SpecAugment.md

SpecAugment

Questions

• SpecAugment is suitable for traditional asr ?
• What does time-warping do ?

Key Concepts

• warping the features(time warp)
• masking blocks of frequency
• masking blocks of time steps
• LAS(listen, attend and spell)
• Similiar to unsupervised pre-training
• Label smoothing

Methods

• Vocal Tract Length Normalization(VTLN)
• Noisy Data Synthesised
• Speed Perturbation
• Acoustic Room Simulator

Reference Repo

• SpecAugment
• las
• clova-speech-hackathon
• sparse_image_warp_pytorch
• espnet.transform.spec_augment

Reference Blog

• SpecAugment

TODO

• Reference paper
• Implement SpecAugment with numpy