PROJECT TITLE:

Multitask Learning of Deep Neural Networks for Low-Resource Speech Recognition - 2015

ABSTRACT:

We propose a multitask learning (MTL) approach to enhance low-resource automatic speech recognition using deep neural networks (DNNs) while not requiring further language resources. We 1st demonstrate that the performance of the phone models of one low-resource language can be improved by coaching its grapheme models in parallel beneath the MTL framework. If multiple low-resource languages are trained along, we investigate learning a collection of universal phones (UPS) as a further task again within the MTL framework to boost the performance of the phone models of all the involved languages. In both cases, the heuristic guideline is to select a task that will exploit further data from the coaching information of the primary task(s). In the primary technique, the additional information is that the phone-to-grapheme mappings, whereas in the second method, the UPS helps to implicitly map the phones of the multiple languages among each different. In an exceedingly series of experiments using 3 low-resource South African languages within the Lwazi corpus, the proposed MTL strategies obtain important word recognition gains compared with single-task learning (STL) of the corresponding DNNs or ROVER that mixes results from many STL-trained DNNs.