DFA-Based Numeric Normalization and KcBERT-NER Hybrid Ensemble for PII Detection in Korean STT Transcripts

Speech-To-Text (STT) systems convert spoken utterances into textual transcripts for various speech-based applications.During this process, numeric information such as telephone numbers, resident registration numbers, bank account numbers, and credit card numbers may be transcribed as Korean Numeric Expressions (KNE) rather than standard Arabic Numeric Strings (ANS).This representation mismatch can cause conventional Named Entity Recognition (NER) models trained primarily on written text to fail to detect Personally Identifiable Information (PII), resulting in false negatives and potential privacy exposure.This study analyzes the causes of such detection failures.It proposes a hybrid ensemble framework combining Deterministic Finite Automaton (DFA)-based Korean numeric normalization, rule-based structured PII detection, and KcBERT-based NER.For controlled evaluation, ANS expressions contained in a public Korean PII text dataset were converted into KNE using an STT-like simulation procedure.The generated KNE expressions were subsequently normalized back into canonical ANS representations using the proposed DFA-based normalization method.Rule-based detection and KcBERT-based NER were then independently applied to the normalized input and combined their confidence scores through decision-level fusion.Experimental results showed that the proposed hybrid model achieved a recall of 1.0000 and a false negative rate (FNR) of 0.0000 on the test set, demonstrating its effectiveness in minimizing missed PII detections under the controlled experimental setting.