinek 85, Kimball86] should be used. Test-set perplexity is simply the geometric mean ofprobabilities at each decision point for the test set sentences, or:

I-log P(w l ,w2, ... ,wn) (5)n

where P(wl'w2, ... ,wn) is the probability of generating a string of n words.These n words are the concatenation of many sentences, with an end-ofsentence counting as one word. In the case of a bigram grammar, for onesentence:

P(w l ,w2, ... ,wm

) :::::

P(w l I Sent-begin)· P(w2 1wI)· .... P(Sent-end I wm)

For tasks whose true language models are known, it can be shown thatas the number of test sentences approaches infinity, test-set perplexity(Equation 5) is equivalent to perplexity (Equation 4). However, for taskswhose true language models are not known, test-set perplexity will be higherthan perplexity measured from the language model because it is inaccurate.In our word-pair and bigram grammars, the test-set perplexity should be thesame as the perplexity measured on the finite state grammars, because thetrue language models are known. The word-pair grammar has a perplexity ofabout 60 (which makes sense because there are 57,878 word pairs and 997words, and 60 is about 5~~8), and the bigram has a perplexity of about 20.

1.2. Computing Error RateFor isolated-word recognition, computing the error rate is

straightforward because the only possible type of error is substitution (anincorrect word was substituted for the correct word). However, in

EVALVATING SPEECH RECOGNIZERS 147

(8)

continuous speech, there are three types of errors: substitution, deletion (acorrect word was omitted in the recognized sentence), and insertion (an extraword was added in the recognized sentence). Clearly, substitutions anddeletions are errors. But it is not clear whether insertions should be countedas errors. On the one hand, a word not expected was inserted. This is clearlyundesirable, and should be penalized. On the other hand, it is not really fairto consider (recognize ~ wreck a nice) as three errors. Earlier work tendednot to count insertions as errors, while more recent ones counted them aserrors. In order to enable comparison against all systems, and because of thedoubtful nature of insertion errors, we report both results : percent correct,which doesn't consider insertions as errors, and word accuracy, which does.

To detennine the recognition accuracy, we first align the recognizedword string against the correct word string, and then compute the number ofwords Correct, Substitutions, Deletions, Insertions. This alignment can beobtained using a dynamic programming algorithm.18 Finally,Percent Correct and Word Accuracy are computed by:

CorrectPercent Correct= 100· (7)Correct Sent Length

E R 100Subs+Dels+Ins

rror ate= .....",....--~---~Correct Sent Length

Word Accuracy = I-Error Rate

Since Correct Sent Length = Correct+Subs + Dels,

Correct-InsWord Accuracy=100· -..",....-- --~

Correct Sent Length(9)

Therefore, Percent Correct and Word Accuracy differ by the number ofinsertions.

Confusions between homonyms are considered correct recognitionswhen no language model is used, because homonyms have the identicaldictionary entries, and are indistinguishable. When a grammar is used,homonym confusions are counted as substitution errors.

l8The actual dynamic programming program was provided by the National Bureau ofStandards.

Appendix IIThe Resource Management Task

11.1. The Vocabulary and the SPHINX PronunciationDictionary

AA42128AAWABERDEENABOARDABOVEADDADDEDADDINGAFFECTAFTERAGAINAJAXAJAX'SALASKAALERTALERTSALEXANDRIAALLANANCHORAGEANDANYANYBODYAPALACHICOLAAPALACHICOLA'SAPRILARABIANARCTICAREAREAAREASAREN'TARKANSASARKANSAS'SAROUNDARRIVALARRIVEARRIVEDARRIVING

AXEY F AO R T UW W AH N T UW EY TDEY EY D AH B AX Y UWAE B ER D IY NAX B AO R DDAXBAHVAE DDAE DX IX DDAE DX IX NGAX F EH K TDAE F T ERAX G EH NEY JH AE K SEY JH AE K S IX ZAXLAESKAXAX L ER TDAX L ER TSAE L IX G Z AE N D R IY AXAALAXNAE NG K R IH JHEH N DDEH N IYEH N IY B AH DX IYAE P AX L AE CH IX K OW L AXAE P AX L AE CH IX K OW L AX Z

EY P R LAX R EY B IY IX NAA R KD T IX KDAAREH R IY AXEH R IY AX ZAARNTDAA R K AX N S AOAA R K AX N S AO ZAX RAW N DDAX R AY V LAX RAY VAX R AY V DDAX RAY V IX NG

150

ARROWASASTORIAASUWASWATATLANTICAUGUSTAVAILABLEAVERAGEBADBADGERBADGER'SBAINBRIDGEBAINBRIDGE'SBANGKOKBARGEBASSBAYBEBEAMBEAMSBEENBEFOREBELOWBERINGBETTERBETWEENBIDDLEBIDDLE'SBISMARKBOMBAYBOTHBOXBRIGHTBRITISHBROOKEBROOKE'SBRUNSWICKBRUNSWICK'SBUD-TESTBUMP

BYC-CODEC-CODESC-RATINGC-RATINGSCl


AE ROWEH ZAE S D AO R IY AXEY EH S Y UW D AH B AX Y UWEY EH S 0 AH B AX Y UWEH TDAE TO L AE N IX KDAA G AX S TDAXVEYLAXBLAE V AX R IX JHB AE DDBAEJHERB AE JH ER ZB EY N B R IH JHB EY N B R IH JH IX ZB AE NG K AA KDBAARJHBAESB EYB IYB IY MB IY M ZB IH NB IX F AO RB IX L OWB EH R IX NGB EH DX ERB IH T W IY NB IH DX LB IH DX L ZB IH Z MAAR KDBAAMBEYB OW THB AA K SBRAY TDB R IH DX IX SHB R UH KDB R UH K SB R AH N Z W IH KDB R AH N Z W IH K SBAHT EH S TOB AH M PDB AYS IY K OW DDS IY K OW D ZS IY R EY DX IX NGS IY R EY DX IX NG ZS IY W AH N

mE RESOURCE MANAGEMENT TASK 151

C2C3C4CSCALIFORNIACAMDEN

CAMDEN'SCAMPBELLCAMPBELL'SCANCANADACAPABILITIESCAPABILITYCAPABLECAPACITIESCAPACITYCARRIERCARRIER'SCAlUUERSCAlUUERS'SCARRYCASREPCASREPEDCASREPSCASUALTYCAT-2CAT-3CAT-4CATEGORIESCATEGORYCENTERCENTEREDCENTERINGCEPCHANGECHANGEDCHANGINGCHANNELCHARTCHARTSCHATTAHOOCHEECHATTAHOOCHEE'SCHESHIRECHINACHOPCHOPPEDCHOPPINGCITRUS

S IY T UWS IY TH R IYS IY F AO RS IY F AY VIt AE L AX F AO R N Y AXIt AE M D IX NIt AI!: M D IX N ZItAEMBLKAEMBLZK IX NIt AE N AX DX AXIt EY P AX B IH L AX DX IY ZK EY P AX B IH L AX DX IYKEYPAXBLK AX P AE S IX DX IY ZK AX P AE S IX DX IYK EH R IY ERK EH R IY ER ZK EH R IY ER ZIt EH R IY ER ZIt AE R IYIt AE Z R EH PDIt AE Z R EH P DDItAEZREHPSIt AE SH L DX IYItAETUWIt AE TH R IYIt AE 'l'D F AO RK AE DX AX G AO R IY ZK AE DX AX G AO R IYS EH N ERS EH N ER DDS EH N ER IX NGS IY IY P IYCH EY N JHCH EY N JH DDCH EY N JH IX NGCHAENLCHAARTDCH AA R TSCH AE DX AX HH UW CH IYCH AE DX AX HH UW CH IY ZCH EH SH ERCH AY N AXCH AA PDCHAAPTDCH AA P IX NGS IH T R AX S

152

CITRUS'SCITYCLEARCLEAREDCLEARINGCLEVELANDCLEVELAND'SCLOSECLOSERCLOSESTCODAGCODECODESCOLORCOMPAREDCONFIDENCECONFIDENCE'SCONIFERCONIFER'SCONQUESTCONQUEST'SCONSTANTCONSTANT'SCONSTELLATIONCONSTELLATION'SCONVENTIONALCOOKCOPELANDCOPELAND'SCORALCOULDCOULDN'TCOUNTCOUNTEDCOUNTINGCROVLCROVLSCRUISERCRUISER'SCRUISERSCRUISERS'SCRUISINGCURRENTCURRENTLYDALEDALE'SDARWINDATA


S IH T R AX S IX SS IH DX IYK L IH RK L IH R DDK L IH R IX NGK L IY V L AX N DDK L IY V L AX N D ZK L OW SK L OW S ERK L OW S IX S TDKOWDAEGK OW DD

K OW D ZK AH L ERK AX M P EH R DDK AA N F IX DX EH N SK AA N F IX DX AX N S IX ZKAANAXFERKAANAXFERZIt AA N K W EH S TOIt AA N K W EH S TSItAANSTAXNTOK AA N S T AX N TSK AA N S T AX L EY SH AX NK AA N S T AX L EY SH AX N ZK AX N V EH N SH AX N LK UH KDK OW P L AX N DDItOWPLAXNDZK AO R LIt OR DDK OR D IX N TOK AW N TOIt AW N IX DDK AW N IX NGKROWVLKROWVLZItRUWZERKRUWZERZKRUWZERZKRUWZERZK R UW Z IX NGK ER AX N TOK ER AX N TD L IYD EY LD EY L ZD AA R W IX ND EY DX AH

THE RESOURCE MANAGEMENT TASK

DATE D EY TDDATED D EY DX IX DDDATES D EY TSDAVIDSON D EY V IX DD S AXNDAVIDSON'S D EY V IX DD S AXN ZDAY D EYDAYS D EY ZDDD992 D IY D IY D IY N AY N N AY N T OWDECEMBER D IX S EH M B ERDECREASE D IY K R IY SDECREASED D IY K R IY S TDDECREASING D IY K R IY S IX NGDEFAULT D IX F AO L TDDEFAULTS D IX F AO L TSDEFINE D IX F AY NDEFINED D IX F AY N DDDEFINING D IX F AY N IX NGDEFINITION D EH F IX N IH SH IX NDEFINITIONS D EH F IX N IH SH IX N ZDEGRADATION D EH G R AX D EY SH IX NDEGRADATIONS D EH G R AX D EY SH IX N ZDEGRADE D IY GREY DDDEGRADED D IY GREY DX IX DDDEGRADING D IY GREY DX IX NGDEGREES D IX G R IY ZDELETE D IX L IY TDDELETED D IX L IY DX IX DDDELETING D IX L IY DX IX NGDENVER D EH N V ERDENVER'S D EH N V ER ZDEPLOYED D IX P L OY DDDEPLOYMENT D IX P L OY M AX N TODEPLOYMENTS D IX P L OY MAX N TSDEPTH D EH P THDEPTHS D EH P TH SDESTINATION D EH S TAX N EY SH IX NDESTINATIONS D EH S T AX N EY SH IX N ZDID D IH DDDIDN'T D IH DD EN TDDIEGO-GARCIA D IY EY G OW G AA R S IY AXDIESEL D IY Z LDIM D IH MDISPLACEMENT D IH S B L EY S MAXN TDDISPLACEMENTS D IH S B L EY S MAXN TSDISPLAY D IH S B L EYDISPLAYED D IH S B L EY DDDISPLAYING D IH S B L EY IX NGDISTANCE D IH S T IX N S

153

154

DIXONDIXON'SDMDSDODOESDOESN'TDON'TDOWNESDOWNES'SDOWNGRADEDOWNGRADEDDRAFTDRAFTSDRAWDUBUQUEDUBUQUE'SDUEDURINGEACHEARLIEREARLIESTEARLYEASTEASTERNEASTPACEASTPAC'SECGOn

ECHOECONOMJ:CEDITEDITEDEDITINGEIGHTEIGHTEENEIGHTEENTHEIGHTHEIGHTYEISENHOWEREISENHOWER'SELEVENELEVENTHEMPLOYEDENDENDINGENGLANDENGLAND'SENGLISH


D IH It S IX ND IH It S IX N ZD IY EH M D IY EH SDOWDAHZDAHZAXNTOD OW N TDD AW N ZD AW N Z IX ZD AW N GREY DDD AW N GREY DX IX DDDRAEFTOD R AE F TSD R AODAXBYOWKDDAXBYOWItSDOWD ER IX NGIY CHER L IY ERER L IY IX S TDER L IYIY S TDIY S T ER NIY S TD P AE ItIY STOP AE It SIY S IY JH IY Z IY R OWFAORWAHN

