eduzhai > Applied Sciences > Engineering >

Semi-Supervised Learning with Data Augmentation for End-to-End ASR

  • king
  • (0) Download
  • 20210505
  • Save

... pages left unread,continue reading

Document pages: 5 pages

Abstract: In this paper, we apply Semi-Supervised Learning (SSL) along with DataAugmentation (DA) for improving the accuracy of End-to-End ASR. We focus on theconsistency regularization principle, which has been successfully applied toimage classification tasks, and present sequence-to-sequence (seq2seq) versionsof the FixMatch and Noisy Student algorithms. Specifically, we generate thepseudo labels for the unlabeled data on-the-fly with a seq2seq model afterperturbing the input features with DA. We also propose soft label variants ofboth algorithms to cope with pseudo label errors, showing further performanceimprovements. We conduct SSL experiments on a conversational speech data setwith 1.9kh manually transcribed training data, using only 25 of the originallabels (475h labeled data). In the result, the Noisy Student algorithm withsoft labels and consistency regularization achieves 10.4 word error rate (WER)reduction when adding 475h of unlabeled data, corresponding to a recovery rateof 92 . Furthermore, when iteratively adding 950h more unlabeled data, our bestSSL performance is within 5 WER increase compared to using the full labeledtraining set (recovery rate: 78 ).

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×