Aapd  @ p0` `@0PP 00p0@p`@0 'HH $d HHHH̀̀̀ff@  'd Footnote TableFootnote**.\t.\t /:;,.!?%9bU, bT LORUnresolved Cross-RefsLOT TableTitleLOFFigureTOCHeading1Heading2Heading3Alt BambergerBiCDsCorradaDeshmukhDoD Doddington DownArrowFaxFinkeFiscus GanapathirajuGleesonGoelHCopyHMMsHTKsHViteHamakerHartwellHemphillHollimanISIPs KhudanpurKirchoff LeftArrowMarkovMeadeMispronuncations ModellingNockOrdowskiPallettPentiumPicone PrzybockiResegmentation RightArrowSaraclarSimrallSolaris SparcstationsTashaTclTkUpArrowUtt ValidatorWERsWaibelWheatleyWootersWorkflow ZavaliagkosZengbaseform bookmarking breakpointsbtwncasesensitivecdcepstralcfg check_boundscheck_dictionary check_lexicon check_silence coarticulatedcoarticulationconfigconfirm_word_files create_configcreate_exciselistcreate_htk_mlfcrossvalidation crosstalk delimiterdev disfluentdoc echo_cancelemaileval excise_signalftpgonnagotta hardlimitedhrshttp: implementableinhouseinternetkinda linguistcallyloginmetricsmfccminsmiscapitalizedmisidentificationmlfmrk mrk_to_transmsecnahnawnonmonosyllabic outofband perutterance phonebasedpostprocessedpp prepare_dataprev raw_to_nist readme_files realtime recognizer reestimationremapped resegment resegmentatinresegmentations resegmented resegmenting retranscribedrogosclite screenshotsecsecs segmenter singlemostsorta spectrogramsplacestateoftheartthemselftimeconsuming timestampstimelinetriphone twochanneltxtuh validators validatorswannawavwaveform waveformswebsite wordbyword workforcexRTyall   Default FontDefault FontEquationVariablesEquationVariablesSymboldqUp^Va      a^p `t"dl 6>g6l 6_Jn}J^6 N   9  % l66666 6 6 6 6 6 6 6 6 6 b.bbb#/b%b&bZb[b`bacScUdd0eqeres1euevex2nLeze{7ZZ777Y88888Z\ e3`i!_L+^eee4e 8!8!eei| 8& 8(#8)#j 8-$8.$808=8D%86&87&8E%8G8I'8J'8L8N(8O(8Q8S)8T)8V8X*8Y*eee58`8b,8c,_M+eee68l.8m.8o8q/8r/8w8y08z08|8~1818E2E2a`eY&l8e`<`(e7ea7a7`)e`Fi~ `*i k8i"G#G$i"k-G%k-`<k9k3n~Jk3k:k4a9k4a>a>a9a?a?a@a@a+l5a,a-l5l8l$l&:l':l3;l4;l8=l9=lUAlVAlgliBljBm3lClCm5Dm6Dm7;m9Em:Em<mFmFm=mGmGmHmHmImIm>nKnKnLnMnMnNnNn o5o7Oo8OoPoPo pQ pSQpTQpaRpbRp'pSpS>< k39933: isip_reference: [15] J.Hamaker, Y.Zeng, and J.Picone, Rules and Guidelines for Transcription andc 12692: isip_table_title: Table 3: Word error rates for each validators transcriptions before manual word alignments, along with theA k22270: isip_reference: [2] Linguistic Data Consortium, SWITCHBOARD: A Users Manual, available at the URLD $34627: isip_heading_0: 1. ABSTRACTG 141681: isip_heading_0: 2. HISTORICAL BACKGROUNDJ $42443: isip_heading_0: 3. SOFTWAREB o15648: isip_reference: [1] A Statistical Guide to SWITCHBOARD: Topic Statistics, http://WWW.ISIP.MsState.Edu/E }24344: isip_figure: Figure9. An example of word alignments before and after manual word alignment review is performed. TheN :40100: isip_heading_1: 5.2. Validator speed and accuracyV K36623: isip_heading_1: 2.1. Development of an Efficient Segmentation Tool[ 34041: isip_figure: Figure4.