Deep Learning-Based Multi-Aspect Pronunciation Assessment for Individuals with Down Syndrome

Abstract

This paper explores the use of an annotated speech corpus to assess multiple dimensions of speech quality—particularly phonetic, fluency and prosody—in individuals with Down syndrome, with the aim of informing the development of automated assessment tools. We conducted a series of experiments using the GOPT model, together with representations extracted from fine-tuning Wav2Vec models focused on phoneme classification. Model predictions were compared against expert annotations from a speech-language pathologist using Pearson correlation. Results demonstrate significant improvements over prior work, with correlations up to 0.49 in certain aspects, particularly for phonetic and fluency dimensions, while prosody remained more challenging to model. The study highlights the potential of Transformer-based architectures for atypical speech assessment and underscores the challenges inherent in assessing atypical speech, particularly due to variability linked to specific disfluency types.

Resources

Details