EH It OWEH It IX N AA M IX KDEH DX IH TOEH DX IH DX IX DDEH DX IH DX IX NGEY TDEY T IY NEY T IY N THEY THEY D IYAY Z IX N HH AW ERAY Z IX N HH AW ER ZAX L EH V IH NAX L EH V IH N THEH M PLOY DDEH N DDEH N D IX NGIH NG G L AX N DDIH NG G L AX N D ZIH NG G L IH SH

lHE RESOURCE MANAGEMENT TASK 155

ENOUGHENROUTEENTERPRISEENTERPRISE'SEQUl:PMENTEQUl:PPEDESTEEMESTEEM'SESTl:MATEDETAETREVEREVERETTEXPECTEDFANNl:NGFANNl:NG'SFARFARTHERFARTHESTFASTFASTERFASTESTFEBRUARYFEET1'1'1'088Fl:FTEENFl:FTEENTHFl:FTHFl:FTYFl:GUREFl:GURESF:IJl:Fl:NDFl:REBUSHF1:REBUSH'SF1:RSTFIVEFl:XEDFLASHERFLASHER'SFLEETFLEETSFLl:NTFLl:NT'SFOOTERFORFORMOSAFORTY

l:H N AH FEH N R OW TOEH N ER PRAY ZEH N ER PRAY Z IX Zl:X It W l:H PD M AX N TOl:X It W l:H P TOEH S D l:Y MEH S D l:Y M ZEH S T AX M EY DX l:X DDl:Y T IY EYl:Y T l:Y AA REH V EREH V AX R l:H TOEH It S B EH It T IX DDI' AE N IX NGF AE N IX NG Z

FAARF AA R DH ERF AA R DH AX S TDFAESTDF AE S T ERI' AE S T IX S TDF EH B Y OW EH R IYI' l:Y TDEH F EH F EH I' Z IY R OW EY TD EY TDF l:H I' T IY NF 1:H I' T IY N THF IH I' THF IH F T IYF IH G ERF l:H G ER ZI' IY JH :IYF AY N DDF AY R B UH SHF AY R B UR SH IX ZF ER S TOF AY VF IH It S TOF L AE SH ERF L AE SH ER ZF L IY TDF L IY TSI' L l:H N TOI' L IH N TSF UH DX ERF ERF ER M OW S AXF AO R DX IY

156

FOURFOURTEENFOURTEENTHFOURTHFOXFOX'SFREDERICKFREDERICK'SFRENCH-POLYNESIAFRESNOFRESNO'SFRIDAYFRIDAY'SFRIGATEFRIGATE'SFRIGATESFRIGATES'SFROMFUELFULLGALLONSGALVESTONGASGETGIVEGLACIERGLACIER'SGNOMONICGOGOINGGONEGREATGREAT-CIRCLEGREATERGREATESTGREENGRIDGRIDLEYGRIDLEY'SGRILLGROSSGROUPGROUPSGUARDFISHGUARDFISH'SGUITARROGUITARRO'SGULF


F AO RF AO R T IY NF AO R T IY N THF AO R THF AA It SF AA It S IX ZF R EH 0 R IH KDF R EH 0 R IH K SF R EH N CH P AA L AX N IY SH AXF R EH Z N OWF R EH Z N OW ZFRAY OX EYFRAY OX EY ZF R IH G IH TOF R IH G IH TSF R IH G IH TSF R IH G IH TSF R AX MF YOWLFUHLGAELAXNZGAELVAXSTAXNGAESG EH TOG IH VG L EY SH ERG L EY SH ER ZN OW M AA N IH KDG OWG OW IX NGGAANGREY TOGREY TO S ER It LGREY OX ERGREY OX IX S TOG R IY NG R IH DOG R IH DO L IYG R IH DO L IY ZG R IH LGROW SGROW POGRUWPSG AA R DO F IH SHG AA R DO F IH SH IX ZG IH T AA R OWG IH T AA R OW ZG AA L F

TIlE RESOURCE MANAGEMENT TASK

HAD HH AE DDHALF HH AE FHARBOR HHAARB ERHARPOON HHAARP UW NHAS HH AE ZHASN'T HH AE Z IX N TOHAVE HH AE VHAVEN'T HHAEVAXNTDHAWKBILL HH AA I<D B IH LHAWKBILL'S HH AA 1m B IH L ZHE HH IYHE'S HH IY ZHECTOR HH EH I<D T ERHECTOR'S HH EH 1m T ER ZHEPBURN HH EH PO B ER NHEPBURN'S HH EH PD B ER N ZHER HH ERHERS HH ER ,ZHFDF EY CH EH F D IY EH FHIGH HH AYHIGHER HH AY ERHIGHEST HH AY IX S TDHIM HH IH MHIS HH IH ZHOME HH OW MHOMER HH OW M ERHONG-KONG HH AA NG K AA NGHONOLULU HHAANAHLUWLUWHOOKED HHUHKTDHORNE HH AO R NHORNE'S HH AO R N ZHOUR AW ERHOURS AW ER ZHOW HH AWHUNDRED HHAHND RAX DDICE-NINE AY S N AY N10 AY 0 IYIDENTIFICATION AY D EH N IH F AX K EY SH IX NIDENTIFICATIONS AY D EH N IH F AX K EY SH IX N Z

IF IH FIN IX NINCLUDE IH N K L UW DOINCLUDED IH N K L UW OX IX DOINCLUDING IH N K L UW OX IX NGINCREASE IH N K R IY SINCREASEO IH N K R IY S TDINCREASING IH N K R IY S IX NGINDEPENDENCE IH N D AX P EH N 0 IX N S

157

158

INDEPENDENCE'SINDIANINDONESIAINFORMATIONINSTALLEOINSTEADINSUFFICIENTINVOLVEDINVOLVINGI RONWooOIRONWOOD'SISISLANDSISN'TITIT'SITSJANUARYJAPANJARRETTJARRETT'SJARVISJARVIS'SJASONJASON'S.JULYJUNEJUPITERJUPITER'SKENNEOYKENNEDY'SICILOMETERICILOMZTERSICIRKICIRK'SICISICAICISKA'SKNOTKNOTSICOOIAKICOREAICOREANLAMPSLANTFLTLARGELARGERLARGESTLAST


IH N D AX P EH N 0 IX N S IX ZIH N 0 IY IX NIH N 0 AX N IY SH AXIH N F ER M EY SH IX NIH N S 0 AA L 00IH N S 0 EH 00IH N S AX F IH SH AX N TOIH N VAAL V 00IH N VAAL V IX NGAY R N W UH 00AY R N W UH 0 ZIH ZAY L AX N 0 ZIH Z EN TOIH TOIH TSIH TSJH AE N Y OW EH R IYJHAXPAENJH EH R IX TOJH EH R IX TSJH AA R V IX SJH AA R V IX S IX ZJH EY S IX NJH EY S IX N ZJH AX L AYJH OW NJH OW P AX OX ERJH OW P AX OX ER ZIC EH N AX DX IYIt EH N AX DX IY ZIt AX L AA M AX OX ERItAXLAAMAXOXERZIt ER ltDIt ER It SIt .IH S It AXIC IH S IC AX ZN AA TON AA TSIC OW OX IY AE !CDIt ER IY AXIt ER IY AX NLAEMPSL AE N TO F L IY TDLAARJHLAARJHERL AA R JH IX S TOLAESTO

TIlE RESOURCE MANAGEMENT TASK 159

STOT UW DDT UW D Z

L AO NL AO N ZERIXIHIH

L AE TDL AE TOL AE TOL EY DXL EY DXL AE DXL AE DXL AE TSL IY S TDL EH F TDL EH NG THL EH NG TH SL EH N IH N G R AE DDL EH SL EH DX ERL EH DX ER ZL EH V LL EH V L ZL IH NG K AX L EH V AX NL IH S TOL OW K EY DX IX DDL OW K EY SH IX NL OW K EY SH IX N ZL AA K W UH DDLAAKWUHDZLAANL AA NGL AA NG G ERLAANGGAXSTOL AA N JH IX T UW DDL AA N JH IX T UW D ZL AA N ZL OWL OW ERL OW IX S TOEH M K OW DDEH M K OW D ZEH M R EY DX IX NGEH M R EY DX IX NG ZEHMWAHNEH M T UWEH M TH R IYEH M F AO REH M F AY VM AE DX AX G AE S K ERM EY DDM EY 1mM EY K IX NG

LATLAT-LONLAT-LONSLATERLATESTLATITUDELATITUDESLATSLEASTLEFTLENGTHLENGTHSLENINGRADLESSLETTERLETTERSLEVELLEVELSLINK-llLISTLOCATEDLOCATIONLOCATIONSLOCKWOODLOCKWOOD'SLONLONGLONGERLONGESTLONGITUDELONGITUDESLONSLOWLOWERLOWESTM-CODEM-CODESM-RATINGM-RATINGSM1M2M3M4MSMADAGASCARMADEMAKEMAKING

160

MANCHESTERMANHATTANMANHATTAN'SMANILAMANYMARCHMARSMARS'SMAXMAXIMUMMAYMCCLUSKYMCCLUSKY'SMEMEASUREMERCATORMERCURYMERCURY'SMETEORMETEOR'SMETERSMETRICMEXICOMIAMIMIDGETTMIDGETT'SMIDPACM!DPAC'SMIDWAYMIDWAY'SMILEMILESMINDMINUTEMINUTESMISHAWAKAMISHAWAKA'SMISSIONMISSIONSMISSISSIPPIMISSISSIPPI'SMIWMOBMONDAYMONDAY'SMONGOLIAMONTHMONTH'S


M AE N CH EH S T ERMAENHHAETOENMAENHHAETOENZM AX N IH L AXM EH N IYMAAR CHMAAR SMAAR Z IX ZMAE K SM AE K S IX M AX MM EYM AX K L AH S K IYM AX K L AH S K IY Z

M IYM EH SH ERM ER K EY DX ERM ER K Y ER IYM ER K Y ER IY ZM IY DX IY ERM IY DX IY ER ZM IY DX ER ZM EH T R IX KDM EH K S IX K OWMAY AE M IYM IH JH IX TOM IH JH IX TSM IH DD P AE KDM IH DD P AE K SM IH DD W EYM IH DD W EY ZMAY LMAY L ZM AY N DDM IH N AX TOM IH N AX TSM IH SH AX W AA K AHM IH SH AX W AA K AH ZM IH SH AX NM IH SH AX N Z

M IH S IX S IH P IYM IH S IX S IH P IY ZEH M AY D AH B AX Y OWEH M OW B IYM AH N D EYM AH N D EY ZM AA N G OW L IY AXM AH N THM AH N TH S


MONTHSMONTICELLOMONTICELLO'SMOREMOSTMOZAMBIQUEMUCHN92762

NAMENAMESNAPLESNASHUANASHUA'SNEARNEARERNEARESTNEVERNEW

NEW-CALEDONIANEW-YORKNEW-ZEALANDNEWCASTLENEWERNEWESTNEXTNINENINETEENNINETEENTHNINETYNINTHNONOMENORTHNORTHERNNOTNOVANOVEMBERNOWNTDSNUCLEARNUMBEROAKLANDOCEANOCTOBEROFOFFOLD

M AH N TH SM AA N IX S EH L OWM AA N IX S EH L OW ZMAO RM OW S TOM OW Z AE M B IY KDM AH CHEH N N AY N T UW S EH V AX NS IH K S T UW

N EY MN EY M ZN EY P L ZN AE SH UW AXN AE SH UW AX ZN IY RN IY R ERN IY R AX S TDN EH V ERNUWNUWKAELAXDOWNYAXN UW Y AO R KDN UW Z IY L AX N DDNUWKAESLN UW ERN UW AX S TDNEHKSTDNAY NN AY N T IY NN AY N T IY N THN AY N DX IYN AY N THN OWN OW MN AO R THN AO R DH ER NN AA TDN OW V AXN OW V EH M B ERN AWEH N T IY D IY EH SN UW K L IY ERN AH M B EROWKLAXNDOW SH IX NAA KD T OW B ERAXVAO FAO L DD

162

OLDEROLDESTOLYMPIAONONEONLYOPENORORANGEOSGPOVERALLOVERLAYOVERLAYSPACPACFL'1'PACIFICPANAMAPARAMETERPARAMETERSPEARL-HARBORPEORIAPEORIA'SPERCEN'l'PERSIANPERSONNELPHILIPPINEPHILIPPINESPIGEONPIGEON'SPLH003

PLUCKPLUCK'SPLUNGERPLUNGER'SPOLLACKPOLLACK'SPORTPORT-ELIZABETHPORT-VICTORIAPORTSPOSITPOSITIONPOSITIONSPOSITSPOUGHKEEPSIEPOUGHKEEPSIE'SPOWERED


AOLDERAO L D AX S '1'DOW L IH M P IY AXAANWAHNOW N L IYOW P AX NAO RAO R AX N JHOW EH S JH IY P IYOW V ER AA LOWVERLEYOW V ER L EY ZP AE KDP AE KD F L IY '1'DP AX S IH F IX !CDPAENAXMAAP ER AE M AX DX ERP ER AE M AX DX ER ZP ER L HH AA R B ERP IY AO R IY AXP IY AO R IY AX ZP ER S EH N 'I'DP ER SH ENP ER S EN EH LF IH L AX P IY NF IH L AX P IY N ZP IH JH IX NP IH JH IX N ZP IY EH L EY CH Z IY R OWZ IY R OW '1'H R IY

