Brief

Researchers have introduced MoDiCoL, a new dataset designed to improve the robustness of Automatic Speech Recognition (ASR) systems. Unlike existing datasets that isolate factors like accents or noise, MoDiCoL allows for the controlled analysis of linguistic content, speaker characteristics, and acoustic environments, including their co-occurrence. The dataset is paired with a continual learning curriculum to simulate real-world incremental updates and study how ASR models acquire, transfer, and forget robustness under evolving conditions. AI

Researchers have developed a new approach to improve Automatic Speech Recognition (ASR) systems by incorporating continual learning techniques to better handle disfluent speech. The method involves introducing explicit disfluency tokens into pre-trained ASR models and then fine-tuning them on diverse datasets. This process aims to prevent catastrophic forgetting of general knowledge while enhancing the model's ability to recognize and process speech disfluencies, addressing a key challenge in current ASR technology. AI