eduzhai > Applied Sciences > Engineering >

Boosting Active Learning for Speech Recognition with Noisy Pseudo-labeled Samples

  • Save

... pages left unread,continue reading

Document pages: 8 pages

Abstract: The cost of annotating transcriptions for large speech corpora becomes abottleneck to maximally enjoy the potential capacity of deep neuralnetwork-based automatic speech recognition models. In this paper, we present anew training pipeline boosting the conventional active learning approachtargeting label-efficient learning to resolve the mentioned problem. Existingactive learning methods only focus on selecting a set of informative samplesunder a labeling budget. One step further, we suggest that the trainingefficiency can be further improved by utilizing the unlabeled samples,exceeding the labeling budget, by introducing sophisticatedly configuredunsupervised loss complementing supervised loss effectively. We propose newunsupervised loss based on consistency regularization, and we configureappropriate augmentation techniques for utterances to adopt consistencyregularization in the automatic speech recognition task. From the qualitativeand quantitative experiments on the real-world dataset and under real-usagescenarios, we show that the proposed training pipeline can boost the efficacyof active learning approaches, thus successfully reducing a sustainable amountof human labeling cost.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×