Dysarthric Speech Recognition Using Curriculum Learning and Articulatory Feature Embedding

Recognizing speech in individuals with articulation disorders is a challenging task due to limited resources and diverse speaker characteristics. Domain adaptation is commonly employed to address these issues, and in this paper, we apply curriculum learning, a method within this approach, to Automatic Speech Recognition (ASR). To enhance the efficiency of curriculum learning, we reorganize the dataset. Additionally, we incorporate speaker and articulatory features to capture the pronunciation characteristics of patients. Experimental results demonstrate that our proposed method achieves an 11.37% improvement compared to the baseline.