BN RECOGNITION SYSTEMS - I


BBN CU-HTK CU-Con
Front-end
  • 36 pole LPC
  • 33 features
  • VTN
  • LDA + normalization
  • 12th order PLP + energy
  • Segmentation
  • CI system
  • Gender-dependent models
  • Agglomerative clustering
  • Kullback-Leibler distance
  • Merge adjacent segments
  • CMU's segmentation
  • Acoustic Model
  • Gender-dependent
  • Type-independent
  • Xword triphones
  • Gender-dependent
  • Type-dependent
  • Xword quinphones
  • 5-pass MLLR adaptn.
  • Word-end information used
  • RNNs
  • Gender-independent
  • Type-independent
  • WI triphones
  • Syllable information used
  • Language Model
  • Trigram and 4-gram
  • 1st pass - trigram
  • Adapted remote corpora
  • Trigram and 4-gram interpolated
  • Trigram for initial MLLR phase
  • Adapted remote corpora
  • 65K lexicon
  • Trigram and 4-gram interpolated
  • Adapted remote corpora
  • Decoder
  • Xword trigram
  • 2-pass decoding
  • Fast match with CI models
  • Multipass lattice
  • MLLR + Xword trigram
  • NOWAY stack decoder
  • WI 4-gram
  • Performance 20.4% WER 15.9% WER 27.2% WER