\ :20370: isip_heading_1: 3.3. Word alignment tool overview_ A39657: isip_heading_1: 2.3. Integrated Project Management Toolsb /28448: isip_heading_1: 4.1. Echo cancellationc '28118: isip_heading_0: 4. Preparationh *27734: isip_heading_1: 4.2. Segmentationk 639436: isip_heading_1: 4.3. Transcription correctionn 718775: isip_heading_1: 4.4. Automatic word alignmentsq 413826: isip_heading_1: 4.5. Manual word alignments ;41334: isip_heading_1: 5.3. Preliminary LVCSR experimentsw /37660: isip_heading_0: 5. SUMMARY OF PROGRESS 17312: isip_table_title: Table 1: Word error rates for each validators transcriptions before manual word alignments and originalR 20146: isip_table_title: Table 2: Percentage of each type of error present in the original LDC transcriptions for conversation 3909. ?10909: isip_heading_1: 5.3. Summary of Switchboard statistics 538234: isip_heading_0: 6. SUMMARY, PLANS AND ISSUES ,26038: isip_heading_0: 7. ACKNOWLEDGEMENTS &16679: isip_heading_0: 8. REFERENCESH x18253: isip_figure: Figure10. An example of the information contained on the front page of the SWB Transcription FAQ. k18669: isip_reference: [2] Charles T. Hemphill, John J. Godfrey, and George R. Doddington, "The ATIS Spoken k23133: isip_reference: [1] Charles T. Hemphill, John J. Godgrey, and George R. Doddington, "The ATIS Spoken ~26523: isip_figure: Figure4. In the above waveform, the caller speaks for 21 seconds without an acoustical pause of 0.5 seci #22267: isip_table_title: Table 4.P w33362: isip_figure: Figure1. A SWITCHBOARD resegmentation tool that allows for easy manipulation of segmentation and }39434: isip_figure: Figure2. A word alignment tool that allows for easy manipulation of word boundaries and for quick tran m15013: isip_reference: [1] J. Godfrey, E. Holliman and J. McDaniel, "Telephone Speech Corpus for Research and! o21545: isip_reference: [3] A. Gleeson, J. Hamaker, N. Deshmukh, A. Ganapathiraju, and J. Picone, A Statistical$ p10836: isip_reference: [1] N. Deshmukh, A. Ganapathiraju, A. Gleeson, J. Hamaker, and J. Picone, Resegmentation '38634: isip_heading_0: 9. ATTACHMENTSn m29208: isip_reference: [18] J.Hamaker, A. Ganapathiraju, and J.Picone, SWITCHBOARD Educational Resources, r35765: isip_figure: Figure7. Timeline for remainder of SWB resegmentation project: We are currently on schedule m41565: isip_reference: [3] J. Godfrey, E. Holliman and J. McDaniel, "Telephone Speech Corpus for Research and f26392: isip_reference: [4] SWITCHBOARD:AUsersManual, http://www.cis.upenn.edu/~ldc/readme_files/6 /28169: isip_heading_1: 2.1. Recording History8 318545: isip_heading_1: 2.2. Transcription History9 :18980: isip_heading_1: 2.3. Recognition and Segmentation* h15326: isip_reference: [4] J.Hamaker and J.Picone, The SWITCHBOARD Frequently Asked Questions(FAQ),- )34146: isip_heading_1: 4.7. The SWB FAQ7 q35139: isip_figure: Figure2. The distribution of the amount of data per speaker in SWB is shown. Subjects were: q35736: isip_reference: [7] M.Finke and A.Waibel, Flexible Transcription Alignment, in Proceedings of the IEEE= n25683: isip_reference: [8] D.Pallett, 1997 Hub-5 Workshop: The Evaluation, in Proceedings of the 1997 