UNSUPERVISED HMM ADAPTATION BASED ON SPEECH-SILENCE DISCRIMINATION

摘要

An unsupervised, discriminative, sentence level, HMMadaptation based on speech-silence classification ispresented. Silence and speech regions are determined eitherusing a speech end-pointer or the segmentation obtained fromthe recognizer in a first pass. The discriminative trainingprocedure using a GPD or any other discriminative trainingalgorithm, employed in conjunction with the HMM-basedrecognizer, is then used to increase the discriminationbetween silence and speech.