Evaluating Textual and Acoustic De-Identification for Code-Switched Indonesian Speech

Abstract

Sharing sensitive speech corpora demands deidentification beyond text redaction: the voice itself is a biometric identifier, so redacting a transcript does not anonymize the recording. We evaluate a two-layer pipeline on an 18.7-hour Indonesian interview corpus on gender bias spanning eleven ethnicities. Layer 1 redacts textual identifiers and culturally specific quasi-identifiers (clan names, toponyms, kinship terms) over diarized transcripts, comparing newsdomain named-entity recognition (NER) with a locally run large language model (LLM); Layer 2 applies McAdams-coefficient speaker anonymization as a lightweight, lower-bound baseline. We evaluate the layers separately, leaving combined end-to-end risk to future work. In a four-ethnicity pilot, the LLM outperforms NER at similarly low precision (strict-match recall 66.3% versus 43.9%), with the largest gains on cultural quasiidentifiers, though both lose ground to ASR-induced name fragmentation. Sweeping the McAdams coefficient traces a privacy-utility curve: at our operating point it roughly doubles the attacker equal error rate (EER, 8.0% to 17.0%) at 17.1% word error rate, giving only partial protection. We frame the privacy-utility-gender tension as a conceptual contribution, not an empirical result. Release is controlled.