FlowEdit: Associative Memory for Lifelong Pronunciation Adaptation in Flow-Matching TTS
arXiv·medium signal
Singh et al. note that flow-matching TTS achieves strong zero-shot quality but stays static after deployment, repeating pronunciation errors on out-of-distribution words (new names, jargon). FlowEdit adds an associative-memory mechanism enabling lifelong pronunciation correction without full retraining. Practical for production voice systems that must continually learn new terms.