Profile-reward recipe: generate 640 (batched) → ASR-filter+trim → SIDON → keep the top 20% by a per-character quality+profile reward → warm-start FT 3 epochs. Below: the quality range of the selected training set, then example generations after each epoch.
zandalari
Characteristic dimensions (auto, from its own voice profile): ▲ S_STRY ▲ REGS ▲ CLRT ▲ BRGT ▼ EXPL ▼ S_RANT ▼ S_NEWS ▼ GEND · emo: Concentration Interest Awe
Training set = top-20% by reward = z(genu)+z(blend)+z(quality)−z(WER)+0.7·z(profile-match). Kept 127/637.
Best 5 of the kept top-20%
R 6.23 · genu 0.28 · blend 0.58 · qual 0.59 · WER 0.00
R 4.15 · genu 0.14 · blend 0.41 · qual 0.60 · WER 0.00
R 4.14 · genu 0.22 · blend 0.32 · qual 0.58 · WER 0.00
R 4.11 · genu 0.20 · blend 0.38 · qual 0.59 · WER 0.00
R 3.94 · genu 0.16 · blend 0.31 · qual 0.66 · WER 0.00
Weakest 5 still inside the top-20% (the selection boundary)
R 1.21 · genu 0.17 · blend 0.02 · qual 0.62 · WER 0.03
R 1.20 · genu 0.14 · blend 0.13 · qual 0.56 · WER 0.00
R 1.19 · genu 0.17 · blend 0.11 · qual 0.59 · WER 0.03
R 1.18 · genu 0.18 · blend 0.29 · qual 0.42 · WER 0.03
R 1.17 · genu 0.12 · blend 0.08 · qual 0.63 · WER 0.00
Fantasy line — refined ep1 / ep2 / ep3
“The old road winds past the ruined tower where the last dragon knights once kept their watch.”
Sci-fi line — refined ep1 / ep2 / ep3
“The station drifted in low orbit while the reactor cooled and the scanners swept the empty dark.”
Neutral line — refined ep1 / ep2 / ep3
“The meeting is scheduled for three o'clock and the report is on the desk by the printer.”