SWITCHBOARD is a corpus  of spontaneous conversations which  addresses
the growing  need  for  large  multi-speaker databases   of  telephone
bandwidth speech. The   corpus  contains  2430  conversations averaging 6
minutes in length; in other words, over 240  hours of recorded speech,
and about 3 million words of text, spoken by over 500 speakers of both
sexes from every major  dialect of American  English.

<br><br>

<b>Project Goals</b>
<ul>
  <li>
       manually creating a new segmentation of the training database
       that consists of utterances typically 10 seconds in duration
       which are excised at significant pause boundaries and/or turn
       boundaries
       
  <li>
       manually correcting the orthographic transcriptions for the
       training database (the test database was already corrected at
       WS'97 by the PI)
       
  <li>
       automatically creating a new set of word alignments by
       performing supervised training on the new segmentations using
       our best phone-based recognizer, and manually review these
       word alignments for accuracy
       
  <li>
       manually creating a new test set segmentation that is
       consistent with the training database; recalibrating the
       improvement in performance achieved by syllable-sized acoustic
       models
       
  <li>
       performing an error analysis to determine the dominant error
       modalities of these systems.
</ul>

<br><br>

<b>Plan</b>
<br>

Our plan for resegmenting the database consists of a six step
process:
<ul>
  <li>
       echo cancel the data
       
  <li>
       adjust utterance boundaries
       
  <li>
       correct the orthographic transcription of the new utterance
       
  <li>
       readjust boundaries if necessary
       
  <li>
       perform supervised recognition on the new utterances to get a
       time-aligned transcription
       
  <li> 
       review the word boundaries for gross errors
</ul>

<br><br>

<b>We are using the following CDs in this project:</b>
<ul>
  <li>
       "Switchboard-1 Transcriptions: Intermediate Version," August, 1997
       
  <li>
       "Switchboard-1 Telephone Speech Corpus: Release 2," August, 1997
</ul>

<br><br>

An overview of the transcription process is given below:

<br><br>

<center>
<img src="../doc/overview/work_flow_diagram.gif" alt="Work flow diagram">
</center>