The PEM ("partitioned evaluation map") file format is given in the
SCLITE documentation available through NIST's web page
(http://www.nist.gov/speech/tools/index.htm). Each record contains 5
fields: (shown here on separate lines, but one record per line)

 <filename> 
 <channel ("A" or "B")>
 <speaker ("unknown_speaker")>
 <begin time>
 <end time>


A system should process the appropriate PEM file:

CONVERSATIONAL SPEECH RECOGNITION EVALUATION:
---------------------------------------------
  hub5e_01.pem   ::  all 60 conversations
  swbd.pem       ::  original switchboard conversations
  s2p3.pem       ::  switchboard 2 phase 3 conversations
  s2p4.pem       ::  switchboard 2 phase 4 conversations (cellular)

  swbd_s2p3.pem  ::  original switchboard and switchboard 2 phase 3
                     conversations
  swbd_s2p4.pem  ::  original switchboard and switchboard 2 phase 4
                     conversations
  s2p3_s2p4.pem  ::  switchboard 2 phases 3 and 4 conversations


  Participating sites may choose to run their systems on all three
  parts of the EvalSet or on only one or two of these parts.
  Evaluation must be done on all twenty conversations of whichever
  part or parts are chosen.  A single baseline system should be run on
  all parts for which the site is participating.  Additional alternate
  systems are welcome, and these may include systems specifically
  designed or tuned for one part or parts of the data, such as the
  cellular data, and run only on that part.

  The identity of the data sources is not to be provided to the
  baseline system.

PHONETIC ANALYSIS COMPONENT:
----------------------------
  phonetic.pem   ::  all 21 conversations used for the phonetic
                     analysis task