LVCSR of meeting dataMartin KarafiatThis poster is focused in the building of recognizer for largevocabulary continuous speech recognition for meeting data. Basic featureextraction and HMM training methods are used but decoding algorithm isadjusted for a minimization of computation time without waste ofaccuracy. The combination of best first decoding with cross-internaltriphones and time synchronous decoding with cross-external triphones isimplemented for this purpose. This poster also discusses theimportance of selection and good balance of language models trainingdata. All experiments are performed on ICSI meetings database. |