AN AUTOMATIC SPEECH ASSESSMENT ALGORITHM ADAPTING TO THE PRONUNCIATION CHARACTERISTICS OF CONVICTED PERSONS

Abstract

A safety-constrained speaker-adaptive algorithm is proposed to separate stable acoustic speaker properties from actual foreign-language pronunciation competence. Two to six unlabeled enrollment utterances estimate baseline F0, timbre, habitual rate, and channel. A small residual adapter is restricted to this nuisance subspace, while phoneme accuracy, completeness, stress, and rhythm are protected through orthogonal projection and a trust region. A do-no-harm gate compares global and adapted scores. In a reproducible simulation of 960 evaluation samples, the proposed adapter achieved r=0.712, MAE=8.31, and QWK=0.668. No real incarcerated-person recordings were used; external ethical validation is required.