Hub-5O j29445: isip_reference: [9] B.Byrne, M.Finke, S.Khudanpur, J.McDonough, H.Nock, M.Riley, M.Saraclar,U i39528: isip_reference: [10] A.Ganapathiraju, J.Hamaker, and J.Picone, Syllable-Based Large Vocabularyv i33219: isip_reference: [11] A.Ganapathiraju, V.Goel, J.Picone, A.Corrada, G.Doddington, K.Kirchoff,} w17636: isip_figure: Figure3. A SWITCHBOARD resegmentation tool that allows for easy manipulation of segmentation and k33800: isip_reference: [12] N.Deshmukh, A.Ganapathiraju, R.Duncan, and J.Picone, An Efficient Tool For P13330: isip_figure: Figure6. Workflow diagram for SWB resegmentation project.' q19576: isip_reference: [13] J.Picone, M.A.Johnson, and W.T.Hartwell, "Enhancing Speech Recognition Performance+ t24990: isip_reference: [14] A.Ganapathiraju and J.Picone, A Least-Mean Square Error(LMS) Echo Canceller, http:/M z13815: isip_figure: Figure11. An example of an item available for comments on the SWB FAQ page. Users can listen to theP v41647: isip_figure: Figure12. An example of the SWB Progress Report that is generated each week during the project.S -28412: isip_heading_1: 3.6. Quality ControlZ 517658: isip_heading_1: 3.8. The SWB Progress Report     G6l :k18669: isip_reference: [2] Charles T. Hemphill, John J. Godfrey, and George R. Doddington, "The ATIS Spoken6l q35139: isip_figure: Figure2. The distribution of the amount of data per speaker in SWB is shown. Subjects were6l '38634: isip_heading_0: 9. ATTACHMENTS6l $34627: isip_heading_0: 1. ABSTRACT6l 141681: isip_heading_0: 2. HISTORICAL BACKGROUND6l $42443: isip_heading_0: 3. SOFTWARE6l p10836: isip_reference: [1] N. Deshmukh, A. Ganapathiraju, A. Gleeson, J. Hamaker, and J. Picone, Resegmentation6l eq35736: isip_reference: [7] M.Finke and A.Waibel, Flexible Transcription Alignment, in Proceedings of the IEEE6l `n25683: isip_reference: [8] D.Pallett, 1997 Hub-5 Workshop: The Evaluation, in Proceedings of the 1997 Hub-56l !K36623: isip_heading_1: 2.1. Development of an Efficient Segmentation Tool6l Uk33800: isip_reference: [12] N.Deshmukh, A.Ganapathiraju, R.Duncan, and J.Picone, An Efficient Tool For6l t#:20370: isip_heading_1: 3.3. Word alignment tool overview6l x$A39657: isip_heading_1: 2.3. Integrated Project Management Tools6l =%'28118: isip_heading_0: 4. Preparation6l $&/28448: isip_heading_1: 4.1. Echo cancellation6l )q'*27734: isip_heading_1: 4.2. Segmentation6l .K(639436: isip_heading_1: 4.3. Transcription correction6l 3')718775: isip_heading_1: 4.4. Automatic word alignments6l 7*413826: isip_heading_1: 4.5. Manual word alignments6l >j29445: isip_reference: [9] B.Byrne, M.Finke, S.Khudanpur, J.McDonough, H.Nock, M.Riley, M.Saraclar,6l D,/37660: isip_heading_0: 5. SUMMARY OF PROGRESS6l I+;41334: isip_heading_1: 5.3. Preliminary LVCSR experiments6l M.:40100: isip_heading_1: 5.2. Validator speed and accuracy6l S/?10909: isip_heading_1: 5.3. Summary of Switchboard statistics6l W0538234: isip_heading_0: 6. SUMMARY, PLANS AND ISSUES6l \1,26038: isip_heading_0: 7. ACKNOWLEDGEMENTS6l a%2&16679: isip_heading_0: 