PLAHKDPLAHKSPLAHNJHERP L AH N JH ER ZPAALAXKDPAALAXKSP AO R '1'DP AO R '1'D IY L IH Z AX B AX '1'HP AO R '1'D V IH KD '1' AO R 1Y AXP AO R '1'SP AA Z IX '1'DP AX Z IH SH AX NP AX Z 1H SH AX N ZP AA Z IX '1'SP AX K IH P S IYP AX K 1H P S IY ZP AW ER DD

THE RESOURCE MANAGEMENT TASK 163

PRAIlUEPRAIlUE'SPRESENTPREVIOUSPROBLEMPROBLEMSPROJECTIONPROPELLEDPROPULSIONPUFFERPUFFER'SPUGET-lQUARTERQUARTERSQUEENFISHQUEENFISH'SRADARRAMSEYRAMSEY'SRANGERRANGER'SRATEDRATHBUImERATHBURNE'SRATINGRATINGSREADINESSREASONREASONERREASONER'SREASONSRECENTRECLAIMl:RRECLAIMER'SREDREDEFINEREDEFINEDREDEFININGREDISPLAYREDOREDRAWREEVESREEVES'SREMAINI:NGREMARKREMARKSREPAI:RREPAIRED

P R EH R IYP R EH R IY ZP R EH Z AX N TOP R IY V IY AX SPRAABLAXMPRAABLAXMZP R AX JH EH K SH AX NP R AX P EH L DDP R AX P AH L SH AX NP AH I' ERPAHFERZP Y OW JH IH TD W AH NK W AO R DX ERK W AO R DX ER ZK W I:Y N F IH SHK W IY N F IH SH IX ZR EY DX AA RR AE M Z IYR AE M Z IY ZR EY N JH ERR EY N JH ER ZR EY DX IX DDR AE TH B ER NR AE TH B ER N ZR EY DX IX NGR EY DX IX NG ZR EH DX IY N EH SR IY Z AX NR IY Z N ERR IY Z N ER ZR IY Z AX N ZR IY S EN TOR IY K L EY M ERR IY K L EY M ER ZR EH DDR IY D IX F AY NR IY D IX F AY N DDR IY D IX F AY N IX NGR IY D IH S B L EYR IY D OWR IY D R AOR IY V ZR IY V Z IX ZR IY M EY N IX NGR IY MAAR 1mR IY MAAR K SR IY P EH RR IY P EH R DD

164

REPAIRINGREPLACEDREPORTREPORTEDREPORTINGREPORTSRESETRESOLUTIONRESOLVEDRESOURCERESOURCESREVIEWROSSSACRAMENTOSACRAMENTO'SSAILSAMESAMPLESAMPLE'SSAN-DIEGOSAN-FRANSASSAFRASSASSAFRAS'SSATURDAYSATURDAY'SSAVESCHENECTADYSCHENECTADY'SSCREENSCREENSSEASEAWOLI'SEAWOLF'SSECONDSEctnUTYSENSORSENSORSSEPTEMBERSETSETTINGSETTINGSSEVENSEVENTEENSEVENTEENTHSEVENTHSEVENTYSHASTASHASTA'S


R IY P EM R IX NGR IY P L EY S TOR IY P AO R TOR IY P AO R DX IX DDR IY P AO R DX IX NGR IY P AO R TSR IY S EH TOR EH Z AX L OW SH AX NR IX Z AA L V DDR IY S AO R SR IY S AO R S IX ZR IY V Y OWRAASSAEKRAXMEHNTOOWSAEKRAXMEHNTDOWZS EY LS EY MSAEMPLSAEMPLZS AE N D IY EY G OWSAENFRAENSAESAXFRAESS AE S AX F R AE S IX SS AE OX ER OX EYS AE DX ER DX EY ZS EY VS K AX N EH KD T IX DX IYS K AX N EH KD T IX DX IY ZS K R IY NS K R IY N ZS IYS IY W AO L FS IY W AO L I' SS EH K AX N DDS AX K Y UH R IH DX IYS EH N S ERS EM N S ER ZS EH PD T EH M B ERS EH TOS EH OX IX NGS EH DX IX NG ZS EM V AX NS EH V AX N T IY NS EM V AX N T IY N TMS EH V AX N TMS EH V AX N DX IYSH AE S T AXSH AE S T AX Z

THE RESOURCE MANAGEMENT TASK 165

SHESHE'SSHERMANSHERMAN'SSHIPSHIP'SSHIPSSHIPS'SSHOWSHOWINGSHOWNSIBERIANSIONEYSILSINCESINGAPORESIXSIXTEENSIXTEENTHSIXTHSIXTYSIZESLOWSLOWERSLOWESTSLQ-32SMALLSMALLERSMALLESTSOHOSOLOMONSONARSOONSOONERSOUTHSOUTHERNSOVIET-UNIONSPEEDSPEEDSSPS-40SPS-48Soo-23

STARTSTARTEDSTARTINGSTATIONSTATUS

SH IYSH IY ZSH ER M AX NSH ER M AX N ZSH IH POSH IH P SSH IH P SSH IH P SSH OWSH OW IX NGSH OW NS AY B IH R IY AX NS IH 00 N IYSILS IH N SS IH NG AX P AO RS IH It SS IH It S T IY NS IH It S T IY N THS IH It S THS IH It S T IYS AY ZSLOWSLOW ERSLOWAXSTnEH S EH L It Y OW TH ER OX IY T OWS M AO LS M AO L ERS M AO L IX S TnS OW HH OWSAALAXMAXNSOWNAARSOWNSOWN ERS AW THS AH OH ER NS OW V IY EH TO Y OW N Y AX NS B IY 00S B IY 0 ZEH S P IY EH S F AO R OX IYEH S P IY EH S F AO R OX IY EY TOEH S It Y OW It Y OW T W EH N IY

TH R IYSOAARTOS 0 AA R OX IX 00S 0 AA R OX IX NGS 0 EY SH AX NS 0 AE OX AX S

166

STEAMSTEREOGRAPHICSTERETTSTERETT'SSTRAITSUBSUB'SSUBICSUBMARINESUBMARINE'SSUBMARINESSUBMARINES'SSUBSSUBS'SSUFFICIENTSUMMARIZESUNDAYSUNDAY'SSUPPLIESSUPPLYSUPPOSEDSURFACESUSTAINEDSWITCHSWITCHESSWORDFISHSWORDFISH'SSYSTEMT-LAMTACANTAIWANTAKETAKENTASMTENTENTHTESTTEXASTEXAS'STFCCTHAILANDTHANTHATTHETHEIRTHEMTHERETHESE


S D IY MS D EH R IY IX G R AE F IX I<D

S D EH R IX TDS D EH R IX TSSDREYTDSAHBS AH B ZS OW B IH ItS AH B M ER IY NS AH B M ER IY N ZS AH B M ER IY N ZS AH B M ER IY N ZS AH B ZS AH B ZS AX F IH SH AX N TDS AH M ER AY ZS AH N D EYSAHNDEYZS AX PLAY ZS AX PLAYS AX P OW Z DDS ER F IX SS AX S D EY N DDS W IH CHS W IH CH IX ZS AO R DD F IH SHS AO R DD F IH SH IX ZS IH S T AX MT IY L AE MT AE It IX NT AY W AA NT EY ItTEYltAXNTAESAXMT EH NT EH N THT EH S TDT EH It S IX ST EH It S IX S IX ZT IY EH F S IY S IYTAYLAENDDDH EH NDH AE TDDH AXDH EH RDH EH MDH EH RDH IY Z


THEYTHIRDTHIRTEENTHIRTEENTHTHIRTIETHTHIRTYTHISTHOSETHOUSANDTHREATTHREATSTHREETHURSDAYTHURSDAY'STICONDEROGATICONDEROGA'STIMETIMESTOTODAYTODAY'STOGGLETOGGLEDTOGGLINGTOKYOTOMORROWTOMORROW'STONKINTONSTOTALTOWNSVILLETRACKTRACKSTRAININGTlUPOLITlUPOLI'STRUE-VIEWTUESDAYTUESDAY'S'l'URBINETURNTURNEDTURNINGTUSCALOOSATUSCALOOSA'STWELFTHTWELVETWENTIETH

DH EYTH ER DDTH ER 'I' IY NTH ER 'I' IY N THTH ER DX IY AX THTH ER DX IYDH IH SDH OW ZTH AW Z AX N DDTH R EH 'I'DTH R EH TSTH R IYTH ER Z D EYTH ER Z D EY Z'I' AY K AA N D AX R OW G AX'I' AY K AA N D AX R OW G AX Z'I' AY M'1' AY M ZTUW'1' IX DX EY'1' IX DX EY Z'1' AA G L'I' AA G L DD'1' AA G L IX NG'I' OW K IY OWTAXMAAROWTAXMAAROWZ'I' AO NG K IX NT AH N ZT OW DX LT AW N Z V IX LTRAEKDTRAEKSTREY N IX NG'1' R IH P AX L IYT R IH P AX L IY ZTRUWVYUWTUWZDEYT UW Z D EY ZT ER B IX NT ER NT ER N DDT ER N IX NGTAXSKAXLUWSAXTAXSKAXLUWSAXZ'I' W EH L F TH'I' W EH L VT W EH N IY IX TH

168

TWENTYTWOTYPETYPESUNITUNITED-STATESUNTILUOMUPDATEUPDATEDUPDATESUPGRADEUPGRADEDUSAUSEUSINGUSNVALUEVALUESVANCOUVERVANCOUVER'SVANDERGRIFTVANDERGRIFT'SVARIOUSVESSELVESSEL'SVESSELSVESSELS'SVIRGINIAVIRGINIA'SVISUALWABASHWABASH'SWADSWORTHWADSWORTH'SWASWASN'TWASPWASP'SWEWEDNESDAYWEDNESDAY'SWEEKWEEK'SWEEKSWELLINGTONWENTWERE


T W EH N IYTUWT AY PDT AY P SY UW N IH TOY UW N AY DX IX DD S D EY TSAX N T IX LY UW OW EH MAH PD D EY TOAH PD D EY DX IX DDAH PD D EY TSAH PD GREY DDAH PD GREY DX IX DDY UW EH S EYYUWZY UW Z IX NGYUWEHSEHNVAELYUWVAELYUWZV AE N K UW V ERV AE N K UW V ER ZV AE N D ER G R IH F TDV AE N D ER G R IH F TSV EH R IY IX SV EH S LV EH S L ZV EH S L ZV EH S L Z

V ER JH IH N Y AXV ER JH IH N Y AX ZV IH SH UW LW AA B AE SHW AA B All: SH XX ZW AA D Z W ER THW AA D Z W ER TH SWAHZWAHZAXNTDW AA S PDWAASPSW IYW EH N Z D EYW EH N Z D EY ZW IY KDW IY K SW IY K SW EH L IH NG T AX NW EH N TDW ER


WEREN'TWESTWESTERNWESTPACWESTPAC'SWHATWHAT'REWHAT'SWHENWHEN'LLWHEN'SWHEREWHERE'SWHICHWHIPPLEWHIPPLE'SWHOWHO'SWHOSEWHYWHY'SWICHITAWICHITA'SWILLWILLAMETTEWILLAMETTE'SWINAMACWINAMAC'SWINDOWWINDOWSWITHWITHINWITHOUTWON'TWORSEWORSTWOULDWOULDN'TYANKEEYEARYEARSYELLOWYESTERDAYYESTERDAY'SYETYORKTOWNYORKTOWN'SZERO

W ER N TOW EH S TOW EH S T ER NW EH S TO P AE KDW EH S TO P AE K SWAH TOW AH OX ERWAH TSW EH NW EH N IX LW EH N ZW EH RW EH R ZW IH CHW IH P LW IH P L ZHH OWHH OW ZHH OW ZWAYW AY ZW IH CH IX T AAW IH CH IX T AA ZW IH LW IH L AX M EH TOW IH L AX M EH TSW IH N AX M AE lIDW IH N AX M AE It SW IH N 0 OWW IH N 0 OW ZW IX THW IX TH IX NW IX OH AW TOW OW N TOW ER SW ER S TOW UH 00W UH 00 AX N TOY AE NG It IYY IH RY IH R ZY EH L OWY EH S T ER OX EYY EH S T ER OX EY ZY EH TOY AO R lID T AW NYAORlIDTAWNZZ IY R OW

170

ZULU

11.2. The Grammar

ZUWLUW


The resource management grammar is defined by a set of 900templates. To conserve space, we list the first 10 templates to provide aflavor of the task (italicized words bracketed by angle brackets are nonterminals):

<wlult-is> <optthe> <shipname's> <gross-ave> displacement in<long-metric> tons

is <optthe> <shipname's> <earliest> <casrep> rated worse than<optthe> hers

<list> <optthe> <threats>

<list> <optthe> <shipname' s> <casreps> from the last <digit> months

<show-list> <optthe> <shipname's> home port

<draw-show> <optthe> <shipname's> last <digit> <sensor> <posits>

is <optthe> <shipname's> remaining fuel insufficient to arrive in port at<optthe> <current> speed

<list> <optthe> <shipname's> <gross-ave> displacement and capabilities

<draw-show> <optthe> <shipname's> track in <bright-dim> <color>with <optthe> <shipname's> in <bright-dim> <color>

is <optthe> <shipname's> fuel capacity <greater-tluln> <optthe><shipname' s>

11.3. Training and Test SpeakersTable II-I enumerates all 120 speakers released by TI to eMU. Among

these speakers. the 80 training speakers and the first 25 evaluation speakerswere used to train SPHINX. The 10 March-87 evaluation speakers and the 6Oct-87 evaluation speakers were used to test SPHINX. Since one speaker isoverlapped between the two evaluation sets, there are actually 15 testspeakers. The first four characters of a speaker ID identify the speaker, thenext digit encodes the dialect of the speaker, and the last lener indicates male

THE RESOURCE MANAGEMENT TASK

or female. where available.

