BN RECOGNITION IN OTHER LANGUAGES

Spanish and Mandarin

Data sets

  • 30 hours training data, 1 hour development test, 1 hour test data
  • single transcription, no annotation

Systems on Spanish

  • BBN --- BYBLOS, MFCC, 5-state HMM
    • 20.3% WER (20.4% on English)

  • CMU --- SPHINX-III, continuous-density tied senones
    • 23.5% WER (24.0% on English)

Mandarin systems

  • Dragon --- PLP cepstra, IMELDA transform, speaker normalization, rapid adaptation
    • 20.2% character error rate

  • IBM --- turn detector, single pass decoding, iterative adaptation
    • 19.8% character error rate