I N T E R N a T I O N a L C O M P U T E R S C I E N C E an Investigation of Tandem Mlp Features for Asr

This project explores speech feature representations produced by discriminatively trained multi-layer perceptrons. Previous research has demonstrated that such a tandem approach can be successfully exploited for large-vocabulary automatic speech recognition systems. The principal aim of this work is to empirically evaluate some variants of these features. While experimental results validate some of the design choices of the standard implementation, other evidence suggests alternatives that may improve performance. From this exploratory investigation, we hypothesize which of the various modifications are most promising; applied to a Mandarin broadcast news task, the new configuration demonstrates significant improvement. Along with the novel presentation of a " best-case scenario " and other cheating experiments, an interpretation of these results is discussed with the hope of guiding future directions of research. Abstract ii 1 Introduction