INTUITIONISTIC FUZZY SCORING FOR FLUENCY IN TELEPRACTICE SESSIONS

This study proposes an uncertainty-aware framework for remote speech evaluation that represents each segment with a triple of evidence-for, evidence-against, and hesitation. Segment features (e.g., articulation rate, pause ratio) are first direction-aligned and passed through slope-constrained sigmoids; a reliability-sensitive kernel driven by packet loss and signal-to-noise ratio allocates hesitation so that ambiguous or noisy conditions do not force brittle decisions. Feature-level triples are combined by weighted averaging (and an ordered variant), ensuring closure and monotonicity, and reduced to a scalar decision index ​ and a companion confidence C. Multi-rater information is fused by estimating rater reliabilities via expectation–maximization or graded-response modeling, with disagreement promoted to additional hesitation. We provide stability guarantees (Lipschitz bounds) and a complete, executable case study on eight segments. The case study shows clear separation between positive and negative outcomes (e.g., a strongly negative example with , and a clearly positive example with ), predictable response to degraded channels (−5 dB or +0.10 packet loss), and negligible rank shifts when switching from standard to ordered averaging. A practical triage band isolates borderline items and surfaces low-confidence cases for review. The approach is simple to audit, fast enough for near–real-time feedback, and designed to integrate into clinical and educational workflows as decision support rather than diagnosis. This study proposes an uncertainty-aware framework for remote speech evaluation that represents each segment with a triple of evidence-for, evidence-against, and hesitation. Segment features (e.g., articulation rate, pause ratio) are first direction-aligned and passed through slope-constrained sigmoids; a reliability-sensitive kernel driven by packet loss and signal-to-noise ratio allocates hesitation so that ambiguous or noisy conditions do not force brittle decisions. Feature-level triples are combined by weighted averaging (and an ordered variant), ensuring closure and monotonicity, and reduced to a scalar decision index ​ and a companion confidence C. Multi-rater information is fused by estimating rater reliabilities via expectation–maximization or graded-response modeling, with disagreement promoted to additional hesitation. We provide stability guarantees (Lipschitz bounds) and a complete, executable case study on eight segments. The case study shows clear separation between positive and negative outcomes (e.g., a strongly negative example with , and a clearly positive example with ), predictable response to degraded channels (−5 dB or +0.10 packet loss), and negligible rank shifts when switching from standard to ordered averaging. A practical triage band isolates borderline items and surfaces low-confidence cases for review. The approach is simple to audit, fast enough for near–real-time feedback, and designed to integrate into clinical and educational workflows as decision support rather than diagnosis.