8. REFERENCES6l hVi39528: isip_reference: [10] A.Ganapathiraju, J.Hamaker, and J.Picone, Syllable-Based Large Vocabulary6l n^i33219: isip_reference: [11] A.Ganapathiraju, V.Goel, J.Picone, A.Corrada, G.Doddington, K.Kirchoff,6l sw17636: isip_figure: Figure3. A SWITCHBOARD resegmentation tool that allows for easy manipulation of segmentation and6l {k33800: isip_reference: [12] N.Deshmukh, A.Ganapathiraju, R.Duncan, and J.Picone, An Efficient Tool For6l 34041: isip_figure: Figure4.6l "}39434: isip_figure: Figure2. A word alignment tool that allows for easy manipulation of word boundaries and for quick tran6l 7m41565: isip_reference: [3] J. Godfrey, E. Holliman and J. McDaniel, "Telephone Speech Corpus for Research and6l 9/28169: isip_heading_1: 2.1. Recording History6l -P13330: isip_figure: Figure6. Workflow diagram for SWB resegmentation project.6l 6m41565: isip_reference: [3] J. Godfrey, E. Holliman and J. McDaniel, "Telephone Speech Corpus for Research and6l X<f26392: isip_reference: [4] SWITCHBOARD:AUsersManual, http://www.cis.upenn.edu/~ldc/readme_files/6l 3q19576: isip_reference: [13] J.Picone, M.A.Johnson, and W.T.Hartwell, "Enhancing Speech Recognition Performance6l >k23133: isip_reference: [1] Charles T. Hemphill, John J. Godgrey, and George R. Doddington, "The ATIS Spoken6l O?318545: isip_heading_1: 2.2. Transcription History6l @:18980: isip_heading_1: 2.3. Recognition and Segmentation6l ]h15326: isip_reference: [4] J.Hamaker and J.Picone, The SWITCHBOARD Frequently Asked Questions(FAQ),6l P)34146: isip_heading_1: 4.7. The SWB FAQ6l zw33362: isip_figure: Figure1. A SWITCHBOARD resegmentation tool that allows for easy manipulation of segmentation and6l 24t24990: isip_reference: [14] A.Ganapathiraju and J.Picone, A Least-Mean Square Error(LMS) Echo Canceller, http:/6l 5p10836: isip_reference: [1] N. Deshmukh, A. Ganapathiraju, A. Gleeson, J. Hamaker, and J. Picone, Resegmentation6l 6~26523: isip_figure: Figure4. In the above waveform, the caller speaks for 21 seconds without an acoustical pause of 0.5 sec6l 8f26392: isip_reference: [4] SWITCHBOARD:AUsersManual, http://www.cis.upenn.edu/~ldc/readme_files/6l :k39933: isip_reference: [15] J.Hamaker, Y.Zeng, and J.Picone, Rules and Guidelines for Transcription and6l ;h15326: isip_reference: [4] J.Hamaker and J.Picone, The SWITCHBOARD Frequently Asked Questions(FAQ),6l A=i33219: isip_reference: [11] A.Ganapathiraju, V.Goel, J.Picone, A.Corrada, G.Doddington, K.Kirchoff,6l GA}39434: isip_figure: Figure2. A word alignment tool that allows for easy manipulation of word boundaries and for quick tran6l B}24344: isip_figure: Figure9. An example of word alignments before and after manual word alignment review is performed. The6l Ch15326: isip_reference: [4] J.Hamaker and J.Picone, The SWITCHBOARD Frequently Asked Questions(FAQ),6l Dx18253: isip_figure: Figure10. An example of the information contained on the front page of the SWB Transcription FAQ.6l %WEz13815: isip_figure: Figure11. An example of an item available for comments on the SWB FAQ page. Users