receiving, by an analog-to-digital converter, an input signal representing input speech;

converting the input speech to digital input speech by the analog-to-digital converter;

generating a set of multiple-mixture monophone models via a set of single-mixture monophone models;

training the set of multiple-mixture monophone models with a set of single-mixture monophone transcripts to generate a set of multiple-mixture context independent models;

generating a set of single-mixture triphone models;

training the set of single-mixture triphone models with a set of triphone transcripts to generate a set of context dependent models;

estimating parameters of the set of context dependent models for the digitized input speech via a data dependent maximum a posteriori (MAP) adaptation method, wherein parameters of tied states of the set of context dependent models are derived by adapting corresponding parameters of the set of multiple-mixture context independent models through the MAP adaptation method via training data associated with the corresponding tied states; and

outputting recognized speech, of the input speech, based on the estimated parameters.