SpecMix : A Mixed Sample Data Augmentation Method for Training with Time-Frequency Domain Features

• DOI: 10.21437/interspeech.2021-103
• Publication Date: 2021-08-27T05:59:39Z
• Authors: Gwantae Kim; David K. Han; Hanseok Ko
• Keywords: Sample (material); Training set

A mixed sample data augmentation strategy is proposed to enhance the performance of models on audio scene classification, sound event and speech enhancement tasks.While there have been several methods shown be effective in improving image classification performance, their efficacy toward time-frequency domain features not assured.We propose a novel approach named "Specmix" specifically designed for dealing with features.The method consists mixing two different samples by applying masks preserving spectral correlation each sample.Our experiments acoustic tasks show that Specmix improves various neural network architectures maximum 2.7%.