NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·2h ago
AI Research
arXiv CS.AI·Jun 10·5 min read
Automated Pronunciation Evaluation for Korean Toddler Speech using Speech Diarization and Self-Supervised Learning
TL;DR
Korean pediatric speech disorders lack effective automated assessment tools. An end-to-end pipeline was developed using neural speaker diarization and self-supervised learning for pronunciation evaluation.
✦ Why It Matters
Engineers can leverage this automated evaluation pipeline to improve speech assessment tools for young children.
Key Takeaways
- Explore neural speaker diarization for speech analysis.
- Utilize self-supervised learning for pronunciation scoring.
- Consider the implications of automated tools in pediatric care.
- Investigate the effectiveness of cross-model ensembles in speech tasks.
Full Summary
Speech sound disorders are prevalent in Korean children, affecting nearly half of pediatric communication cases. To address this, a novel automated pronunciation evaluation pipeline was created, integrating neural speaker diarization—specifically the NeMo SortFormer model—and self-supervised learning techniques.
A new corpus of 53 recordings from children aged 2-5 was annotated, resulting in 1,190 consonant and 748 vowel correctness labels. The NeMo SortFormer model achieved 88.69% speaker count accuracy and a 33.04% diarization error rate, effectively distinguishing between caregiver and toddler speech.
For pronunciation scoring, a cross-model ensemble was employed, routing consonant predictions to HuBERT-large and vowel predictions to WavLM-large, yielding balanced accuracies of 0.720 and 0.845, respectively. These findings indicate significant potential for automated tools in assessing toddler speech, which could enhance early intervention strategies.