Skip to Content
Applied SciencesApplied Sciences
  • Article
  • Open Access

18 November 2026

53 Pages

An End-to-End Speaker Re-Identification System for Romanian Speech: Model Adaptation, Closed-Set and Open-Set Evaluation

,
and
Computer and Cyber Security Department, Faculty of Information Systems and Cyber Security, Military Technical Academy ‘Ferdinand I’, 050141 Bucharest, Romania
*
Author to whom correspondence should be addressed.

Abstract

This study evaluates ReDimNet2 and ECAPA-TDNN fine-tuned on Romanian Common Voice 22 and tests a deployed speaker re-identification pipeline. We compared six models for verification on Common Voice 22 and RoDigits and for closed-set and open-set identification on both corpora and ECHO. Across 30 single-enrollment splits per corpus, RO-ReDimNet2 achieved mean Rank-1 true positive identification rates (TPIRs) of 93.64%, 84.65%, and 90.20% on combined Common Voice 22, combined RoDigits, and screened ECHO, respectively, with mean false positive identification rates (FPIRs) near 1%. Its combined-set verification equal error rates were 3.56% on Common Voice 22 and 4.36% on RoDigits. Training with augmentation improved ECAPA-TDNN verification on degraded speech. Increasing enrollment from three to five recordings improved mean Rank-1 identification for both adapted models on matched probes. Threshold transfer across corpora substantially changed TPIR and FPIR, and voice activity detection reduced TPIR on combined probes at fixed thresholds. On nine continuous multi-speaker recordings from Romanian Parliamentary Speeches, 5-s windows yielded a pooled temporal F1 of 92.40% for RO-ReDimNet2 versus 87.67% for RO-ECAPA-TDNN. RO-ReDimNet2 achieved 97.72% window-level TPIR and 3.10% FPIR. At 100,000 enrollments, mean pgVector search latency stayed below 3.4 ms under sequential querying.

Article Metrics

Citations

Article Access Statistics

Multiple requests from the same IP address are counted as one view.