Tweeted By @zacharylipton

Simple technique: When performing data augmentation, don't just coerce same *label*, coerce the same *representation*. Significant gains (CER) both in-sample and on out-of-domain noisy speech. Joint work w Davis Liang and Zhiheng Huang at @awscloud https://t.co/cixxqyopjF

— Zachary Lipton (@zacharylipton) July 19, 2018