BN RECOGNITION SYSTEMS - II

CMU Dragon IBM
Front-end
  • 39 features (MFCC, energy)
  • 36 features (MFCC, energy)
  • PLP
  • LDA + normalization
  • -
    Segmentation
  • Cross-entropy measure
  • GMM classification
  • CI phone recognition
  • Kullback-Leibler distance
  • GMMs + BIC
  • Acoustic Model
  • Tied states
  • MLLR adaptation
  • Baum-Welch adaptation
  • Word-end in clustering
  • Speaker normalization
  • Only F0 and F1
  • Focus-specific adaptn.
  • DT clustering
  • OFS transforms
  • MLLR adaptation
  • Language Model
  • CU-CMU SLM toolkit
  • 64K words
  • 1st pass - trigrams
  • Rescoring - 4-grams
  • 57K vocabulary
  • Interpolated trigrams
  • Adapted remote corpora
  • 64K lexicon
  • 4-gram mixture LM
  • Decoder
  • N-best lists
  • Rescoring with 4-grams
  • Multipass wordgraph
  • MLLR + Xword trigram
  • -
    Performance
  • 18.5% WER
  • 26.9% WER
  • 17.7% WER