<?xml version="1.0" standalone="yes"?>
<Paper uid="H94-1007">
  <Title>Spontaneous</Title>
  <Section position="3" start_page="0" end_page="27" type="metho">
    <SectionTitle>
1. MATERIAL DESIGN
</SectionTitle>
    <Paragraph position="0"> The prospective applications for the Macrophone data partly determined the linguistic design of the material and the population of speakers to be recorded. Examples of the applications include: * voice interactive systems to support telephone services like collect calls, third-party billing, or rate inquiries * database information retrieval services that might provide schedule or availability information about transportation or other public services in a limited semantic domain * systems for ordering theater or stadium tickets, or for making medical or other appointments * systems for manipulating bank accounts or other financial resources.</Paragraph>
    <Paragraph position="1"> I.I. Goal The goal of the Macrophone project was to provide a basic set of common spoken material suitable for training and evaluation of speech recognition systems for telephone-based applications, particularly those that use names, places, times, and numbers in a North American context.</Paragraph>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
1.2. Sources
</SectionTitle>
      <Paragraph position="0"> The material collected came from pools of prompt texts.</Paragraph>
      <Paragraph position="1"> Materials were selected from these pools automatically and combined into a prompting sheet that is mailed to a person. In the Macrophone corpus, 45 responses were solicited on each sheet.</Paragraph>
      <Paragraph position="2"> Of the 45 responses, 34 were read and 11 were spontaneous. The prompt material for the spontaneous utterances was designed to elicit particular responses or types or ranges of responses. The following describes the read and spontaneous material presented, with examples.</Paragraph>
      <Paragraph position="3">  Examples of the sentence types (TIMIT, WSL ATIS) are: Will you please confirm government policy regarding waste removal? The budget is a long way from completion, however.</Paragraph>
      <Paragraph position="4"> I'd like to buy a coach class ticket for a flight from Columbus to San Jose.</Paragraph>
      <Paragraph position="5"> For more information on the TIMIT sentences, see Lamel et al. (1986); for the WSJ sentence set, see Paul &amp; Baker (1992); and for the ATIS materials, see Hirschrnan-MADCOW (1992). Spontaneous There were 11 prompts in the interaction that solicited spontaneous speech. Of these, ten were fixed and one was rotating. Some of the fixed questions provided additional demographic information about the speaker. Five of the fixed questions were designed to elicit an answer of yes or no. Six of the fixed questions were printed on each sheet: Are you ready to start7 (y/n) Are you calling from your home phone? (y/n) Do you speak any language besides English at home? (y/n) Please name a major city in your state.</Paragraph>
      <Paragraph position="6"> Would you be willing to participate in another study like this one?(y/n) We would appreciate any comments you may have about this recording session. Please record your comments.</Paragraph>
      <Paragraph position="7"> The four unpfinted questions were: Are you using a cordless phone? (y/n) What is today's date? What time is it now? What is your date of birth? The rotating question was taken from the set: Please say any number from 1 to 100.</Paragraph>
      <Paragraph position="8"> Please say any number from one thousand to one million. What is your house number?</Paragraph>
    </Section>
    <Section position="2" start_page="0" end_page="27" type="sub_section">
      <SectionTitle>
Prompt Pools
</SectionTitle>
      <Paragraph position="0"> The utterances were selected from pools of prompts that were designed for this project. In particular, each pool was designed with knowledge of the material in the other pools; thus, for  example, &amp;quot;Jones&amp;quot; was in the family-name pool, but was excluded from the street-name pool. The prompt pools were: Place Names -- All United States and Canadian cities with population over 150,000, and at least two cities in each state in the United States; all cities worldwide with over 2 million people, and at least one city from most major commercial nations.</Paragraph>
      <Paragraph position="1"> Personal Names--A gender-balanced list constructed from the 600 most common first names and the 600 most common last names in the United States. Some common first and last names that are also common words were omitted.</Paragraph>
      <Paragraph position="2"> Street Names -- The most frequent 971 names compiled from the ZIP+4 directory BODY field by counting block faces. We deleted state names, city names, first and surnames that occur in our other lists, and we deleted the letter names A - Z. Application Words -- A list of 674 control words selected from existing and imagined telephone applications.</Paragraph>
    </Section>
  </Section>
  <Section position="4" start_page="27" end_page="27" type="metho">
    <SectionTitle>
2. DATA FLOW
</SectionTitle>
    <Paragraph position="0"> ThE process is shown in Figure 1 and is described in the foUowing sections.</Paragraph>
