Abstract

We propose a novel neural network architecture based on conformer transducerthat adds contextual information flow to the ASR systems. Our method improvesthe accuracy of recognizing uncommon words while not harming the word errorrate of regular words. We explore the uncommon words accuracy improvement whenwe use the new model and/or shallow fusion with context language model. Wefound that combination of both provides cumulative gain in uncommon wordsrecognition accuracy.

 

Quick Read (beta)