Article details

Research area
Speech recognition

Presented at the 2nd MCM of COST278, Zurich, Switzerland


Qian Yang, Jean-Pierre Martens

The ELIS-RUG Pronunciation Modeling Approach


This approach consists of the following steps:

  1. Create standard recognizer: one typical pronunciation per word (exceptions)
  2. Create expert transcription which is aligned with lexicon transcription: use acoustic models
  3. Learn pronunciation rules that can explain most of the discrepancies between aligned transcriptions
  4. Use these rules to transform typical word models into new ones: multiple pronunciations + probabilities
  5. Put new word models in the lexicon of the recognizer
  6. Retrain acoustic models and return to step 2
Read/download now