Domain adaptation for accented speech

Speaker: Robin San Roman

Data and place: March 31, 2022, at 10:30 – Hybrid

Abstract: Automatic speech recognition (ASR) systems tend to have a performance drop with accented speech compared to standard speech. This is due to linguistic differences between the different domains of speech. In this talk, I will present a method to adapt CPC models to new domains with a small amount of unlabeled out-of-domain data. I will also present ideas to move towards a speaker-invariant speech representation that allows easier generalization to new domains.