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.
lightforged_draenei
Characteristic dimensions (auto, from its own voice profile): ▲ S_MONO ▲ S_STRY ▲ STRU ▲ ESTH ▼ S_RANT ▼ EXPL ▼ DARC ▼ S_TECH · emo: Concentration Emotional_Numbness Interest
Training set = top-20% by reward = z(genu)+z(blend)+z(quality)−z(WER)+0.7·z(profile-match). Kept 123/618.
Best 5 of the kept top-20%
R 6.78 · genu 0.24 · blend 0.27 · qual 0.65 · WER 0.00
R 6.53 · genu 0.13 · blend 0.64 · qual 0.57 · WER 0.00
R 6.01 · genu 0.28 · blend 0.53 · qual 0.55 · WER 0.00
R 5.85 · genu 0.09 · blend 0.68 · qual 0.65 · WER 0.00
R 5.72 · genu 0.20 · blend 0.21 · qual 0.64 · WER 0.00
Weakest 5 still inside the top-20% (the selection boundary)
R 1.34 · genu 0.15 · blend 0.22 · qual 0.48 · WER 0.00
R 1.33 · genu 0.03 · blend 0.00 · qual 0.71 · WER 0.00
R 1.30 · genu 0.00 · blend 0.20 · qual 0.63 · WER 0.03
R 1.29 · genu 0.04 · blend 0.11 · qual 0.65 · WER 0.07
R 1.28 · genu 0.01 · blend 0.05 · qual 0.67 · 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.”