Tied mixture modeling optimization for Korean-digit in the embedded ASR system

Abstract

In the embedded Automatic Speech Recognition (ASR) system, Semi-Contimrorrs Hidrlen Markov Model (SCHMM) or Tierf-Mi.rtirre (TM) model is one of tlie most promisirig acoustic modeling metliods that solve the size problem of the existing Continirons Hirldcri Markov Model (CHMM) while minimizirig the recognition peifiinnancr rlegra(iation. Moreover. f o r a geiierul isolated n,ord task, coiite.rt rlepenrlent nior1el.v sirch us tri-phones are nsed to guarantee high recognition performance of the embedded sy tem. However. tu nse the models constructed only in this way alone cannot be siifJicienr to render improved recognition rate in Korean-digit speech task 4 w r e a lurge niirtrral similarin e.rists. Hence. w e consfrnct new deilicated HMM ' S f o r all or parts of Korean-digit that has exclusive srafes using the same Gaussian pool of previoirs tri-phone mode1.s. This remedial actiori allows rlie strncture qf entire HMM.s maintained while minimizing the occupied memory space. Representative esperiments are rrpecred to reduce worderror-rate on the Korean-digit task by about 56% in enniporison with using only general rr-plione models. ' Mixture Model, Embedded ASR System. Index Terms Exclusive HMM's, Korean Digits, Tied