Clinical conversations embed identity both in spoken identifiers and in the speaker's voice, while pauses, delivery style, and role shifts also convey valuable clinical cues. Protecting such speech therefore requires more than word‑level anonymization. AegisVox introduces a factorized framework that assigns a consistent surrogate voice to each document and replaces identifier‑bearing turns with document‑consistent alternatives, preserving contextual coherence.
We evaluated the system on 390 public simulated or non‑patient dialogues (totaling 63.6 hours). Two speech renderers were compared: a text‑to‑speech (TTS) pipeline and a codec‑based speech editor. The TTS renderer realized 535 out of 601 planned replacements, whereas the codec editor achieved 453/601. Although aggregate clinical preservation appeared similar, the TTS approach exhibited more model‑labeled contradictions, and conversation‑wide speaker linkage scores revealed stronger edited‑speech connections under a score‑direction sensitivity analysis.
Fresh transcription of the original audio served as a reference for interpreting the clinical measurement chain, and an independent judge assessed sensitivity to model‑based evaluation. The findings show that higher replacement realization can coexist with poorer clinical or speaker‑linkage metrics. AegisVox therefore supports use‑specific trade‑off evaluation rather than guaranteeing complete identifier removal or clinical safety.
Review