Reproduce “Towards few-shot isolated word reading assessment” on OpenResearch

About this paper

Reuben Smit, Retief Louw, Herman Kamper

Researchers at Stellenbosch University developed an ASR-free, few-shot method for isolated word reading assessment of child speech in Afrikaans, which compares child utterances to a small set of adult templates using self-supervised learning features and distance metrics. The approach achieved a balanced accuracy of 65.3% for child-adult matching, highlighting the inherent challenge of acoustic mismatch between child and adult speech for current general-purpose SSL models.

Key insights

• While self-supervised learning (SSL) representations are highly effective for reading assessment when acoustic conditions are consistent (adult speech vs. adult templates), their performance consistently degrades when applied to child speech inputs compared to adult templates due to the high acoustic variability of child speech.
• Current general-purpose SSL models, despite their robustness in zero-resource and cross-lingual adult speech scenarios, are not robust enough to bridge the acoustic gap between adult-provided templates and child speech inputs, indicating a need for child-speech-specific SSL models.
• The choice of evaluation metric is critical for practical reading assessment; balanced accuracy provides a more realistic indicator for practical deployment than AUC or F1 when a single threshold must be used across all word classes.

Summary by alphaXiv.