can listen to the6l *Fv41647: isip_figure: Figure12. An example of the SWB Progress Report that is generated each week during the project.6l /G-28412: isip_heading_1: 3.6. Quality Control6l 4H)34146: isip_heading_1: 4.7. The SWB FAQ6l :I517658: isip_heading_1: 3.8. The SWB Progress Report6l AJ17312: isip_table_title: Table 1: Word error rates for each validators transcriptions before manual word alignments and original6l G2K20146: isip_table_title: Table 2: Percentage of each type of error present in the original LDC transcriptions for conversation 3909.6l MLm15013: isip_reference: [1] J. Godfrey, E. Holliman and J. McDaniel, "Telephone Speech Corpus for Research and6l RMo21545: isip_reference: [3] A. Gleeson, J. Hamaker, N. Deshmukh, A. Ganapathiraju, and J. Picone, A Statistical6l X9N '38634: isip_heading_0: 9. ATTACHMENTS6l ^O12692: isip_table_title: Table 3: Word error rates for each validators transcriptions before manual word alignments, along with the6l f,Pi33219: isip_reference: [11] A.Ganapathiraju, V.Goel, J.Picone, A.Corrada, G.Doddington, K.Kirchoff,6l lQ#22267: isip_table_title: Table 4.6l qRr35765: isip_figure: Figure7. Timeline for remainder of SWB resegmentation project: We are currently on schedule6l wSm29208: isip_reference: [18] J.Hamaker, A. Ganapathiraju, and J.Picone, SWITCHBOARD Educational Resources,  <$paratext[Title]> <$paratext[Heading1]> <$curpagenum><$lastpagenum>"<$monthnum>/<$daynum>/<$shortyear><$monthname> <$daynum>, <$year>;<$monthname> <$daynum>, <$year> <$hour>:<$minute00> <$ampm>"<$monthnum>/<$daynum>/<$shortyear> (Continued)+ (Sheet <$tblsheetnum> of <$tblsheetcount>) <$marker1> <$marker2> isip_paratext <$paratext>  isip_name/INSTITUTE FOR SIGNAL AND INFORMATION PROCESSINGisip_toc<$paratext>\t\t<$pagenum> isip_paranum<$paranumonly><$monthname> <$daynum>, <$year>"<$monthnum>/<$daynum>/<$shortyear> <$fullfilename> <$filename>isip_document_name2Improved Monosyllabic Word Modeling on SWITCHBOARDisip_document_dateAUGUST 15, 1998MHTMLVVHeadings   ACcc isip_title}}>>??A@@ isip_prefaceDDTOCEEIXFFLOFGGLOT isip_preface&&TTVVXXPPbCwwdd isip_prefaceH Figure7.D \::4.1.q?,%:2.2.l CC Figure8.,0 >Yt fW>**4.: 9_I[)4.3.9aa6G9#9=9,9-]p:"4.3. K99$ P9.9>9?9@9A9B\ ]q :EL%7.Y7H[2]Zp7H[5]Yu3 Yv" G ] *]  Yw" ^X&++>>Yx1 4q@qAqBqCqD^Z&(\ :qI''7+9+ R T V X Z \ ^ `'j"*(R'l(f                                     ; = ? A C E G I K M# Q#Y*[ (T(h]_a c           e       g       i     *km&o(z(|'~'      ' ) + - / 1 3 5 7 9 ; = ? A C E G I K M i n p r t v x          ***                                                    ! # % ' ) + - / 1 3 5 7 9 ; = ? A C E G I K M O Q S U W Y [ ] _ a c e# i k m o  s u w y }                                                               "    *   **&**'  U W Y #%'7] *aH qJ'''XGEn&7 >'`H[7]Y7H[3]X) +D+E+F+G+H+I+J+K+L+MY7H[4]7M 7L%1.8L%2.n&l?CC Figure9._ 8[%2.1.8%[)2.2.8*[)2.3.8<L)3.84[%3.1.8F[)3.2.8K[)3.3.8P[)3.4.,<:02.3.,=*3.,>:3.1.,? ,@:3.2.,A ,I",F",J Yy 8U[)3.5.^\2]: 3.6.Y:1.2.-Y7H[6]Y atHH[1]Yz .*6..*7.8_L)4.] :8i[)4.1.8n[)4.2.8vL)5.ZM%9@ ^^2.