A new framework named FlowEdit addresses a critical limitation in current flow-matching text-to-speech (TTS) systems by enabling lifelong pronunciation adaptation without requiring full model retraining1. While these advanced TTS models deliver high-quality zero-shot audio, they typically remain static post-deployment, struggling with accurate pronunciation of out-of-vocabulary proper nouns and other complex terms. Persistent errors necessitate costly and time-consuming model retraining to correct. FlowEdit circumvents this by learning pronunciation corrections through latent conditioning edits, rather than modifying the underlying model weights. This approach allows for dynamic, continuous improvement of pronunciation accuracy in a frozen TTS system. Published on arXiv AI on June 18, 2026, this research indicates a step towards highly adaptable synthetic voice generation. For practitioners and informed readers, the capability of AI to seamlessly adapt pronunciation without extensive retraining signals enhanced realism in voice synthesis, which could amplify the sophistication of deepfake audio and voice impersonation attacks, raising concerns for information integrity and geopolitical stability if leveraged by state-aligned threat actors in influence operations.