International Conference on Computer and Knowledge Engineering

Home / 15th International Conference on Computer and Knowledge Engineering

Adaptive Pronunciation Scoring: Aligning Automated Assessments with Human Expert Evaluations

Authors :

Omid Aghdaei¹ Mohammad Sadegh Safari² Mohammad Hassan Rasoolizadeh³ Abedeh Mirzaee⁴

1- Part AI Research Center 2- Part AI Research Center 3- Part AI Research Center 4- Part AI Research Center

Keywords :

Pronunciation evaluation،adaptive scoring،speech assessment،Computer-assisted Pronunciation Training (CAPT،phonetic error analysis

Abstract :

Conventional pronunciation evaluation systems aggregate metrics such as accuracy, fluency, completeness, and prosody into a composite score. These systems apply uniform penalization when assessing accuracy, neglecting the linguistic and perceptual significance of different errors. Consequently, minor phonetic variations are penalized as severely as critical intelligibility errors, leading to inaccurate assessments. This paper introduces an adaptive scoring framework that adjusts penalty weights based on the severity and linguistic relevance of errors. By categorizing errors into contextual, structural, phonetic, and alignment types, the framework enables more precise, context-sensitive evaluation. The approach is evaluated on the Speechocean762 and proprietary datasets. Experimental results show stronger alignment with expert judgments compared to traditional models, improving the reliability of Computer-Assisted Pronunciation Training (CAPT) and automated speech evaluation.