Improving deep neural network acoustic modeling for audio corpus indexing under The IARPA Babel program

Abstract

This paper is focused on several techniques that improve deep neural network (DNN) acoustic modeling for audio corpus indexing in the context of the IARPA Babel program. Specifically, fundamental frequency variation (FFV) and channelaware (CA) features and data augmentation based on stochastic feature mapping (SFM) are investigated not only for improved automatic speech recognition (ASR) performance but also for their impact to the final spoken term detection on the pre-indexed audio corpus. Experimental results on development languages of Babel option period one show that the improved DNN acoustic models can reduce word error rates in ASR and also help the keyword search performance compared to already competitive DNN baseline systems.