Learning Problem-agnostic Speech Representations from Multiple Self-supervised Tasks

6 Apr 2019Santiago PascualMirco RavanelliJoan SerràAntonio BonafonteYoshua Bengio

Learning good representations without supervision is still an open issue in machine learning, and is particularly challenging for speech signals, which are often characterized by long sequences with a complex hierarchical structure. Some recent works, however, have shown that it is possible to derive useful speech representations by employing a self-supervised encoder-discriminator approach... (read more)

PDF Abstract

Evaluation Results from the Paper


TASK DATASET MODEL METRIC NAME METRIC VALUE GLOBAL RANK COMPARE
Distant Speech Recognition DIRHA English WSJ PASE-FineTuned Word Error Rate (WER) 29.8 # 2