Learning to Hear Hesitation: Continual Learning for Disfluency-Aware ASR
The study uses continual learning (CL) with explicit disfluency tokens to improve Automatic Speech Recognition (ASR) performance when dealing with variable speech disfluencies, preventing knowledge loss. By introducing tokens like 'um' or 'uh' and using methods like Weight Averaging, the researchers found a trade-off: better ASR accuracy versus better marker recognition. This suggests that learning markers requires parameter updates that can conflict with maintaining general speech recognition skills.