Opening the Black Box of wav2vec Feature Encoder

Kwanghee Choi, Eun Jung Yeo

Self-supervised models, namely, wav2vec and its variants, have shown promising results in various downstream tasks in the speech domain. However, their inner workings are poorly un…