< 8{L%6.Y *Y*1.a ^`2\_ ^b2a:1.1.Y * ^d2a[%1.1.+*2.+:2.1.]CC Figure7..A f].I ^f2] :,x:3.3.,y ]  ,{",|",}",~",", nV]! Kp{ ,:43.4.]" >]# \" >^h2@ ^j22:(4.2.^l2>B YY [1]^n2^oR]x *]CC Figure5.]C Figure1.`H[4]^r&^t&^v2^x2^z22 nV^|22 ^~2-*5.2E^2- .[2]2^2^2^Roa^\ Table 1.o_\F Table 3.o o~p:Vo^\ Table 2.^CC Figure13.dрH[6]a[)1.2.a[)1.3.a Y E#""a Y7H[7]a eH [9]Y a bV bH[5]b\CC Figure3.b b::3.7.[F /\""\ "fW\ \: 1.3.`H[3]EmC Figure11.c: dCC Figure2.dH[17]nVewH[8]@" e| e @""eH [10]eH [11]e eȀH[16]eʀH [12]fWgVW<iW<i]<g XiXTgXgXi]<656D656^6΀5g<gWi]Xi]XgXhCXhDXi!]Xg<hEXhFXgXi"XXg<hG<?::3.5.hHWi#]Xg"Wi$]Xg$Xg%Xg&Xg'Xg(Xg)Xg*<g+Wi%WXg-Xg.Xg/Xg0Xg1Xg2Xg3XE$""g4Xg5Xg6<g7WhJXg9XhKXhL<hMWi&]XhOXi'XXhQXhRXi(WXhTXhU<hVWil]XgG<hXXhYXhZXh[XiyCC Figure4.h]Xh^Xh_<h`Wi;XXi]hcXhdXi<XXhfXhgXi=WXhiXhjXi>]WhlXhmXhnXho<i?]Xi@]XiA]XiB]XiC]XiDXXiEWXiF]WiG]XiH]XiIXXiJWXiK]Xi j  j : 3.8.j% j J kC Figure6.iR]XiS]XiT]XiU]XiV]XiW]Xi]]i]iWi]i]]i]iWi]iWi]i]i]i]i]i]i]i]i]nVk߀H [13]k6[14]l!H[15]lfCC Figure10.lzNlNlNmNlOlOlOi]]i]]i]]i]]i]]i]]i]]iW]i]]i]]i]]i]]i]]i]]lOlOlO5lO5mCC Figure12.m[[3.6.mHIm[[3.7.mJmKmLmMmNmOmPKlNlOlOlOlOmQSm[[3.8.mSmKnVnVmWnVmYKmZInVm\m]Um^UoVlQl omambUmcUnRoVooVmhmiUmjKmkIommnVmoKmpInVmrnVmtKmuInVmwmxmymzKm{InVm}m~mmmmmKmKnVoVopH\\ Table 4.pj p~H[18]oRo~ooooop pZVp=ZVpDVpEVpFVpGVd9>@d] &d*  HHˆ* HHˆ…))<`EXECUTIVE SUMMARY ($ YThe SWITCHBOARD(SWB) Corpus consists of 2430 conversations digitally recorded over long 02^distance telephone lines. The SWB Corpus totals over 240conversation hours (elapsed time) of ]data. The average conversation duration is six minutes. The transcriptions contain more than ^3million words of text. The SWB Corpus includes more than 500adult-aged speakers and covers Vmost major American English dialects. Such impressive statistics make SWB the premier \database for telephone bandwidth large vocabulary conversational speech recognition(LVCSR) eresearch. The goal of this project is to resegment the speech data and correct the transcriptions in @5an effort to significantly advance LVCSR technology.  ^We have completed the first six months of the SWB project and have released 525 conversations 0`with corrected segmentation, transcriptions, and automatic word alignments. Additionally, there _are 275 conversations awaiting release with automatic word alignments. These 800 conversations Ycomprise 41% of the conversations used in the WS97 partition, and 33% of the entire SWB Xcorpus. We have also performed a major overhaul of the lexicon by removing incorrect or ^unnecessary entries and making the lexicon case sensitive. Finally, we have created extensive ]documentation including a statistical analysis of the conversations and a description of the @[transcription conventions. All such information is on-line and available via the Internet. & VIn an effort to make the resegmentation process highly efficient, we have developed a 0.fsegmentation tool that is specifically tailored to the needs of the SWB project. It is written in C++ cand uses Tcl-Tk (v8.0) for the user interface. It is highly portable across