Abstract

Low-resource ASR self-training seeks to minimize resource requirements such as manual transcriptions or language modeling text. This is accomplished by training on large quantities of audio automatically labeled by a small initial model. By analyzing our previous experiments with the conversational telephone English Fisher corpus, we demonstrate where self-training succeeds and under what resource conditions it provides the most benefit. Additionally, we will show success on Spanish and Levantine conversational speech as well as the tougher English Callhome set, despite initial WER of more than 60%. Finally, by digging beneath average word error rate and analyzing individual word performance, we show that self-trained models successfully learn new words. More importantly, self-training benefits most words which appear in the unlabeled audio but do not appear in the manual transcriptions.