INTERSPEECH 2009

Using the central observation that marginbased weighted classification error (modeled using Minimum Phone Error (MPE)) corresponds to the derivative with respect to the margin term of marginbased hinge loss (modeled using Maximum Mutual Information (MMI)), this article subsumes and extends marginbased MPE and MMI within a broader framework in which the objective function is an integral of MPE loss over a range of margin values. Applying the Fundamental Theorem of Calculus, this integral is easily evaluated using finite differences of MMI functionals; latticebased training using the new criterion can then be carried out using differences of MMI gradients. Experimental results comparing the new framework with marginbased MMI, MCE and MPE on the Corpus of Spontaneous Japanese and the MIT OpenCourseWare/MITWorld corpus are presented.
Bibliographic reference. McDermott, Erik / Watanabe, Shinji / Nakamura, Atsushi (2009): "Marginspace integration of MPE loss via differencing of MMI functionals for generalized errorweighted discriminative training", In INTERSPEECH2009, 224227.