Leveraging Phonemic Transcription and Whisper toward Clinically Significant Indices for Automatic Child Speech Assessment

Diagnosing speech sound disorders (SSD) in children requires professional assessment by speech-language pathologists. Detecting and diagnosing a medical condition takes time and is usually expensive in terms of labor. However, early identification and treatment are essential for subsequent care. ASR-based child speech assessment prioritizes semantic understanding over phonetic accuracy, making it unsuitable for pronunciation assessment. This study uses phonemic transcription available in a normative dataset and utilizes pre-trained speech models to develop an automatic phoneme recognition model with a Phoneme Error Rate (PER) as low as 3.76%. Clinically relevant indices calculated from the model prediction are highly correlated with those from the original normative data. We regard these experimental results as solid evidence that validates the feasibility of our evaluation workflow for practical application in early screening for phonological development delays.