SpecAugment : A Simple Data Augmentation Method for Automatic Speech Recognition

https://arxiv.org/pdf/1904.08779.pdf https://ai.googleblog.com/2019/04/specaugment-new-data-augmentation.html https://www.notion.so/SpecAugment-A-Simple-Data-Augmentation-Method-for-Automatic-Speech-Recognition-ece0a0be49844c7c93ab4f0045ee1562 https://github.com/WindQAQ/listen-attend-and-spell https://github.com/DemisEom/SpecAugment Abstract 음성 인식을 위한 간단한 data augmentation 방법 제안 SpecAugment  log mel spectrogram (이미지 처럼 다룸)을 input으로 3가지 방법으로 data augmentation Time warping Frequency masking Time masking 음성인식 네트워크로 Listen, Attend and Spell을 사용 LibriSpeech 960h,  SwitchBoard 300h dataset 대상으로 성능 개선, Language Model 과 shallow fusion 시 WER 좀 더 좋음 […]

Read More