Disfluency Correction

Disfluency correction (DC) aims to remove hesitations, repetitions, and fillers from spoken language, improving the quality of transcribed speech for various applications. Current research focuses on developing robust DC models, employing techniques like reinforcement learning for personalized medication adjustments to mitigate speech disfluencies and leveraging large-scale datasets and advanced architectures such as sequence-tagging models and end-to-end systems incorporating pretrained acoustic language models to achieve high accuracy across multiple languages. The improved accuracy and multilingual capabilities of DC systems are significantly impacting downstream tasks like machine translation and speech-to-speech applications, particularly benefiting language learning and accessibility for individuals with speech impairments.