Acoustic Modeling for Under-resourced Language using Mismatched Transcriptions

Mismatched crowdsourcing is a technique to derive speech transcriptions using crowdworkers unfamiliar with the language being spoken. This technique is especially useful for under-resourced languages since it is hard to hire native transcribers. In this paper, we demonstrate that using mismatched transcription for adaptation improves performance of speech recognition under limited matched training data conditions. We show that using previously published methods for training data augmentation improves the utility of mismatched transcription. Finally, we show that a mismatched transcription can be used to train one neural network in two forms, in two sequential steps: first as a probabilistic transcription, and second as the auxiliary task of a multi-task learner.