eduzhai > Applied Sciences > Engineering >

Sequence to Multi-Sequence Learning via Conditional Chain Mapping for Mixture Signals

  • Save

... pages left unread,continue reading

Document pages: 15 pages

Abstract: Neural sequence-to-sequence models are well established for applicationswhich can be cast as mapping a single input sequence into a single outputsequence. In this work, we focus on one-to-many sequence transduction problems,such as extracting multiple sequential sources from a mixture sequence. Weextend the standard sequence-to-sequence model to a conditional multi-sequencemodel, which explicitly models the relevance between multiple output sequenceswith the probabilistic chain rule. Based on this extension, our model canconditionally infer output sequences one-by-one by making use of both input andpreviously-estimated contextual output sequences. This model additionally has asimple and efficient stop criterion for the end of the transduction, making itable to infer the variable number of output sequences. We take speech data as aprimary test field to evaluate our methods since the observed speech data isoften composed of multiple sources due to the nature of the superpositionprinciple of sound waves. Experiments on several different tasks includingspeech separation and multi-speaker speech recognition show that ourconditional multi-sequence models lead to consistent improvements over theconventional non-conditional models.

Please select stars to rate!


0 comments Sign in to leave a comment.

    Data loading, please wait...