No abstract

No abstract

No abstract

Abstract. We propose a method for controlling Gaussian mixture splitting in HMM states during the training of acoustic models. The method is based on introducing special criteria of mixture quality in every state. These criteria are calculated over a separate part of the speech database. We back up states before splitting and revert to saved copies of the states whose criteria values have decreased, which makes it possible to optimize the number of Gaussians in the GMMs of the states and to prevent overfitting. The models obtained by such training demonstrate improved recognition rate with a significantly smaller number of Gaussians per state.

No abstract