    <Paragraph position="1"> J---If D ..... ~ Collect Tele- rN Verify &amp; Package mnoute h~phone Signals ~ Transcribe ~ &amp; Deliver Figtwe 1. Data Collection Process 2.2. Hardware and Software for Phone  SRI designed and implemented a set of systems for this data collection. Dialogie hardware resident in an IBM-compatible PC provided a digital connection to 10 telephone lines that are available toll-free to callers. The Dialogie/PC system concentrated the data and sent it to the disk of a small Sun Microsystems SPARCstation. As convenient, the data was moved to an archive disk from which the data was archived onto Exabyte tapes and pressed onto CD-ROMs for further manipulation. The labor-intensive part of the Macrophone corpus development was the transcription and verification of the spoken material. This was performed with special-purpose software running on Sun ELCs.</Paragraph>
    <Paragraph position="2"> 2.3. Time and Resources Used The design and selection of material, set up and monitoring of the telephone collection system, the verification/transcription of the utterances, and finally the delivery of the resultant files took about seven months of professional labor and about eight months of semi-skilled labor. After the linguistic material had been decided upon, the project yielded 200,000 utterance files with verified transcriptions and demographic headers in five calendar months.</Paragraph>
    <Section position="1" start_page="27" end_page="27" type="sub_section">
      <SectionTitle>
2.4. Material Distribution
</SectionTitle>
      <Paragraph position="0"> The data collection process began with the distribution of the material to prospective callers. The material was presented in the form of unique prompting sheets to guide the caller through the telephone interaction. The sheets were designed following a fixed format, but provide different read material. Each prompting sheet contained a different set of read material, and each caller received a different sheet. A sample prompting sheet is shown in</Paragraph>
    </Section>
  </Section>
  <Section position="5" start_page="27" end_page="27" type="metho">
    <SectionTitle>
BEFORE YOU C/U.L
</SectionTitle>
    <Paragraph position="0"> Please write y~xr six-dlglt panel idmatiJicatinn nnmbc~r in the spree provided in the mld~.o of this page. (It i~ I~'inted at the top of your cover letter.)</Paragraph>
  </Section>
  <Section position="6" start_page="27" end_page="29" type="metho">
    <SectionTitle>
PLEASE C~.L 1-81Xl-XXX-XXXX
</SectionTitle>
    <Paragraph position="0"> A eomlmter ~ amwer end ask you the follovdng questinm.</Paragraph>
    <Paragraph position="1"> ~ l&amp;quot;nankyou for mll~ .f~tr~ ~o~ reeor~t~ syswm. Your ~ict will b~ record~ and u~ed\[~ ~sm~ and dewlopmmt of speech le~nolo~y, lf yoa do not w~h to Imv* your yore recorded and real for th~a * purpa~, you may ~ang up now.</Paragraph>
    <Paragraph position="2"> ~h# ~oMon will begin with a few queJgonl, Your anawera provide m with hr~rtant in~ormmion about vocal qmllly cnd ~peec~ pert~rn~. Aft2 of your r~Ip~ will b~ kept canfldenKat Are you ram'y to ston? (yot~ re~me) Are you eall~ from your hom~ ph~t? (your re~te) Do yalt ~ any li~,nSua~* b~Mea E, rq~qLQ al ~or~? ~our reapome) Plmae read your paa el ldent~ion number which you ~l~d in below: (write your pad fdenll|mllan number here) Thank you. You will now be a~.ed m read ea~ of Oie O.ems in tlle rlg,ht-hand cohunn.  1. (a wordJ mblma 2. (axnt~nee) Ceuld you give me a lid ct -n tflemeon fliShm fram Loa Angelea 3. (a lel~onc number) ~79) 528-5883 4. (a place) Ctnwn. CYaina 5. (a~l~ll~word) M.A.N. If. O. S. K. I 6. (a number) 51.611 melaa 7. (anaraC/) Dmmy Peyne. of 251 Ironwood W,y &amp; (a number) 4 9. (a smlence) I~lit~ policy wm to keep the reut~ opea end prote~ the aettled aree&amp; 10. (a time) 9:22 11. (a word) south 12. (a sm/cnee) Ul h~ that only people with money wm be approved.&amp;quot; she mid. 13. (a dollar ~alolmO $219 14. (anamC/) Pamy Ward of the I R S 15. (= dollar ~tounO $6388.12 16, (adam) Tuesdty. Maraa 20. \]990 17. (a #ira/c/we) The Sla~e of Lib~ty end Ellis htlend ere within the waters of New york Bay. 18. (a word) flow 19. (a dollar ~nounO $2 ~. (a word) leave 21. (a C/~ card number) 1495-1772-1515 22. (a am/cnce) The C/~tteway reded abropay at the d-~re. 23. (a name) Leuise Blmlford, of 143 Wexlneld Roed 24. (a place) ~ Jemeiea 25. (a1~aatm) 118 26. (a doUar a~na~O $600 27. (a word) slower 28. (a apelled wotrl) B.R.U.S,H 29. (a name) Dim RidL at 388 Bea~ 349. (a ~en/cnccC/) How am I get from the Ta~ma ailport to downtow~ 31. (a number) 17~905 kilosrerm 32. (aJen/cneC/) irs all psydlolcgicaL 33. (a word) ~ee 27;cart you  WouM you be willing to parffc~ate in anainer awdy likJC/ ~1~ one? (your re~aae) WC/ would appreciate any ccornmen~ you may haw aboat inb F~contln8 action. PleaaC/ recotd your corm. (yatr re~e) you wry much for parllcipatin8 in ~ data collection C/~bn. You may now hang up the/clephone.  