Tag
GRAFT is a per-word pronunciation conditioning mechanism for zero-shot text-to-speech that uses a spoken sample of a target word to control its pronunciation, achieving significant improvements in target-word phoneme error rates across multiple languages while preserving speaker similarity.