Discriminative Training of Acoustic Models - From MPE to fMPE

Discriminative Training of Acoustic Models - From MPE to fMPE

Learn about discriminative training of acoustic models in this 37-minute lecture from IBM researcher Daniel Povey at Johns Hopkins University's Center for Language & Speech Processing. Gain insights into Maximum Mutual Information Estimation (MPE) techniques and discover fMPE, an innovative feature-space transformation method developed at IBM that maximizes the MPE criterion for acoustic model training.