171

Each of the speakers spoke 40 sentences. For the training speakers. all40 sentences were used to train the HMMs. For the testing speakers. 10 ofthe sentences were designated by DARPA as evaluation sentences. and wereused only for fmal tests. The remaining 30 sentences were used to tune theparameters of SPHINX.

80 training speakers

adg04f ahh05m aksOlf apv03m bar07m bas04f bef03m bjk02mbma04m bmh05f bns04m bom07m bwm03m bwp04m cal03m cef03mceq08f cft04f cke03f cmb05m cmr02f crc05m csh03m cth07mcyl02f das05m daw18m dhs03m djh03f dlb02m dlh03m dlr07mdlr17m dms04f dmt02m drd06m dsc06m dtb03m eeh04f ejs08mers07m etbOlf fwk04m qjd04f qmd05f qxp04m hbs07m hes05fhpq03m jcs05f jemOlf jma02m jpq05m jrk06m jws04m jxm05fkes06m kkh05f lih05m ljc04m mah05f mcc05m mdm04m mqk02mmju06f mmh02f pqhOlm pql02m rcqOlm rqm04m rkmOlm rtk03mrwsOlm sdc05f tju06m tlb03m tpfOlm utb05f vlo05m wem05m

First 25 evaluation speakers

ajp06 bgt05 bpm05 cae06 chh07 cpmOl dtd05 ejl06 esd06esj06 qrlOl hjb03 hxs06 jlm04 jln08 jmd02 jsa05 laq06rav05 rddOl rjml2 sds06 tab07 tdp05 wbtOl

March-87 evaluation speakers

awf05 bcq18 bth07 ctt05 dabOldlc03 qwt02 jfc06 jfr07 sahOl

Oct-87 evaluation speakers

ctm02 dpkOl qwt02 ljd03 lmk05 sjk06

Table n·l: The list of all 120 speakers released by Tl.

Ap~endix IIIExamples or-SPHINX Recognition

In this appendix, we enumerate the results of SPHINX on the 150 testsentences, using the best SPHINX configuration described in Table 6-11. Foreach sentence, we show the correct sentence, as well as the recognizedsentence using the bigram grammar, the word-pair grammar, and nogrammar. Each word error is italicized, and insertions are designated with**

Correct:Bigram:Word-pair:None:


Correct:Bigram:Word·palr:None:






what's the mercury's average cruising speedwhat's the mercury's average cruising speedwhat's the mercury's average cruising speedwhat's the mercury's average cruising speed

what is the eta at her destination of fanningwhat is the eta at her destination of fanningwhat is the eta at her destination of fanningwhat is iJ eta at her destination of farming

how soon can esteem chop to atlantic fleethow soon can esteem chop to atlantic fleethow soon can esteem chop to atlantic fleethow sea again esteem chop to atlantic fleet

are there no ships that are in the mozambique channel•• jind any ships that are in the mozambique channeljind the nine ships that are in the mozambique channelby end as ships that their and the mozambique channel

draw a chart of ross seadraw·· chart of ross seadraw·· chart of ross seadraw·· chart overall •• sea

what was peoria's location and asuw area mission code july onewhat was peoria's location and asuw area mission code july onewhat was peoria's location and asuw area mission code july onewhat was peoria's location and asuw area mission ten july one

is rathbume located in wellington or aberdeenis rathbume located in wellington or aberdeenis rathbume located in wellington or aberdeenis rathbume located in wellington more aberdeen

what frigates in bering sea have both lamps and sps-48what frigates in bering sea have both lamps and sps-48what frigates in bering sea have both lamps and sps -48what frigates the bering sea have bad lamps in sps-48

174

Correct:Digram:Word-pair:None:




Correct:Digram:Word·palr:None:






Correct:Blgram:Word-pair:None:


what frigates in bering sea have both lamps and sps-48what frigates in bering sea have both lamps and sps-48what frigates in bering sea have both lamps and sps-48what frigates tM bering sea have bad lamps in sps-48

display the tracks and speeds of ships that are in solomon seadisplay the tracks and speeds of ships that are in solomon seadisplay the tracks and speeds of ships that are in solomon seadisplay the track since speeds tM ships that are dim solomon sea

display a chart centered aroWld jarrett using stereographic projectiondisplay a chart centered aroWld jarrett using stereographic projectiondisplay·· chart centered around jarrett using stereographic projectiondisplay·· chart set add around jarrett using stereographic projection

what link-II cruisers are in sea of japanwhat link-II cruisers are in sea of japanwhat f1tds and cruisers are in sea of japanwhat the yafl!cee addiflg cruisers are in sea Ofl japan

display the tracks of any cruisers in eastpacdisplay the tracks of any cruisers in eastpacdisplay the tracks of any cruisers in eastpacdisplay the tracks would any cruisers in eastpac

will firebush be at miami tomorrowwill firebush be at miami tomorrowwill firebush be at miami tomorrowwill firebush be yet miami tomorrow

what is mishawaka's percent fuelwhat is mishawaka's percent fuelwhat is mishawaka's percent fuelwhat Ihese mishawaka's to sel fuel

did mob mission area of the copeland ever go to m4 in nineteen eighty onedid mob mission area of the copeland ever go to m4 in nineteen eighty onegive mob mission area of .. copeland ever go to m4 in nineteen eighty onedid mob mission carry echo the code weill ever go to m4 in nineteen east one

are there more than four sps-40 capable frigates in port noware there more than four sps-40 capable frigates in port noware there more than four sps-40 capable frigates in port nowby their more badfull arefl't sps-40 capable frigates thafl port now

total the ships that will arrive in diego-garcia by next monthtotal the ships that will arrive in diego-garcia by next monthtotal the ships that will arrive in diego-garcia by next monthtotal a ships's pac will arriving" diego-garcia by next Ofl

redisplay overlay sobo turning on echOredisplay overlay soho turning on aprilredisplay overlay soho turning on plude flOW

redisplay overlays soho turning 10fl pac flOW

EXAMPLES OF SPHINX RECOGNITION 175












is mcclusky's destination the same as ramsey'si. mcclusky'. destination the same as ramsey'sis mcclusky's destination the same as ramsey'sis mcclusky's destination ** esteem has ramsey's

show on data .creen arkansas's track since four octobershow 1011 data screen arkansas's track since four octobershow 1011 data screen arkansas's track since four octobershow Ion beam screen arkansas's track since four october

is jason's maximum sustained speed slower than jupiter'sis jason's maximum sustained speed slower than jupiter'sis jason's maximum sustained speed slower than jupiter'sis jason's maximum sustained speeds slower than jupiter's

which submarines in bismark sea have tacanwhich submarines in bismark sea have tacanwhich submarines in bismark sea have tacanwhich submarine. on bismark sea have if tonlcin

what is the name and c-eode of the carrier in siberian seawhat is the name and c-eode of the carrier in siberian seawhat is the name and c-eode of the carrier in siberian seawhat is the name and c-eode of the carrier in siberian sea

do any ships that are in bass strait have an m5 miw m-ratingdo any ships that are in bass strait have an m5 miw m-ratingdo any ships that are in bass strait have an m5 miw m-ratingto rated ships that ** during by strait have the the m5 ofmiw won't m-rating

set the color of hooked track to bright redset •• color of hooked track to bright redset the color of hooked track to bright redset the color *. coolc chart to broolr.e red

is constant's last location closer than denver's to pac alertis constant's last location closer than denver's to pac alertis constant's last location closer than denver's to pac alertis constant's last the location closer than denver's to currellt alert

is copeland farther from sidney than the davidsonis copeland farther from sidney than the davidsonis copeland farther from sidney than the davidsonis copeland's earlier from sidney than the davidson

how far is the meteor from the midgellhow far is the meteor from •• midgellhow far is •• meteor from •• midgetlhow far is ** meteor from ** midgell

redefine area pacredefine area pacredefine area pacret:leji1ll!d area pac

176













how close is •• gulf of california to davidsonhow close is the gulf of california to davidsonhow close is gulf of california to davidsonhow close of is gulf ifcalifornia to davidson

when did seawolf degrade from her previous equipment c-ratingwhen did seawolf degrade from her previous equipment c-ratingwhen did seawolf degrade from her previous equipment c-ratingwhen did seawolf degrade from her previous equipment c-rating

what frigate in north atlantic ocean has the slowest current speedwhat frigate in north atlantic ocean has the slowest current speedwhat frigate in north atlantic ocean has the slowest current speedwhat frigate the north {at ocean has •• slowest kirk speed

how many submarines were in pan-victoria on the twentieth of marchhow many submarines were in pan-victoria on the twentieth of marchhow many submarines were in pan-victoria ofhomer twentieth of marchhow many submarines weren't •• pan-victoria •• homer twentieth of march

display a new chan projection using mercatordisplay a new chan projection using mercatordisplay a new chan projection using mercalordisplay" new chart projection using mercator

show grillshow grillshow grillshow grill

get the cruiser's locations for aprilget the cruiser's locations for aprilgel the cruiser's locations for aprilget •• cruisers locations for april

what is the average training rating code for usn ships that are in arctic oceanwhat is the average training rating code for usn ships that are in arctic oceanwhat is the average training rating codes for usn ships •• are in arctic oceanwhat using •• average training grun code for usn ships •• slerell •• arctic ocean

reset the switches to defaultsreset •• switches to defaultsreset •• switches to defaultsreset" switches to defaults

edit location data for track a42128edit location data for track a42128edit location data for track a42128edit location data for track a42128

do any vessels thaI are in gulf of lonkin have asw mission area of m4do any vessels that are in gulf of tonkin have asw mission area of m4do any vessels that are in gulf of tonkin have asw mission area of m4to any vessel's •• .. bering gulf both lonkin have asw mission area both m4






Correct:Digram:Word-paIr:None:





Correct:DIgram:Word-pair:None:


what's glacier's maximum draftwhat's glacier'. maximum draftwhat's glacier's maximum draftwhat'. glacier'. maximum draft

where was the brooke on january sixteenwhere was the brooke on january sixteenwhere was the brooke were ill january sixteenwMre's -- the brooke which years sixteen

clear data screenclear data screenclear data screencleared did screen

how many vessels are in indian oceanhow many vessels are in indian oceanhow many vessels are in indian oceanhow many vessels are -- indian ocean

what is the midway's fuel levelwhat is the midway's fuel levelwhat is the midway's fuel levelwhat/is/ the midway's fuel/elle/s

show the conquest's position seventeen august of eighty sixshow the conquest's position seventeen august /ell eighty sixshow the conquest'. position seventeen august/ell eighty sixshow the COIIqU/!SI position seventeen - - more ulimaled six

why did queenfish change equipment readiness twenty three maywhy did queenfish change equipment readiness twenty three maywhy did queenfish change equipment readiness twenty three maywhy eight queenfish change equipment readiness twenty three may

show percent fuel aboard mercuryshow percent fuel aboard mercuryshow percent fuel aboard mercuryshow percent fuel aboard were /hree

fmd crovls and tracks for tfcc frigates in north pacific oceanfmd crovls and tracks for tfcc frigates in north pacific oceanfmd crovls and tracks for tfcc frigates in north pacific oceanfmd crovls end tracks for tfcc frigates - - north pacific ocean

toggle sail and save switchestoggle sail and save switchestoggle sail and save switchestoggle sail all sea switches

where was frederick's destination november founeenthwhere was frederick's destination november founeenthwhere was frederick's destination november founeenthwhere was frederick's destination november founeenth

