eduzhai > Applied Sciences > Engineering >

Deep Graph Random Process for Relational-Thinking-Based Speech Recognition

  • king
  • (0) Download
  • 20210505
  • Save

... pages left unread,continue reading

Document pages: 16 pages

Abstract: Lying at the core of human intelligence, relational thinking is characterizedby initially relying on innumerable unconscious percepts pertaining torelations between new sensory signals and prior knowledge, consequentlybecoming a recognizable concept or object through coupling and transformationof these percepts. Such mental processes are difficult to model in real-worldproblems such as in conversational automatic speech recognition (ASR), as thepercepts (if they are modelled as graphs indicating relationships amongutterances) are supposed to be innumerable and not directly observable. In thispaper, we present a Bayesian nonparametric deep learning method called deepgraph random process (DGP) that can generate an infinite number ofprobabilistic graphs representing percepts. We further provide a closed-formsolution for coupling and transformation of these percept graphs for acousticmodeling. Our approach is able to successfully infer relations among utteranceswithout using any relational data during training. Experimental evaluations onASR tasks including CHiME-2 and CHiME-5 demonstrate the effectiveness andbenefits of our method.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×