SpecAugment++; A Hidden Space Data Augmentation Method for Acoustic Scene Classification

SpecAugment++: A Hidden Space Data Augmentation Method for Acoustic Scene Classification (Interspeech, 2021)

https://arxiv.org/pdf/2103.16858.pdf

Contents

1. Abstract

Abstract

SpecAugment++

• a novel DA for acoustic scene classification (ASC)

• SpecAugment, mixup : only work on the input space

• SpecAugment++ : applied to both the input space and the hidden space

• hidden state) consist of …
  • masking blocks of frequency channels
  • masking blocks of time frames

• Imputing masked values

  • previous) zero ( = ZM )

  • proposed) 2 approaches

    • based on the use of other samples within the minibatch

      • a) mini-batch based mixture masking ( = MM )
      • b) mini-batch based cutting masking ( = CM )

    • can be seen as introducing additional noises generated from the dataset

      & guide the networks to be more discriminative for classification

• Experimental results

  • DCASE 2018 Task1 dataset … 3.6% gain
  • DCASE 2019 Task1 dataset … 4.7% gain

1. Introduction

Different from the mixup [15] and BC learning [17] …

\(\rightarrow\) Labels of the augmented data are not changed!

2. SpecAugment++

3. Experiments