Each example asks the model to speak an English sentence in the voice of a non-English reference — keeping the speaker's timbre without carrying their accent across. The trade-off to listen for: more aggressive masking removes more accent but increasingly destroys the voice, sometimes flipping the apparent gender. The samples below are drawn from the FLEURS evaluation set used in the paper, and the scores shown are the per-clip objective metrics and human listening-test ratings.
Accent is the mean listener rating: 0 = none, 1 = very slight, 2 = mild, 3 = moderate, 4 = strong, 5 = very strong.