178













has swordfish reported any training problemshas swordfish reported any training problemshas swordfish reponed any training problemshas swordfish reponed any training problems

is tripoli's fuel capacity larger than tuscaloosa'sis tripoli's fuel capacity larger than tuscaloosa'sis tripoli's fuel capacity larger than tuscaloosa'sis tripoli's fuel capacity larger than tuscaloosa's

how many vessels were deployed since thiny one octoberhow many vessels were deployed since thiny one octoberhow many vessels were deployed since thiny one octoberhow many vessels were deployed since thiny one october

set switches to their defaultsset switches to their defaultsset switches to their defaultsset switches to their defaults

display virginia's displacement in metric tonsdisplay virginia's displacement in metric IOnsdisplay virginia's displacement in metric tonsdisplay virginia's displacement in metric tons

show posits of frigates that are in westpacshow posits of frigates that are in westpacshow positl of frigates that are in westpacshow posits of frigates that are again westpac

show supplies readiness of ironwood august oneshow supplies readiness of ironwood august oneshow supplies readiness of ironwood august oneshow supplies readiness ofof ironwood's august one

how early can fox be therehow early can fox be therehow early can fox be therehow all early can fox beam there

fmd full position data for all tracksfmd full position data for all tracksfmd full position data for all tracksfmd full position data four all tracks

what is midgett's percent fuelwhat is midgeu's percent fuelwhat is midgett's percent fuelwhat he'smidgett's percent fuel

which ships in manchester have a supplies readiness rating of c5which ships in manchester have an supplies readiness rating of c5which ships due in manchester have an supplies readiness rating of c5which ships's dim manchester have •• supplies readiness rating of c5













what casrep did firebush have on twenty seven maywhat casrep did flrebush have on twenty seven maywhat casrep did flrebush have on twenty seven maywhat casrep did firebush have one twenty seven may

how many ships are not ntds capablehow many ships are not ntds capablehow many ships are not ntds capablehome many ships are knot ntds capable

clear displayclear displayclear displayclear display

is pigeon's test depth greater than pluck'sis pigeon's test depth greater than pluck'sis pigeon's test depth greater than pluck'sis pigeon's test depth greater than pluck's

is citrus more than eighty kilometers from clevelandis citrus more than eighty kilometers from clevelandis citrus more than eighty kilometers from clevelandis citrus more than eighty kilometers from cleveland

what is the distance from the mishawaka to the monticellowhat is the distance from •• mishawaka to the monticellowhat is the distance from •• mishawaka to the monticellowent is the distance from •• mishawaka to the monticello

what is the name and the various ratings of the frigate in west bering seawhat is the name and the various ratings of the frigate in west bering seawhat is the name and the various ratings of the frigate in west bering seatwo what is •• sustained and the various ratings of the frigate in west bering sea

which subs that are c3 are in korean baywhich subs that are c3 are in korean baywhich subs that are c3 are in korean baywhich sub's •• sooner c3 are reporting •• bay

get the ships and their fleet identificationsget the ships in their fleet identificationsget the ships in their fleet identificationsget the ships do their fleet identifications

show the various fleet identifications for frigatesshow the various fleet identifications for frigatesshow the various fleet identifications for frigatesshow the various fleet identifications for frigates

on what day·· could dubuque arrive in port at his maximum sustained speed•• what would it tau dubuque arrive in port at his maximum sustained speedon what day can dubuque arrive in port at his maximum sustained speedher would date the dubuque arriving •• port •• hers maximum sustained speed

180










Correct:Blgram:Word·palr:None:



what is manhattan's fuel capacitywhat is manhattan's fuel capacitywhat is manhattan's fuel capacitywould is manhattan's fuel capacity

what is the asw average rating code for ships in formosa straitwhat is the asw average rating code for ships in formosa straitwhat is the asw average rating code for ships in formosa straitwhat is ** asw average rating code for ship show formosa strait

weren't more than ninety c I ships in pacific fleet todaywere more than ninety c I ships in pacific fleet todaywere more than ninety c1 ships in pacific fleet todaywere more the ninety SM one ships ofpacific fleet to at

how many kilometers is anchorage from new-yorkhow many kilometers is anchorage ofnew-yorkhow many kilometers is anchorage to new-yorkhow many kilometers does anchorage ** new-york

draw the tracks of all subs that are in gulf of tankindraw ** tracks for all subs that are in gulf of tonkindraw ** tracks of all subs that are in gulf of tonkindraw ** tracks ** will subs's ** ** centering gulf ** tonkin

show queenfish's location on twenty two february and its various capabilitiesshow queenfish' s location on twenty two february and its various capabilitiesshow queenfish's location on twenty two february and its various capabilitiesshow quun[lSh location on twenty two from weren', minutes hers capabilities

who has the least fuel remainingwho has ** least fuel remainingwho has ** least fuel remainingwho has ** least fuel remaining

give any cruisers that were c2 on eight augustgive any cruisers that were c2 on eight augustgive any cruisers that were c2 on eight augustgive any cruisers the were she two Ion eight august

give c5 ships in pacific fleetgive c5 ships in pacific fleetgive c5 ships in pacific fleetgive c5 ships an pacific fleet

what is the number of vessels that are in ross sea without slq-32what is the number of vessels that are in ross sea without slq-32what is the number of vessels that are in ross sea with pollack tM slq-32what is ** number ** vessels *. leiter .* ross sea with pollack slq-32

tum groups on and redraw the current areatum groups on and redraw the current areatum groups on and redraw the current areatum groups long thon redraw the current area

EXAMPLES OF SPHINX RECOGNlTION 181












which vessels in korean bay have a supplies readiness that is c3which vessels in korean bay have allY supplies readiness •• ofc3which vessels in korean bay have allY supplies readiness the •• c3which vessels in korean bay have he supplies readiness the .... c3

set the sail parameter to offset •• sail parameter turlled offset"· sail parameter turlled offshow persollnel •• parameter to ofoff

which link-II capable carriers have an equipment resource rating ofmore than c4which link-II capable carriers have an equipment resource rating of more than c4which link-II capable carriers have an equipment resource rating of more than c4which link-II capable carriers havell't·· equipment resource rating of ill were than c4

set switches to defaultsset switches to defaultsset switches to defaultsshow switches to defaults

what if apalachicola's propulsion type was steam turbine instead of gaswhat if apalachicola's propulsion type was steam turbine instead of gaswhat if apalachicola's propulsion type was steam turbine instead of gasone give apalachicola's propulsion type •• steam turbine instead of get bass

list vessels that were deployed on the fll'St of septemberlist vessels that werell't deployed on .... fll'St of septemberlist vessels and werell't deployed on •• first of septemberlist vessels all were deployed cOllifer's •••• of september

fmd frigates in honolulufmd frigates in honolulufmd frigates in honolulufmd frigates end honolulu

show the same chart with novashow the same chart with novashow the same chart with novashow·· sustained chart with nova

show the names of any submarines in yellow sea on twenty eight octobershow the names of any submarines in yellow sea on twenty .... octobershow the names of any submarines in yellow sea on twenty·· octobershow the names have any submarine's beam yellow sea on twenty eighty october

why was mercury's miw m-code changed on april twenty twowhy was mercury's miw m-code changed on april twenty twowhy was mercury's miw m-code changed on april twenty twowhy was where two reeelll miw m-eode challge on april twenty two

which ships in philippine sea are link-II capablewhich ships in philippine sea arell't link-I I capablewhich ships in philippine sea arell'tlink-II capablewhich ships in philippine .... e5 link-II capable

182


Corred:Digram:Word-pair:None:







Correct:Digram:Word-pair:

None:

Correct:Digram:Word.palr:None:



is there a problem with personnel for Ihe camdenis there a problem with personnel from the camdenis there a problem is persoMel for alilat is camdenis bering •• prob/~ms•• persOMelfar lat camden

are there any cruisers longer Ihan nineteen hlUldred meters that are in siberian sea•• find any cruisers longer than nineteen hundred meters that are in siberian sea•• futd any cruisers longer Ihan nineteen hundred meters that are in siberian seago,,~ Ihere many cruisers 10" i" than nineteen hundred me is that ar~as •• siberian sea

increase letter size to maximum value and redrawincrease letter size to maximum value and redrawincrease lener size to maximum value and redrawincrease 1m size •• maximum value in redraw

show downes's radar latitudes and longitudes using novashow downes's radar latitudes and longitudes using novashow downes's radar latitudes and longitudes using novashow dow,,~s radar latitudes i"longitudes using nova

find me the mission area ratings for arkansasfind me the mission area ratings for arkansasfmd me Ihe mission area ratings for arkansasfmd th~s~ than mission area ratings fiv~ arkansas

tum areas off and redraw current areatum areas off and redraw current areatum areas off and redraw current areatum areas half i" redraw current area

what is the mob m-code for samplewhat is Ihe mob m-code for samplewhat is Ihe mob m-code for samplewhat is be mob t~" code for sample

how soon does fresno arrive in townsvillehow soon does fresno arrive in townsvillehow soon does fresno arrive in townsville•• has,,' t is fresno arrive in townsville

were there more Ihan fifteen pacific fleet vessels employed in nineteen eighty threewere there more than fifteen pacific fleet vessels employed in nineteen eighty threewere there more than fifteen pacific fleet vessels b~t" employed in nineteen eighty Ihreefor fifthwere Ihere more than fifteen pacific fleet vessels chploy~d •• nineteen eighty Ihree off

is the wasp's last location closer Ihan wichita's to osgpis the wasp's last location closer Ihan wichita's to osgpis the wasp's last location closer Ihan wichita's to osgpis •• last last location closer Ihan wichita's to thailand list c~p

has •• home's miw mission area gone to m3 before twenty two augusthas mars on miw mission area gone to m3 before twenty two augusthas home's miw mission area gone to m3 before twenty two augustit has home's miw mission area going two m3 before twenty two august


Correct:

Digram:

Word-pair:

None:











display a chart of bering sea with the time window from twenty fourhoodred to eighteen hoodred hours zuludisplay a chart of bering sea with •• time window from twenty fourhoodred to eighteen hoodred hours zuludisplay·· chart of bering sea with •• time window from twenty fourhoodred to eighteen hoodred hours zuludisplay·· chart •• bering sea with •• time in that from twenty fourhad eighl it eighteen hundred hours beam

how many in west philippine sea have more than half their fuel lefthow many in west philippine sea have more than half their fuel lefthow many in west philippine sea have more than half their fuel lefthow any in west philippine sea have more than half the fuel left

how close is seawolf's last location to fifty two degrees north eight degrees easthow close is seawolf's last location to fifty·· degrees north eight degrees easthow close is seawolf's last location to fifty·· degrees north eight degrees easthow close •• seawolf's last location two fifty·· degrees north be decrease east

get the destinations and arrival hour at destination for all subsget the destinations and arrival hour at destination for all subsget •• destinations and arrival hour at destination for all subsget the destinations •• arrival hour bad destination for all subs

get all usn ships that are in coral seaget all usn ships that are in coral seaget all usn ships that are in coral seagive all usn ships that •• bering coral sea

tum off cep switchtum off cep switchtum off cep switchtum above cep switch

is the economic speed of apalachicola less than that of the brunswick•• whose economic speed of apalachicola less than that of the brunswick•• whose economic speed of apalachicola a list the lat of·· brunswickhis •• economic speed do of apalachicola list ...... the level brlUlswick

define area alerts for gulf of californiadefine area alerts for gulf of californiadefine area alerts for gulf of californiadefine area alerts for gulf of california

clear display windowclear display windowclear display windowclear display window

show me home's track in dim orange with reeves's in bright greenshow me home's track in dim orange with reeves in bright greenshow me home's track in dim orange with reeves in bright greenshow me home's track indian·· orange with reeves in bright green

get resource area ratings for enterpriseget resource area ratings for enterpriseget resource area ratings for enterprisedid resource area ratings for enterprise

