Multi-Model ASR Integration With Reliability Weighting for Automated Speech Disorder Screening

초록

Speech sound disorders affect communication development in children, requiring early detection for timely intervention. Traditional clinical assessment relies on manual phonetic transcription and calculation of Percent Consonants Correct, a labor-intensive process limiting accessibility in underserved regions. We present an automated system for Korean child speech sound disorder screening based on standardized picture-naming tasks, using multi-model automatic speech recognition integration with reliability-weighted machine learning. Our approach fine-tunes four state-of-the-art models with multiple independent training runs, employing intra-model ensemble methods to generate robust transcriptions. We then train gradient boosting classifiers using feature vectors that combine model-specific Percent Consonants Correct predictions with reliability indicators capturing prediction uncertainty and inter-model agreement. This allows us to predict human-annotated consonant accuracy and classify children as typically developing or speech-sound-disordered based on age-stratified normative thresholds. Evaluated on 572 Korean children aged 2.5-9 years with 21,878 utterances across 5-fold speaker-stratified cross-validation, our system achieves 82.4% unweighted average recall, 86.9% accuracy, and 0.759 F1-score, improving upon the best single-model performance by 5.3 percentage points in unweighted average recall. Ablation studies confirm that multi-model integration and reliability-based weighting are both critical for accurate consonant accuracy prediction. This work demonstrates that imperfect automatic speech recognition, when combined with ensemble-based uncertainty estimation and multi-model integration, can achieve reliable speech disorder screening, potentially expanding access to diagnostic services in resource-limited settings.

키워드