RESOURCE MANAGEMENT TEST: 2

ORIGINAL TEST DATE: October 1987

RECOGNITION SYSTEMS: speaker-dependent continuous speech transcription
                     speaker-independent continuous speech transcription   

SPEAKER-DEPENDENT TRAINING MATERIAL:
     . ((all 12 speakers from Speaker-Dependent Training Set) X
        (all 600 "sr" and "st" sentence utterances/speaker))
       = 7200 training utterances from CD2-1.1 and CD2-2.1: /rm1/dep_trn
     . Files are listed in CD2-4.2: /rm1/doc/train/6a_deptr.ndx and
       CD2-4.2: /rm1/doc/train/6b_deptr.ndx

SPEAKER-INDEPENDENT TRAINING MATERIAL:
     . no restrictions - can use material from CD2-3.1: /rm1/ind_trn as well
       as DARPA TIMIT and other speech corpora

TEST PARAMETERS:
     . grammar:
          (1) no grammar - full branching

SPEAKER-DEPENDENT TEST MATERIAL:
     . ((8 speakers from Speaker-Dependent Development-Test set) X 
        (25 sentence utterances/speaker)) 
       = 200 test utterances from CD2-4.2: /rm1/dep/dev
     . 4 of these speakers are from the March 87 test
          - 10 retest utterances for each speaker
          - 15 new utterances for each speaker
     . 4 of these speakers were not used in the previous test
     . files are listed in CD2-4.2: /rm1/doc/tests/2_oct87/2_deptst.ndx

SPEAKER-INDEPENDENT TEST MATERIAL:
     . (((6 speakers from Speaker-Independent Development-Test set) X
         (10 sentence utterances/speaker)) +
        ((4 speakers new speakers from Speaker-Dependent Development-Test set
          used in the speaker-dependent test) X 
         (25 sentence utterances/speaker)))
       = 160 test utterances from CD2-4.2: /rm1/ind/dev_aug /rm1/dep/dev
     . files are listed in CD2-4.2: /rm1/doc/tests/2_oct87/2_indtst.ndx

SCORING SOFTWARE SETTINGS:
     The following command should be used when scoring this test:

     wgscore hypothesis-file [system-code] [system-description]

     Note: Homophone substitution errors are normally scored as correct in
     the no-grammar case, and incorrect when a grammar is employed.  In this
     early test, however, homophone errors were scored as incorrect in the
     no-grammar case.  Therefore, this test should be scored using the
     scoring software settings for "with-grammar scoring" (wgscore).