Decoder JULIUS

• Interface with Acoustic & Language Model
• Input
  waveform files (16bit PCM), MFCC files (HTK)
  Microphone (Sun, SGI, PC/Linux, DAT-LINK/netaudio)
• 2-pass decoding (tree-trellis search)
  1st pass: real-time processing with simple constraint
  intermediate: word-trellis with index
  2nd pass: stack-decoder to output N-best candidates

Overview of Decoder JULIUS

cross-word phone model language model search approx.

1st pass approximate 2-gram 1-best

2nd pass accurate 3-gram N-best

Decoding Parameters

• 1st pass: frame-synchronous beam search
  • tree-structured lexical entry vs. (partially) linear entry
  • 1-best approx. vs. word-pair approx.
  • 1-gram factoring vs. 2-gram factoring
  • cross-word triphone handling with 1-best approximation
• 2nd pass: stack decoding search
  • word graph vs. word trellis $\rightarrow$ word-trellis index
  • beam search vs. best-first search $\rightarrow$ enveloped best-first
  • N-best candidates [standard] vs. 1-best candidate [fast]
  • accurate cross-word triphone handling
    without delay [standard] vs. with delay [fast]
• Others
  • beam width
  • LM weight
  • insertion penalty
• Gaussian Pruning in PTM model application
  • safe... already computed k-best values as threshold
  • beam... set up a beam width in intermediate dimensions
  • heuristic... heuristic estimation of yet-to-be-computed dimensions