Can we enable speech data with Differential Privacy?

We study whether speech can be shared with formal privacy guarantees. We present a speaker anonymization mechanism with local differential privacy by perturbing speaker embeddings in Qwen3-TTS. Our analysis identifies practical operating points for differential privacy in speech, reveals worst-case acoustic identity leakage, and quantifies the impact of informed adversaries. We further propose content-aware privacy, a Pufferfish-based formulation that captures identity leakage through linguistic content and exposes the limitations of acoustic-only protection.