Learning to Hear Hesitation: Continual Learning for Disfluency-Aware ASR

In three linesNew continual learning approach to improve Automatic Speech Recognition (ASR) on disfluent speech. Researchers introduce explicit disfluency tokens into a pretrained ASR model, then continue training on datasets with varying disfluency distributions. Analysis reveals trade-off between marker learning and ASR performance.

Summary generated by Claude — human-verified