performing front-end processing on an acoustic input signal, wherein the front-end processing generates MEL frequency cepstral features representative of the acoustic input signal;

performing recognition and adaptation by:

providing the MEL frequency cepstral features to a speech recognizer, wherein the speech recognizer utilizes the MEL frequency cepstral features and a current list of acoustic training models to determine at least one best hypothesis;

receiving, from the speech recognizer, at least one best hypothesis, associated acoustic training models, and associated probabilities;

computing a pre-adaptation acoustic score by recognizing an utterance using the associated acoustic training models;

choosing acoustic training models from the associated acoustic training models;

performing adaptation on the chosen associated acoustic training models;

computing a post-adaptation acoustic score by recognizing the utterance using the adapted acoustic training models;

comparing the pre-adaptation acoustic score with the post-adaptation acoustic score to check for improvement; modifying the current list of acoustic training models to include the adapted acoustic training models, if the acoustic score improved after performing adaptation; and performing recognition and adaptation iteratively until the acoustic score ceases to improve;

choosing the best hypothesis as recognized words once the acoustic score ceases to improve; and

outputting the recognized words.