Usefulness of Automatic Speech Recognition Assessment of Children With Speech Sound Disorders: Validation Study (Preprint)

• DOI: 10.2196/preprints.60520
• Publication Date: 2025-01-14T19:00:30Z
• Authors: Do Hyung Kim; Joo Won Jeong; Dayoung Kang; Taekyung Ahn; Yeonjung Hong; Younggon Im; Jaewon Kim; Min Jung Kim; Dae-Hyun Jang
• Keywords: Preprint; Speech sound

<sec> <title>BACKGROUND</title> Speech sound disorders (SSDs) are common communication challenges in children, typically assessed by speech-language pathologists (SLPs) using standardized tools. However, traditional evaluation methods time-intensive and prone to variability, raising concerns about reliability. </sec> <title>OBJECTIVE</title> This study aimed compare the outcomes of SLPs an automatic speech recognition (ASR) model two SSD assessments South Korea, evaluating ASR model’s performance. <title>METHODS</title> A fine-tuned wav2vec 2.0 XLS-R model, pretrained on 436,000 hours adult voice data spanning 128 languages, was used. The further trained 93.6 minutes children’s voices with articulation errors improve error detection. Participants included children referred Department Rehabilitation Medicine at a general hospital Incheon, from August 19, 2022, June 14, 2023. Two assessments—the Assessment Phonology Articulation for Children (APAC) Urimal Test (U-TAP)—were used, transcriptions compared SLP transcriptions. <title>RESULTS</title> 30 aged 3-7 years who were suspected having SSDs. phoneme rates APAC U-TAP 8.42% (457/5430) 8.91% (402/4514), respectively, indicating discrepancies between across all phonemes. Consonant 10.58% (327/3090) 11.86% (331/2790) U-TAP, respectively. On average, there 2.60 (SD 1.54) 3.07 1.39) per child correctly produced phonemes, 7.87 3.66) 7.57 4.85) incorrectly based correlation terms percentage consonants correct excellent, intraclass coefficient 0.984 (95% CI 0.953-0.994) 0.978 0.941-0.990) UTAP, &lt;i&gt;z&lt;/i&gt; scores showed more pronounced differences than 8 individuals showing 2 U-TAP. <title>CONCLUSIONS</title> results demonstrate potential assessing its performance varied or word characteristics, highlighting areas refinement. Future research should include diverse samples, clinical settings, strengthen refinement ensure broader applicability.