184













is anybody in westpac ntds capableis anybody in westpac ntds capableis anybody in westpac ntds capableis anybody in I~ss pac ntds capable

list monday's and tuesday's casualty reports for frigates in bass straitlist monday's and luesday's casualty repons for frigates in bass straitlist monday's and tuesday's casualty reports for frigates in bass straitlist monday's and tuesday's casualty reports far frigates can bass strail

how many cruisen thai are sqq-23 capable are there at bombayhow many cruisen thai ar~n't sqq-23 capable are there al bombayhow many cruisen thaltM sqq-23 capable are there at bombayhow many cruisen •• stU~1I sqq-23 capable by there hav~ bombay

which of the cruisen that are in korean bay have Ips-48which of the cruisen thai are in korean bay have sps-48which of the cruisen that are in korean bay have sps-48which of a cruisers that are in korean bay have sps-48

does the campbell have four open cat-3 problemswas the campbell have four open cat-3 problemsdoes the campbell have four open cat-3 problemsdoes the campbell have four open cat-3 problems

has jason been downgraded yethas jason been downgraded yethas jason been downgraded yethas jason in downgraded yet

is there a gulf of thailand ship rated m5 on miwis there a gulf of thailand ship rated m5 on miwis there a gulf of thailand ship rated m5 on miwis their •• gulf of thailand ship rated tlUning thailand Ion miw

what's hawkbill' s fleet identificationwhat's hawkbill's fleet identificationwhat's hawkbill' s fleet identificationwhat's hawkbill's fleet identification

show on data screen ranger's track since october thirteenthshow Ion data screen ranger's track since october thineenthshow Ion data screen ranger's track since october thineenthshow on~ dat~d screen queenflSh track since october ifthirtccnth

what is the aaw rating of the virginiawhat is the aaw rating of the virginiawhat is fiji in aaw rating of the virginiawhat is ~ta go b~am rating both the virginia

list all the alertslist all •• alenslist all •• alertslist all~ alens

EXAMPLES OF SPHINX RECOGNmON 185




Correct:Bigram:Word.palr:None:








get names and latitudes for sps-40 carriers in arabian sea twenty seven novemberget names and latitudes for sps-40 carriers in arabian sea twenty seven novemberget names and latitudes for sps-40 carriers in arabian sea twenty seven novemberget names and latitudes for sps-40 carriers in arabian sea twenty seven november

give vessels in indian ocean and their destinationsgive vessels in indian ocean and their destinationsgive vessels in indian ocean and their destinationsgive vessel's in ollly ocean end their destinations

where's ponac1c nowwhere'. ponac1c nowwhere'. ponac1c nowwhere'. ponac1c now

toggle the mit ofmeasure parametertoggle •• unit ofmeasure parametertoggle •• unit ofmeasure parametertoggle •• unit •• measure parameter

are any ships in bismark sea below ninety percent of their fuel capacityare any ships in bismark sea below IIW percent of their fuel capacityare any ships in bismark sea below IIW percent of their fuel capacitywhere me ships in bismark sea below mind hers letter their fuel capacity

increase letter size to the max value and redrawincrease leller size to the max value and redrawincrease leller size to •• max value and redrawincrease letter size •••• max value •• redraw off

what's the cleveland's current readinesswhat's the cleveland's current readinesswhat's the cleveland's current readinesswhat's the cleveland slarl readiness

get latitudes and longitudes and names of ships in the arabian seagive latitudes and longitudes and names of ships in •• arabian seagive latitudes and longitudes ofnames of ships in •• arabian seagive latitudes illiangitudes •• names an ship. lIear •• arabian sea

fmd mission. edited todayfmd missions edited todayfmd missions edited todayfmd missions edited today

get the various capabilities for gas turbine ships in the gulf of tonkinget the various capabilities for gas turbine ships in •• gulf of tankinget the various capabilities for gas turbine ships in •• gulf of tankingive do various capabilities for gas turbine ships all •• gulf •• tonkin

was lockwood's location an sunday in sea of japanwas lockwood's location on sunday in sea of japanwas lockwood's location an sunday in sea ill japanwas lockwood's location on sunday in sea" japan

186




Correct:Billram:Word-pair:None:





Correct:B1llram:Word-pair:None:




get c2 ships that are in diego-garciaget c2 ships that are in diego-garciaget c2 ships that are in diego-garciaget c2 ships •••• c~lIurillg diego-garcia

review alerts within the last ten hours for the ships that are in gulf of alaskareview alerts within the last ten hours for the ships that are in gulf of alaskareview alerts within the last ten hours for the ships that are in gulf of alaska~t~or 110 arctic than the last ten hours from the ships that are in gulf of alaska

what speed is eisenhower goingwhat speed is eisenhower goingwhat speed is eisenhower goingwhat speed is eisenhower going

define an alert for the formosa straitdefine an alert for the formosa straitdefine an alert for the formosa straitd~!,Ulillg all ar~II'1 for the formosa strait

show the same chart with time started at nineteen hundred zulushow the same chart with time started at nineteen hundred zulushow the same chart with time started at nineteen hundred zulushow the same chart w~r~II'1 time started hav~ nineteen ~nd i/$ ~lIgland

ia eiaenhower's beam amaller than mississippi 'ais 1M eisenhower's beam smaller than mississippi'sis eisenhower's beam smaller than mississippi'sis eisenhower's beam small than mississippi's

show locations for subs in eastpac that went to cl on eleven januaryshow locations for subs in eastpac that went to cl on eleven januaryshow locations for subs in eastpac that went to cl on eleven januaryshow locations for subs and eastpac that OM to cl on eleven january

give current equipment readiness of the hectorgive current equipment readiness of the hectorgive current equipment readiness •• lhal hectorgive current equipment ralillg does of the hector

show the new definitions involving vancouvershow the new definitions involving vancouvershow the new definitions involving vancouvershow Ihall~ definitions involving vancouver

show latitude and longitude of seawolfshow latitude and longitude of seawolfshow latitude a/longitude of seawolfshow latitude •• longitude of seawolf

show carriers that are in china sea •• and m3 on miwshow carriers that are in china sea hav~ all m3 on miwshow carriers that are in china sea and m3 on miwshow carriers that orallg~ •• all sea 1~1I m3 on miw

References[Adams 86]

[BaW 78a]

[BaW 78b]

[BaW 80a]

[BaW 80b]

[BaW 81a]

[BaW 81b]

Adams, D. A., Bisiani, RThe Carnegie Mellon University Distributed SpeechRecognition System.

In Speech Technology, pages 14-23. March/April, 1986.

Bahl, L. R, Baker, J. K., Cohen, P. S., Jelinek, F.. Lewis,B. L., Mercer, R L.Recognition of a Continuously Read Natural Corpus.In IEEE International Conference on Acoustics, Speech,

and Signal Processing. April, 1978.

Bahl, L.R, Baker, J.K., Cohen, P.S., Cole, A.G., Jelinek,E, Lewis, B.L., Mercer, RL.Automatic Recognition of Continuously Spoken Sentencesfrom a Finite State Grammar.

In IEEE International Conference on Acoustics, Speech,and Signal Processing, pages 418-421. April, 1978.

Bahl, L. R, Bakis, R., Cohen, P. S., Cole, A. G., Jelinek,F., Lewis, B. L., Mercer, R L.Further Results on the Recognition of a ContinuouslyRead Natural Corpus.

In IEEE International Conference on Acoustics, Speech,and Signal Processing. April, 1980.

Bahl, L.R, Bakis, R, Jelinek, F., Mercer, RL.Language-Model/Acoustic Channel Balance Mechanism.IBM Technical Disclosure Bulletin 23(7B):3464-3465,December, 1980.

Bahl, L. R, Bakis, R, Cohen, P. S., Cole, A. G., Jelinek,F., Lewis, B. L., Mercer, R. L.Speech Recognition of a Natural Text Read as IsolatedWords.

In IEEE International Conference on Acoustics, Speech,and Signal Processing. April,1981.

Bahl, L. R, Bakis, R, Cohen, P. S., Cole, A. G., Jelinek,E, Lewis, B. L., Mercer, R L.Continuous Parameter Acoustic Processing forRecognition of a Natural Speech Corpus.


188

[Bahl83a]

[Bahl83b]

[Bahl88a]

[Bahl88b]

[Baker 75a]

[Baker 75b]

[Bakis 76]

[Baum 72]


BaW, L. R, Jelinek, F., Mercer, R.A Maximum Likelihood Approach to Continuous SpeechRecognition.

IEEE Transactions on Pattern Analysis and MachineIntelligence PAMI-5(2):179-190, March, 1983.

BaW, L. R, Cole, A. G., Jelinek, F., Mercer, R L., Nadas,A., Nahamoo, D., Picheny, M. A.Recognition of Isolated-Word Sentences from a 5000Word Vocabulary Office Correspondence Task.


BaW. L.R, Brown, P.F., De Souza, P.V., Mercer, RL.Obtaining Candidate Words by Polling in a LargeVocabulary Speech Recognition System.


Bahl. L.R, Brown, P.F., De Souza, P.V., Mercer, RL.Acoustic Markov Models Used in the Tangora SpeechRecognition System.


Baker, 1. K.The DRAGON System -- An Overview.IEEE Transactions on Acoustics, Speech, and Signal

Processing ASSP-23(l):24-29, February, 1975.

Baker, J. K.Stochastic Modeling as a Means ofAutomatic Speech

Recognition.PhD thesis, Computer Science Department, CarnegieMellon University, April, 1975.

Bakis, RContinuous Speech Recognition via Centisecond AcousticStates.

In 9Jst Meeting of the Acoustical Society ofAmerica.April, 1976.

Baum, L. E.An Inequality and Associated Maximization Technique inStatistical Estimation of Probabilistic Functions ofMarkov Processes.

Inequalities 3:1-8, 1972.

REFERENCES

[Brown 83]

[Brown 87]

[Chigier 88]

[Chow 86]

[Chow 87]

[Cohen 74]

[Cole 80]

189

Brown, P. F., Lee, C-H., Spohr, 1. C.Bayesian Adaptation in Speech Recognition.In IEEE International Conference on Acoustics, Speech,

and Signal Processing, pages 761-764. April,1983.

Brown, P.The Acoustic-Modeling Problem in Automatic Speech

Recognition.PhD thesis, Computer Science Department, CarnegieMellon University, May, 1987.

Chigier, B. Brennan, R.Broad Class Network Generation Using a Combination ofRules and Statistics for Speaker IndependentContinuous Speech.


Chow, Y. L., Schwartz, R., Roucos, S., Kimball, 0., Price,P., Kubala, F., Dunham, M., Krasner, M., Makhoul, J.The Role of Word-Dependent Coarticulatory Effects in aPhoneme-Based Speech Recognition System.


Chow, Y.L., Dunham, M.O., Kimball, O.A., Krasner,M.A., Kubala, G.F., Makhoul, J., Roucos, S., Schwartz,RM.BYBLOS: The BBN Continuous Speech RecognitionSystem.


Cohen, P. S., Mercer, R. L.The Phonological Component of an Automatic SpeechRecognition System.

In Proceedings of the IEEE Symposium on SpeechRecognition, Pittsburgh, PA, pages 177-187. 1974.

Cole, R A., Rudnicky, A. I., Zue, V. W., Reddy, D. RSpeech as Patterns on Paper.In R. A. Cole (editor), Perception and Production of

Fluent Speech. Lawrence Erlbaum Associates,Hillsdale, N.J., 1980.

190

[Cole 83]

[Cole 86a]

[Cole 86b]

[Cravero 84]

[Cravero 86]

[D'Orta 87]

[Das 83]

[Davis 80]


Cole, R. A, Stem, R. M., Phillips, M. S., Brill, S. M.,Specker, P., Pilant, A P.Feature-Based Speaker Independent Recognition ofEnglish Letters.

In IEEE International Conference on Acoustics, Speech,and Signal Processing. October, 1983.

Cole, R. APhonetic Classification in New Generation SpeechRecognition Systems.

In Speech Tech. 86, pages 43-46. 1986.

Cole, R. A., Phillips, M., Brennan, B., Chigier, B.The C-MU Phonetic Classification System.In IEEE International Conference on Acoustics, Speech,


Cravero, M., Fissore, L., Pieraccini, R., Scagliola, C.Syntax Driven Recognition of Connected Words byMarkov Models.

In IEEE International Conference on Acoustics, Speech.and Signal Processing. April, 1984.

Cravero, M, Pieraccini, R, Raineri, F.Definition and Evaluation of Phonetic Units for SpeechRecognition by Hidden Markov Models.


D'Orta, P, Ferretti, M., Scarci, S.Phoneme Classification for Real Time SpeechRecognition of Italian.

In IEEE International Conference on Acoustics, Speech,and Signal Processing, pages 81-84. April,1987.

Das, S.K.Some Dimensionality Reduction Studies in ContinuousSpeech Recognition.


Davis, S.B, P. Mennelstein.Comparison of Parametric Representations ofMonosyllabic Word Recognition in ContinuouslySpoken Sentences,

IEEE Transactions on Acoustics, Speech, and SignalProcessing ASSP-28(4):357-366, August, 1980.

REFERENCES 191

[Deng 88] Deng, L, Lennig, M., Gupta, V.N., Mennelstein, P.Modeling Acoustic-Phonetic Detail in an HMM-basedLarge Vocabulary Speech Recognizer.


[Derouault 87] Derouault, A.-M.Context-Dependent Phonetic Markov Models for LargeVocabulary Speech Recognition.

In IEEE International Conference on Acoustics, Speech.and Signal Processing, pages 360-3. April, 1987.

[Duda 73] Duda, R. 0., Hart, P. E.Pattern Classification and Scene Analysis.John Wiley & Sons, New York, N.Y., 1973.

[Feng 88] Feng, M.W., Kubala, F., Schwartz, R.Improved Speaker Adaptation Using Text DependentMappings.


