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.
gravelly-snarling-overseer
Characteristic dimensions (auto, from its own voice profile): ▲ S_STRY ▲ REGS ▲ S_MONO ▲ ESTH ▼ S_RANT ▼ EXPL ▼ S_NEWS ▼ 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 127/638.
Best 5 of the kept top-20%
R 5.72 · genu 0.17 · blend 0.31 · qual 0.70 · WER 0.00
R 5.19 · genu 0.27 · blend 0.43 · qual 0.53 · WER 0.00
R 4.81 · genu 0.10 · blend 0.35 · qual 0.71 · WER 0.00
R 4.68 · genu 0.20 · blend 0.33 · qual 0.67 · WER 0.00
R 4.57 · genu 0.22 · blend 0.15 · qual 0.62 · WER 0.00
Weakest 5 still inside the top-20% (the selection boundary)
R 1.35 · genu 0.11 · blend 0.12 · qual 0.72 · WER 0.00
R 1.34 · genu 0.14 · blend 0.31 · qual 0.65 · WER 0.57
R 1.34 · genu 0.09 · blend 0.09 · qual 0.57 · WER 0.03
R 1.32 · genu 0.06 · blend 0.20 · qual 0.60 · WER 0.00
R 1.30 · genu 0.16 · blend 0.01 · qual 0.55 · WER 0.03
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.”