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.
human
Characteristic dimensions (auto, from its own voice profile): ▲ GEND ▲ S_MONO ▲ S_STRY ▲ STRU ▼ S_RANT ▼ EXPL ▼ S_ASMR ▼ S_TECH · emo: Concentration Interest Emotional_Numbness
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 8.26 · genu 0.59 · blend 0.40 · qual 0.46 · WER 0.07
R 7.02 · genu 0.39 · blend 0.16 · qual 0.64 · WER 0.00
R 6.28 · genu 0.45 · blend 0.27 · qual 0.55 · WER 0.06
R 6.03 · genu 0.24 · blend 0.60 · qual 0.52 · WER 0.00
R 5.24 · genu 0.18 · blend 0.86 · qual 0.46 · WER 0.00
Weakest 5 still inside the top-20% (the selection boundary)
R 1.35 · genu 0.07 · blend 0.05 · qual 0.64 · WER 0.00
R 1.35 · genu 0.18 · blend 0.35 · qual 0.45 · WER 0.00
R 1.35 · genu 0.17 · blend 0.20 · qual 0.51 · WER 0.00
R 1.35 · genu 0.05 · blend 0.14 · qual 0.65 · WER 0.07
R 1.35 · genu 0.15 · blend 0.26 · qual 0.51 · WER 0.04
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.”