Speech recognition using deep neural networks trained with non-uniform frame-level cost functions

Abstract

The aim of this paper is to present two new variations of the frame-level cost function for training a Deep neural network in order to achieve better word error rates in speech recognition. Minimization functions of a neural network are salient aspects to deal with when researchers are working on machine learning, and hence their improvement is a process of constant evolution. In the first proposed method, the conventional cross-entropy function can be mapped to a nonuniform loss function based on its corresponding extropy (a complementary dual function), enhancing the frames that have ambiguity in their belonging to specific senones (tied-triphone states in a hidden Markov model). The second proposition is a fusion of the proposed mapped cross-entropy and the boosted cross-entropy function, which emphasizes those frames with low target posterior probability. The developed approaches have been performed by using a personalized mid-vocabulary speaker-independent voice corpus. This dataset is employed for recognition of digit strings and personal name lists in Spanish from the northern central part of Mexico on a connected-words phone dialing task. A relative word error rate improvement of 12.3% and 10.7% is obtained with the two proposed approaches, respectively, regarding the conventional well-established crossentropy objective function.