environments including @PWindows95. Our validation staff uses this tool to execute the following tasks: ' ^ fsegmentation+: creation of a new segmentation that consists of utterances typically 10seconds in #j@Mduration and excised at significant pause boundaries and/or turn boundaries; + |`Ntranscription validation+: correction of the orthographic transcriptions; 0 dword alignment+: adjustment of word boundaries produced by a forced alignment that uses the new #@7transcriptions with our best phonebased LVCSR system. \ aOur cross-validation tests on relatively clean utterances have shown that our validators have an 0_average word error rate(WER) of 2.6% (this number varies dramatically with the convention one ]uses for scoring). This is a substantial improvement from the 8%WER (measured under similar ^conditions) present in the current best transcriptions recently released by LDC. After manual Xword alignments our final quality control stepthe WER is reduced to 1.5%. Our best `validators are able to reduce the WER to less than 0.5%. We are currently implementing measures eto reduce the average error rate to less than 1%. To place this in perspective, a typical six minute `conversation has approximately 1200words, which implies that the final transcription will have @7approximately 12words in error for each conversation. e; [To further underscore the importance these new transcriptions, we have demonstrated a 1.9% pI\absolute improvement in recognition performance (from 49.7% to 47.8%) simply by training on bthe new transcriptions. Equally exciting is the fact that recognition error rates on monosyllabic Uwords dropped a similar amountfrom 49.6% to 47.7% (and performance on other words Zdropped from 49.1% to 47.4%). Since monosyllabic words dominate the SWB Corpus, this is a @!particularly significant result. HHˆ*HHˆ ld. Rdb  dX  rkK<4z_o  ssrkK<4zĂR6^ R 6m !l 6^ 6 @(: KTimeline for the remainder of the SWB resegmentation P@;project. Our anticipated completion date is December1999. IH`5^L IH`533 l IH`5^M IH`5}S^ W UTUT@h  DH^N  DH33l DH^O DH _7W7UTUT@h  H/(د^P H/(دl H/(د^Q H/(د9#(+ Word error rates for each validators p.transcriptions before manual word alignments, (along with the WER for the original LDC @(transcriptions for conversation sw3909. DT^RDTbl DT^S DT?#(-A  breakdown of the error modalities for p0the original LDC transcriptions of conversation -sw3909. Consistency of contractions has been @ observed to be a major problem. ll7+)8ll7 l ll7+* 8ll7 gg-`Bi-Annual Status Report For .` 6` :` !1 'Improved Monosyllabic Word Modeling on @ SWITCHBOARD ` 2` 5` 6` `  h  3` /A` 3` 4` `submitted by: 1` :`9J. Hamaker, N. Deshmukh, A. Ganapathiraju, and J.Picone ;`0Institute for Signal and Information Processing <`2Department of Electrical and Computer Engineering =`Mississippi State University >` Box 9571 ?`413 Simrall, Hardy Road @`%Mississippi State, Mississippi 39762 A`Tel: 601-325-3149 B`Fax: 601-325-3149 AC`*email:{hamaker, picone}@isip.msstate.edu }Nf^n ZuxNf^`blW?#`Conversation Statistics