[Fisher 87] Fisher, W.M., Zue, V., Bernstein, J., Pallett, D.An Acoustic-Phonetic Data Base.In 113th Meeting of the Acoustical Society ofAmerica.May, 1987.

[Furui 86] Furui, S.Speaker-Independent Isolated Word Recognition UsingDynamic Features of Speech Spectrum.

IEEE Transactions on Acoustics, Speech. and SignalProcessing ASSP-34(1):52-59, February, 1986.

[Gray 84] Gray, R.M.Vector Quantization.IEEE ASSP Magazine 1(2):4-29, April, 1984.

[Gupta 87] Gupta, V.N., Lennig, M., Mennelstein, P.Integration of Acoustic Infonnation in a Large VocabularyWord Recognizer.


[Hamming 86] Hamming, R.W.Coding and Information Theory.Prentice-Hall, Englewood Cliffs NJ, 1986.

192

[Haton 84]

[Hon 88]

[Hunt 80]

[Hwang 88]

[IBM 85]

[ltakura 75]

[Jelinek 76]

[Jelinek 80]


Haton, J.-P.Knowledge-based and Expert Systems in AutomaticSpeech Recognition.

In DeMori, R. (editor), New Systems and ArchitecturesforAutomatic Speech Recognition and Synthesis.Dordrecht, Reidel. Netherlands, 1984.

Hon,H.W.Personal Communication.unpublished.1988

Hunt. M. 1., Lennig, M., Mermelstein, P.Experiments in Syllable-Based Recognition of ContinuousSpeech.


Hwang,M.Y.Personal Communication.unpublished.1988

IBM speech recognition group.A Real-Time, Isolated-Word, Speech Recognition Systemfor Dictation Transcription.

In IEEE International Conference on Acoustics, Speech.and Signal Processing. March,1985.

Itakura, F.Minimum Prediction Residual Principle Applied toSpeech Recognition.

IEEE Transactions on Acoustics, Speech, and SignalProcessing ASSP-23(1):67-72, February, 1975.

Jelinek, F.Continuous Speech Recognition by Statistical Methods.Proceedings a/the IEEE 64(4):532-556, April, 1976.

Jelinek, F., Mercer, R.L.Interpolated Estimation ofMarkov Source Parametersfrom Sparse Data.

In E.S. Gelsema and L.N. Kanal (editor), PatternRecognition in Practice, pages 381-397. NorthHolland Publishing Company, Amsterdam, theNetherlands, 1980.

REFERENCES

[Jelinek 85]

[Jelinek 87]

[Juang 8Sa]

[Juang 8Sb]

[Juang 8Sc]

[Kimball 86]

[Klatt 72]

193

Jelinek, F.Self-Organized Language Modeling for SpeechRecognition.

Unpublished.1985

Jelinek, F.Personal Communication.unpublished.1987

Juang, RH., Rabiner, L.R., Levinson, S.E., Sondhi, M.M.Recent Developments in the Application of HiddenMarkov Models to Speaker-Independent IsolatedWord Recognition.

In IEEE International Conference on Acoustics, Speech,and Signal Processing, pages 9-12. April,198S.

Juang, B. H., Rabiner, L. R.Mixture Autoregressive Hidden Markov Models forSpeech Signals.

IEEE Transactions on Acoustics, Speech, and SignalProcessing ASSP-33(6):1404-13, December, 1985.

Juang, B.H., Rabiner, L.R.A Probabilistic Distance Measure for Hidden MarkovModels.

The Bell System Technical Journal 64(2):391-408,February, 1985.

Kimball, 0., Price, P., Roucos, S., Schwartz, R., Kubala,F., Chow, Y.-L., Haas, A., Krasner, M., Makhoul, J.Recognition Performance and Grammatical Constraints.In Lee S. Baumann (editor), Proceedings o/the DARPA

Speech Recognition Workshop, pages 53-59.February, 1986.

Klatt, D.H., Stevens, K.N.Sentence Recognition from Visual Examination ofSpectrograms and Machine-Aided Lexical Searching.

In Proceedings 1972 Conference on SpeechCommunication and Processing, pages 315-318.IEEE and AFCRL, 1972.

194

(Klatt 86]

[Kubala 88]

(Lamel86]

[Lea 80]

(Lee 85a]

[Lee 85b]

[Lee 86]

AUTOMATIC SPEECH RECOGNITIOl\

Klatt, D.Problem of Variability in Speech Recognition and inModels of Speech Perception.

In J.S. Perkell and D.M. Klatt (editor), Variability andInvariance in Speech Processes, pages 300-320.Lawrence Erlbaum Assoc, Hillsdale, N.J., 1986.

Kubala, G.F., Chow, Y., Derr, A., Feng, M., Kimball, 0.,Makhoul, 1., Price, P., Rohlicek, 1., Roucos, S., Schwartz,R., Vandegrift, J.Continuous Speech Recognition Results of the BYBLOSSystem on the DARPA lOOO-Word ResourceManagement Database.


Lamel, L.F., Kassel, R.H., Seneff, S.Speech Database Development: Design and Analysis ofthe Acoustic-Phonetic Corpus.

In Baumann, L.S. (editor), Proceedings of the DARPASpeech Recognition Workshop, pages 100-109.February, 1986.

Lea, W.A.Trends in Speech Recognition.Prentice-Hall, Englewood aiffs, NJ, 1980.

Lee, K.F.Incremental Network Generation in Template-Based Word

Recognition.Technical Report CMU-CS-85-181, Computer ScienceDepartment, Carnegie Mellon University, December,1985.

Lee, K.F.Network Representation of Templates in WordRecognition.

In The J09th Meeting of the Acoustical Society ofAmerica. April, 1985.

Lee, K.F.Incremental Network Generation in Word Recognition.In IEEE International Conference on Acoustics, Speech,


REFERENCES 195

[Lee 87] Lee, K.F.Towards Speaker-Independent Continuous SpeechRecognition.

In 1987 NATO ASI on Speech Recognition and DialogUnderstanding. 1987.

[Lee 88a] Lee, K.F., Hon, H.W.Large-Vocabulary Speaker-Independent ContinuousSpeech Recognition.

In IEEE International Conference on Acoustics. Speech.and Signal Processing. April, 1988.

[Lee 88b] Lee, K.F., Hon, H.W.Speaker-Independent Phoneme Recognition Using Hidden

Markov Models.Technical Report CMU-CS-88-121, Computer ScienceDepartment, Carnegie Mellon University, Pittsburgh,PA, April, 1988.

[Lee 88c] Lee, K.F.On Large-Vocabulary Speaker-Independent ContinuousSpeech Recognition.

Journal of the Eurpopean Association ofSignalProcessing, September, 1988.

[LeeCH 88a] Lee, C.H., Rabiner, L.R.A Network-based Frame-synchronous Level BuildingAlgorithm for Connected Word Recognition.


[LeeCH 88b] Lee, C.H., Soong, F.K., Juang, B.H.A Segment Model Based Approach to SpeechRecognition.


[Lesser 75] Lesser, V. R., Fennell, R. D., Erman, L. D., Reddy, R. D.The Hearsay II Speech Understanding System.IEEE Transactions on Acoustics, Speech, and Signal

Processing ASSP-23(l):11-24, February, 1975.

[Levinson 77] Levinson, S. E., Rosenberg, A. E., Flanagan, 1. L.Evaluation of a Word Recognition System Using SyntaxAnalysis.


196 AUTOMATIC SPEECH RECOGNTIlON

[Levinson 79] Levinson, S. E., Rabiner, L. R., Rosenberg, A. E., Wilpon,J.G.Interactive Clustering Techniques for Selecting SpeakerIndependent Reference Templates for Isolated WordRecognition.

IEEE Transactions on Acoustics, Speech, and SignalProcessing ASSP-27(2):134-41, April, 1979.

[Levinson 83] Levinson, S. E., Rabiner, L. R., Sondhi, M. M.An Introduction to the Application of the Theory ofProbabilistic Functions on a Markov Process toAutomatic Speech Recognition.

The Bell System Technical Journal 62(4), April, 1983.

[Linde 80] Linde, Y., Buzo, A., Gray, R.M.An Algorithm for Vector Quantizer Design.IEEE Transactions on Communication COM-28(l):84-95,January, 1980.

[Lippmann 87] Lippmann, R.P., Martin, E.A., Paul, D.P.Multi-Style Training for Robust Isolated-Word SpeechRecognition.


[Lowerre 76] Lowerre, B. T.The HARPY Speech Recognition System.PhD thesis, Computer Science Department, CarnegieMellon University, April, 1976.

[Lowerre 77] Lowerre, B. T.Dynamic Speaker Adaptation in the Harpy SpeechRecognition System.


[Lowerre 80] Lowerre, B.T., Reddy, D.R.The Harpy Speech Understanding System.Trends in Speech Recognition.Prentice-Hall, Englewood Cliffs, NJ, 1980.

[Lucassen 83] Lucassen, J.M.Discovering Phonemic Basejorms: an Information

Theoretic Approach.Research Repon RC 9833, IBM, February, 1983.

REFERENCES

[Makhoul 85]

[Markel 76]

[Meilijson 87]

[Merialdo 87]

[Murveit 88]

[Myers 81]

[Nadas 81]

[Nag 86]

197

Makhoul, J., Roucos, S., Gish, H.Vector Quantization in Speech Coding.Proceedings of the IEEE 73(11):1551-1588, November,1985.

Markel, 1. D., Gray, A H.Linear Prediction ofSpeech.Springer-Verlag, Berlin, 1976.

Meilijson, I.A Fast Improvement to the EM Algorithm on its OwnTenns.

Forthcoming in the Journal of the Royal StatisticalSociety.

1987

Merialdo, B.Speech Recognition With Very Large Size Dictionary.In IEEE International Conference on Acoustics, Speech,

and Signal Processing, pages 364-7. April, 1987.

Murveit, H., Weintraub, M.Speaker-Independent Connected-Speech RecognitionUsing Hidden Markov Models.


Myers, C.S., Rabiner, L.R.Connected Digit Recognition Using a Level BuildingDTW Algorithm.

ASSP ASSP-29(3):351-363, June, 1981.

Nadas, A, Mercer, R. L., Baht, L. R., Bakis, R., Cohen,P. S., Cole, A G., Jelinek, F., Lewis, B. L.Continuous Speech Recognition with AutomaticallySelected Acoustic Prototypes Obtained by EitherBootstrapping or Clustering.


Nag, R., Austin, S.c., Fallside, F.Using Hidden Markov Models to Define Linguistic Units.In IEEE International Conference on Acoustics, Speech,


198 AUTOMATIC SPEECH RECOGNITION

[Ney 87) Ney, H., Mergel, D., Noll, A, Paeseler, AA Data-Driven Organization of the DynamicProgramming Beam Search for Continuous SpeechRecognition.


[Ney 88) Ney, H., Noll, APhoneme Modelling Using Continuous Mixture Densities.In IEEE International Conference on Acoustics, Speech,


[Nilsson 80) Nilsson, N.J.Principles ofArtificial Intelligence.Tioga Publishing Co., Palo Alto, CA, 1980.

[Nishimura 87] Nishimura, M., Toshioka, K.HMM-Based Speech Recognition Using MultiDimensional Mutli-Labeling.


[Noll 87) Noll, A. Ney, H.Training of Phoneme Models in a Sentence RecognitionSystem.


[Oppenheim 72) Oppenheim, A V., Johnson, D. H.Discrete Representation of Signals.The Proceedings of the IEEE 60(6):681-691, June, 1972.

[Paul 86)

[Paul 88)

[Picheny 88)

Paul, D. B., Lippmann, R. P., Chen, Y., Weinstein, C.Robust HMM-Based Techniques for Recognition ofSpeech Produced under Stress and in Noise.

In Speech Tech. April, 1986.

Paul, D.B., Martin, E.A.Speaker Stress-Resistant Continuous Speech Recognition.In IEEE International Conference on Acoustics, Speech,


Picheny, M.Personal Communication.unpublished.1988

REFERENCES

[Polifroni 88]

[Price 88]

[Rabiner 79]

[Rabiner 81 ]

[Rabiner W~l

[Rabiner 84]

[Rabiner 85]

[Rabiner 86]

199

Polifroni, 1.Personal Communication.unpublished.1988

Price, P.J., Fisher, W., Bernstein, J., Pallett, D.A Database for Continuous Speech Recognition in a 1000Word Domain.


Rabiner, L. R, Levinson, S. E., Rosenberg, A. E., Wilpon,J.G.Speaker-Independent Recognition of Isolated WordsUsing Clustering Techniques.

IEEE Transactions on Acoustics, Speech, and SignalProcessing ASSP-27(4):336-349, August, 1979.

Rabiner, L. R, Wilpon, J. G.Isolated Word Recognition Using a Two-Pass PatternRecognition Approach.