Prospective callers were solicited through a market research firm that was able to select a sample from their panel of 400,000 U.S. households. Since no incentive was offered for placing the call, a conservative estimate of a 25% response rate was used. Twenty thousand (20,000) prompting sheets were mailed, which resulted in 6700 calls, at a 33% response rate. The sheets were sent out as six separate mailings of 1000, 2000, 5000, 5000, 5000, and 2000 at approximately one-week intervals. Calls typically started coming in the day following a mailing and peaked three days after a mailing. Although ten lines were available, all ten were never activated at once. The mailings and the response rate are shown in Figure 3. Each vertical bar shows the date and the number of prompting sheets in a mailing, and the lower line shows the number of calls received per day.  The targetpopulation was specified to consist of equal numbers of males and females between the ages of 10 and 80, balanced between ages 20 and 60 and fewer in the 10-19 and 6180 age groups. The target population was also specified to be gengraphieally balanced according to the latest census figures. The sample population was selected by the market research firm to compensate for different expected response rates. Past experience indicated higher expected response from females than from males, lower response rates among young people, higher response rates among elderly people, and lower response rates from people with household incomes above about $40,000. Figure 4 shows the mailings sent out (upper curve) and the responses received (lower curve) as a function of age. The low response rate among people aged 20-30 is due to income skew in the mailing sample, and partly due to low response rates among the people in that 20-30 age group who received sheets.  The recipients of the prompting sheets were instructed to call a toll-free 1-800 number which connected them to one of 10 digital telephone lines set up to receive calls. All data was recorded directly from T1 digital telephone lines in 8-bit mu-law format using Dialogic hardware installed in an IBM-compatible PC. The PC, which operated under Interactive UNIX, was a 33 MHz 386 with 16 MBytes of RAM. Each response was written as a separate file to one of two 2-GByte disks of a Sun Sparestation ELC. A completed call resulted in about 2 MBytes of data. The data collection system is shown in Figure 5.  Software was written to play out prerecorded prompts and record the interaction. As a half duplex system, the data collection system was only capable of either playing out a prompt or recording a response. Care was taken to truncate the end of the prompting text in an attempt to keep callers from responding before the system began recording. It was necessary to remove the written text from the printed sheets for a few of the prompts to force the participant to listen to the entire prompt before responding.</Paragraph>
    <Paragraph position="3"> An average telephone call took about six minutes and resulted in about four minutes of collected speech (including two seconds of silence at the end of each utterance). Approximately one third of the coLlected data was silence.</Paragraph>
    <Paragraph position="4"> 2.6. Verification and Transcription Verification of the read responses and transcription of the spontaneous responses was performed by temporary workers using SRI software written for Sun ELC computers.</Paragraph>
    <Paragraph position="5"> The data verification occurred in two steps. Since each of the 20,000 sheets was unique, it was necessary to supply a unique sheet identifier to bring up the default transcription for the read  material. The sheet identifier was in the form of a 10-digit telephone number and was transcribed in a first pass along with three other responses. These items were then propagated to the headers of all 45 speech files produced from that call. In addition to the demographic responses, a gender indication was included (decision made by the transcriber). The demographic information items are responses to the following prompts: Do you speak any language besides English at home? Are you using a cordless phone? What is your date of birth? The second step of data verification involved providing an orthographic transcription of each utterance. Each utterance waveform was displayed on the transcriber's console and played through the computer's audio port. Default transcriptions were provided for read data, previously transcribed demographic ,d, ata, and predictable responses to spontaneous questions such as' Are you :ready to start?&amp;quot;, &amp;quot;What is today's date?&amp;quot;, &amp;quot;Would you be willing to participate in another study like this one?&amp;quot;. Utterance files which did not contain any speech or contained truncated speech (approximately 7.5% of the total) were discarded. Utterances which were difficult to transcribe -- those containing word fragments, mispronunciations, and other disflueneies -- were set aside for a linguist to review. All other utterances were transcribed according to what was said; the transcriptions also included markings for non-speech events such as background noise, background speech, line noise, mouth noise, and verbal hesitations.</Paragraph>
    <Paragraph position="6"> 2.7. Package and Delivery Each file was written as 8-bit mu-law with a SPHERE header. The headers contain information about the data as well as demographic information about the caller. All data files were written to Exabyte tape and shipped to LDC, where the files will be pressed onto CD-ROMs and made available to members of the Linguistic Data Consortium.</Paragraph>
  </Section>
class="xml-element"></Paper>