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

In the embedded Automatic Speech Recognition (ASR) system, Semi-Continuous Hidden Markov Model (SCHMM) or Tied-Mixture (TM) model is one of the most promising acoustic modeling methods that solve the size problem of the existing Continuous Hidden Markov Model (CHMM) while minimizing the recognition performance degradation. Moreover, for a general isolated word task, context dependent models such as tri-phones are used to guarantee high recognition performance of the embedded system. However, to use the models constructed only in this way alone cannot be sufficient to render improved recognition rate in Korean-digit speech task where a large mutual similarity exists. Hence, we construct new dedicated HMM's for all or parts of Korean-digit that has exclusive states using the same Gaussian pool of previous tri-phone models. This remedial action allows the structure of entire HMMs maintained while minimizing the occupied memory space. Representative experiments are expected to reduce word-error-rate on the Korean-digit task by about 86% in comparison with using only general tri-phone models.