Abstract
Humans can quickly adapt to recognize acoustically degraded speech, and here we hypothesize that the quick adaptation is enabled by internal linguistic feedback - Listeners use partially recognized sentences to adapt the mapping between acoustic features and phonetic labels. We test this hypothesis by quantifying how quickly humans adapt to degraded speech and analyzing whether the adaptation process can be simulated by adapting an automatic speech recognition (ASR) system based on its own speech recognition results. We consider three types of acoustic degradation, i.e., noise vocoding, time compression, and local time-reversal. The human speech recognition rate can increase by >20% after exposure to just a few acoustically degraded sentences. Critically, the ASR system with internal linguistic feedback can adapt to degraded speech with human-level speed and accuracy. These results suggest that self-supervised learning based on linguistic feedback is a plausible strategy for human adaptation to acoustically degraded speech.