METHOD OF CREATING ACOUSTIC MODEL USING MUILTI-SPACE GAUSSIAN MIXTURE MODEL

摘要

Disclosed is a method for generating an acoustic model using a multi-acoustic space Gaussian mixture model (GMM), comprising collecting a small amount of voice database; calculating statistical information from n number of existing acoustic models using the voice database; determining an optimal model unit on the basis of the calculated statistical information and a tying rule of the n number of existing acoustic models; mixing each of the n number of existing acoustic models with 1 GMM to generate n number of n GMMs increased in a Gaussian mixture number; generating a final GMM by applying the optimal model unit to a result of combining the n number of n GMMs; and calculating an optimal weight among final GMMs on the basis of a result of discrimination learning of a final acoustic model including the final GMM as each model unit.