In IEEE International Conference on Acoustics, Speech,and Signal Processing, pages 724-7. March,1981.

Rabiner, L. R, Levinson S. E., Sondhi, M. M.On the Application of Vector Quantization and HiddenMarkov Models to Speaker-Independent, IsolatedWord Recognition.

The Bell System Technical Journal 62(4):1075-1105,April,1983.

Rabiner, L. R, Pan, K. c., Soong, F. K.On the Performance of Isolated Word Speech RecognizersUsing Vector Quantization and Temporal EnergyContours.

AT&T Bell Laboratories Technical Journal63(7):1245-1260, September, 1984.

Rabiner, L. R, Juang, B. H., Levinson, S. E., Sondhi,M.M.Recognition of Isolated Digits Using Hidden MarkovModels With Continuous Mixture Densities.

AT&T Technical Journal 64(6):1211-33, July-August,1985.

Rabiner, L.R., Juang, B.H.An Introduction to Hidden Markov Models.IEEE ASSP Magazine 3(1):4-16, January, 1986.


[Rabiner 88a] Rabiner, L.R., Wilpon, lO., Soong, F.K.High Perfonnance Connected Digit Recognition UsingHidden Markov Models.


[Rabiner 88b] Rabiner, L.R.A Tutorial on Hidden Markov Models and SelectedApplications in Speech Recognition.

IEEE Proceedings, 1988.

[Reddy 77] Reddy, D. R.Speech Understanding Systems: Summary ofResults ofthe Five-Year Research Effort at Carnegie MellonUniversity.

Internal Document.August, 1977

[Reddy 83] Reddy, D.R., ZUe, V.Recognizing Continuous Speech Remains an musiveGoal.

IEEE Spectrum :84-87, November, 1983.

[Richter 86] Richter, A.G.Modeling of Continuous Speech Observations.In Advances in Speech Processing Conference, IBM

Europe Institute. July, 1986.

[Rosenberg 83] Rosenberg, A. E.• Rabiner, L. R.. Wilpon, J., Kahn. D.Demisyllable-Based Isolated Word Recognition System.IEEE Transactions on Acoustics. Speech, and Signal

Processing ASSP-31(3):713-726. June, 1983.

[Roucos 87] Roucos, S., Dunham, M.O.A Stochastic Segment Model for Phoneme-BasedContinuous Speech Recognition.

In IEEE International Conference on Acoustics, Speech.and Signal Processing, pages 73-76. April, 1987.

[Roucos 88] Roucos, S., Ostendorf, M.• Gish, H., Derr. A.Stochastic Segment Modeling Using the EstimateMaximize Algorithm.


REFERENCES 201

[Rudnicky 87] Rudnicky, A., Baumeister, L., DeGraaf, K., Letunann, E.The Lexical Access Component of the CMU ContinuousSpeech Recognition System.


[Ruske 82] Ruske, G.Auditory Perception and Its Application to ComputerAnalysis of Speech.

In C. Y. Suen and R De Mori (editor), Auditory Signals.Volume II: Computer Analysis and Perception. CRCPress, Boca Raton, FL, 1982.

[Russell 85] Russell, MJ., Moore, RK.Explicit Modeling of State Occupancy in Hidden MarkovModels for Automatic Speech Recognition.

In IEEE International Conference on Acoustics. Speech,and Signal Processing, pages 5-8. April, 1985.

[Schwartz 80] Schwartz, R, Klovstad, J., Makhoul, J., Sorensen, J.A Preliminary Design of a Phonetic Vocoder Based on aDiphone Model.


[Schwartz 84] Schwartz, R M., Chow, Y. L., Roucos, S., Krasner, M.,Makhoul, J.Improved Hidden Markov Modeling of Phonemes forContinuous Speech Recognition.


[Schwartz 85] Schwartz,R, Chow, Y., Kimball, 0., Roucos, S., Krasner,M., Makhoul, 1.Context-Dependent Modeling for Acoustic-PhoneticRecognition of Continuous Speech.


[Schwartz 87] Schwartz, R, Chow, Y., Kubala, F.Rapid Speaker Adaptation Using a Probabilistic SpectralMapping.



[Shikano 85] Shikano. K.Evaluation ofLPC Spectral Matching Measuresfor

Phonetic Unit Recognition.Technical Report. Computer Science Department.Carnegie Mellon University. May. 1985.

[Shikano 86a] Shikano, K., Lee. K, Reddy. D. R.Speaker Adaptation through Vector Quantization.In IEEE International Conference on Acoustics, Speech.


[Shikano 86b] Shikano. K., Lee, K. Reddy. D. R.Speaker Adaptation through Vector Quantization.Technical Report CMU-CS-86-102. Computer ScienceDepartment, Carnegie Mellon University, December.1986.

[Shikano 86c] Shikano. K.Evaluation ofLPC Spectral Matching Measures for

Phonetic Unit Recognition.Technical Report. Computer Science Department.Carnegie Mellon University, 1986.

[Shikano 86d] Shikano, K.Text-Independent Speaker Recognition Experiments using

Codebook:s in Vector Quantization.Technical Report. Computer Science Department.Carnegie Mellon University. 1986.

[Shipman 82] Shipman. D.W.• Zue. V.W.Properties of Large Lexicons: Implications for AdvancedIsolated Word Recognition Systems.

In IEEE International Conference on Acoustics, Speech,and Signal Processing. pages 546-549. April. 1982.

[Stem 83] Stem. R. M.• Lasry. M. 1.Dynamic Speaker Adaptation for Isolated LetterRecognition Using MAP Estimation.

In IEEE International Conference on Acoustics, Speech,and Signal Processing. April. 1983.

[Sugawara 85] Sugawara, K.• Nishimura, M.• Toshioka, K.. Okochi. M.,Kaneko. T.Isolated Word Recognition Using Hidden Markov Models.In IEEE International Conference on Acoustics. Speech,

and Signal Processing. April. 1985.

REFERENCES 203

[Thompson 87] Thompson, H.S., Laver, J.D.The Alvey Speech Demonstrator - Architecture,Methodology, and Progress to Date.

In Proceedings ofSpeech Tech. 1987.

[TI 87] TI Speech Recognition Group.TI Speech Recognition Technology Development.In DARPA Speech Recognition Workshop. October, 1987.

[Tohkura 86) Tohkura, Y.A Weighted Cepstral Distance Measure for SpeechRecognition.


[Viterbi 67] Viterbi, A. 1.Error Bounds for Convolutional Codes and anAsymptotically Optimum Decoding Algorithm.

IEEE Transactions on Information TheoryIT-13(2):260-269, April, 1967.

[Waibel 86) Waibel, A. H.Prosody and Speech Recognition.PhD thesis, Computer Science Department, CarnegieMellon University, October, 1986.

[Wilpon 82) Wilpon, 1. G., Rabiner, L. R., Bergh, A.Speaker-Independent Isolated Word Recognition Using a129-Word Airline Vocabulary.

The Journal of the Acoustical Society ofAmerica72(2):390-396, August, 1982.

[Zue 85) Zue, V. W.The Use of Speech Knowledge in Automatic SpeechRecognition.

Proceedings of the IEEE 73(11):1602-1615, November,1985.

[Zwicker 61) Zwicker, E.Subdivision of the Audible Frequency Range into CriticalBands (Frequenzgruppen).

Journal of the Acoustical Society ofAmerica 33:248,February, 1961.

IndexA* search 41Adaptation 10,14, 115, 140Admissibility 42Alpha terminal 24, 118

Bark scaleSee also Mel scale

Baseforms 75Baum Welch algorithmSee also Forward-backward algorithm

Bayes rule 22Beam search 2Bigram grammar 46Bilinear transform 64results 84

Blackboard 2, 63Bottom-up 36,63

Clustering 4, 14Agglomerate clustering 104,117Entropy clustering 104Speaker clustering 116, I I7

Co-aniculation 6, 13, 93Codebook 52See also Vector quantization

Cohons 143Composite distance metric 68Compound phones 78, 81Content words 6Context-dependent phone modeling 13Context-dependent phonesSee also Triphones

Continuous speech recognition 6,38, 146

Deleted interpolation 13,15,58adaptation 119, 122contextual models 97smoothing 31

Delta cepstrum 65See also Differenced coefficients

Demisyllables 8, 13, 93Differenced coefficients II, 66results 84

Diphones 13,94Discriminant analysis 67Discrimination 142Distance metrics 32, 52, 68Duration modeling 54, 72exponential distribution 72results 87semi-Markov models 73word duration modeling 73

Dynamic programmingSee also Dynamic time warp

Dynamic time warp 2,31,38,60,93,147

Entropy 104, 145Error analysis 133Error modeling 87Error rate 146, 147Estimate-maximize algorithm 11,13,26See also Forward-backward algorithm

Finite state grammars 8,36,46, 170Fixed-width parameters 64Formant slope 65Forward algorithm 20, 37Forward-backward algorithm 11,23,57,76,98convergence proof 24for continuous speech recogntion 38for isolated word recogntion 36

Function words 6, 100Function-word-dependent phone modeling 13, 100results 108

Gaussian autoregressive density 32Gaussian mixture density 32Generalized triphones 13, 103results 110

Hamming window 51Hidden Markov models 10, 17continuous density HMM 32, 141decoding 22discrete density HMM 32,141evaluation 20initialization 27, 49leaming 23model topology 82problems 141similarity 104

Homonyms 60Human speech knowledge 138

Insertion/deletion modeling 77explicit modeling 78implicit modeling 78results 86

Integrated search 36, 63Interpolated re-estimation 14,118Interpolation 13,95,96See also Deleted interpolation

Isolated word recognition 6,36,146

K-nearest neighbor window 31Knowledge engineering 12,63Knowledge integration 12,67segment-level integration 73

Language model match factor 59Language models 22,59, 145Laplacian mixture density 32

206

Large vocabulary recognition 8,91Learning 10,14,115,140See also Adaptation, Hiden Markov models

Left-context dependent phones 95, 107Level buildingSee also Viterbi algorithm

Logarithmic compression 28LPC analysis 51LPC cepstral coefficients 51

Markovassumption 19, 142Maximum likelihood estimation 23,33, 142Maximum mutual information estimation 143Mel scale 64Microphone 47Multi-phone units 93Multiple codebooks 69Multiple independent observations 30Multiple pronunciations 79

Natural language grammars 9Neural networks 38Non-phonemic affricates 81Null transitions 27

Output probabilitydefinition 17re-estimation 24

Output-independence assumption 19,142

Panen window 31Pattem recognition 38Pdf 17Percent correct 60, 147Perplexity 2,8,46, 145test-set perplexity 146

Phone models 8,34,54,69,87,92Phone transition modeling 94Phoneme recognition 49, 87Phonemes 13Phones 48,55,83Phonetic models 76Phonological rules 76,81Power 11,66results 84

Pre-emphasis 51Principal component analysis 67Pronunciation dictionary 55, 83Prosody 11, 66

Regression coefficients 65Resource management task 15,45Richter mixture density 32Right-context dependent phones 95,107

Scaling 28Segment-based parametersSee also Variable-width features


Segmental K-means algorithm 28Signal processing 51Silence modeling 57Simiar speakers 121Small vocabulary recognition 91Smoothing 30, 58co-occurrence method 31, 120distance method 30floor method 30

Speaker adaptation 5,143Speaker cluster identification 116, 118Speaker cluster selection 14, 117Speaker clustering adaptation 116results 124, 125

Speaker-dependent recognition 5, 61, 71, 131Speaker-independent recognition 3,71,127,132Spectrogram reading 3,12,100Speech knowledge 10,11,12Speech models 10,138Speech units 10, 12, 34, 139Stack and reduce 67Stack decoding 41Stochastic grammars 9, 46Stochastic segment model 142Stress 100Subword models 8,34, 143acoustic subword models 143

Syllables 13,93

Tag 46Test set perplexity 46Test speakers 47, 170Tied transitions 26, 119TIMIT Database 48,55,82,87TIRM database 47Top-down 36Trainability vs. specificity 78,97,106,119,137Training speakers 47,170Transition phone models 94Transition probabilitydefmition 17re-estimation 24

Trellis 21Trigram grammars 9,61Triphones 95, 108Typing vs. speaking 7

Variable-width parameters 72Vector quantization 12,14,32,52,69,141distortion 32,69,70, 124

Viterbi algorithm 22,39,76beam search 40level building 40time-synchronous Viterbi search 40

Viterbi search 141beam search 59

Word accuracy 60, 147

INDEX

Word bolUldary detection 6, 38Word models 8,34,91Word pair grammar 46Word-dependent co-articulatory effects 91Word-dependent phones 95

207

Appendix I Evaluating Speech Recognizers - Home - …978-1-4615-3650-5/1.pdf · Appendix I...

Documents

Transcript of Appendix I Evaluating Speech Recognizers - Home - …978-1-4615-3650-5/1.pdf · Appendix I...