Deep Auto-encoder Based Multi-task Learning Using Probabilistic Transcriptions

• http://www.isca-speech.org/archive/Interspeech_2017/pdfs/0582.PDF
• http://www.fit.vutbr.cz/research/groups/speech/publi/2017/das_interspeech2017_IS170582.pdf PDF

cross-lingual speech recognition, probabilistic transcription, deep neural networks, multi-task learning

This article is about deep auto-encoder based Multi-task Learning using probabilistic transcriptions.

We examine a scenario where we have no access to native transcribers in the target language. This is typical of language communities that are under-resourced. However, turkers (online crowd workers) available in online marketplaces can serve as valuable alternative resources for providing transcripts in the target language. We assume that the turkers neither speak nor have any familiarity with the target language. Thus, they are unable to distinguish all phone pairs in the target language; their transcripts therefore specify, at best, a probability distribution called a probabilistic transcript (PT). Standard deep neural network (DNN) training using PTs do not necessarily improve error rates. Previously reported results have demonstrated some success by adopting the multi-task learning (MTL) approach. In this study, we report further improvements by introducing a deep auto-encoder based MTL. This method leverages large amounts of untranscribed data in the target language in addition to the PTs obtained from turkers. Furthermore, to encourage transfer learning in the feature space, we also examine the effect of using monophones from transcripts in well-resourced languages. We report consistent improvement in phone error rates (PER) for Swahili, Amharic, Dinka, and Mandarin.