SpecAugment; A Simple Data Augmentation Method for Automatic Speech Recognition

SpecAugment: A Simple Data Augmentation Method for Automatic Speech Recognition (arxiv, 2019)

https://arxiv.org/pdf/2103.06695.pdf

Contents

1. Abstract
2. Introduction
3. Augmentation Policy

Abstract

SpecAugment,

• simple DA for speech recognition
• applied directly to the feature inputs of a NN (i.e., filter bank coefficients)
• consists of ..
  • warping the features
  • masking blocks of frequency channels
  • masking blocks of time steps

1. Introduction

Data augmentation for ASR

( ASR = Automatic Speech Recognition )

• Vocal Tract Length Normalization [11]
• Synthesize noisy audio [12]
• Speed perturbation for LVSCR tasks in [13]
• Use of an acoustic room simulator [14]
• Data augmentation for keyword spotting in [15, 16]
• Feature drop-outs for training multi-stream ASR systems [17]

SpecAugment

• operates on the log mel spectrogram of the input audio

  ( rather than the raw audio itself )

• simple & computationally cheap

• consists of three kinds of deformations of the log mel spectrogram

  • (1) Time warping
    • a deformation of the TS in the time direction
  • (2) Time masking
  • (3) Frequency masking

2. Augmentation Policy

• individual augmentations applied to a single input
• log mel spectrograms are normalized ( = zero mean )
  • \(\therefore\) imputing masked values to zer o = setting it to mean value

• can apply multiple masks