THE BROADCAST NEWS RECOGNITION TASK

Television and radio news data and speeches

Various focus conditions

  • F0 - Baseline
    		•
  • F1 - Spontaneous
    		•
  • F2 - Telephone
    		•
  • F3 - Degraded
    		•
  • F4 - Music
    		•
  • F5 - Non-native
    		•
  • FX - Other
    		•

Test data selection

  • To minimize year-to-year variability

  • To use recent broadcasts for keeping LMs current

  • Randomized excerpts selection

  • To represent the training set (not to maximize focus condition coverage)

  • 166 minutes of news, 144 variable-length excerpts

  • 14 one-minute speech excerpts

  • Audrey Le got commended for her work on test data selection