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.
kul_tiran_human
Characteristic dimensions (auto, from its own voice profile): ▲ S_STRY ▲ GEND ▲ S_MONO ▲ S_DRAM ▼ S_TECH ▼ S_ASMR ▼ S_WHIS ▼ EXPL · 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 8.14 · genu 0.21 · blend 0.62 · qual 0.69 · WER 0.43
R 7.50 · genu 0.11 · blend 0.45 · qual 0.66 · WER 0.00
R 7.32 · genu 0.17 · blend 0.35 · qual 0.66 · WER 0.07
R 6.54 · genu 0.38 · blend 0.60 · qual 0.52 · WER 0.77
R 6.10 · genu 0.42 · blend 0.10 · qual 0.51 · WER 0.00
Weakest 5 still inside the top-20% (the selection boundary)
R 1.38 · genu 0.04 · blend 0.10 · qual 0.55 · WER 0.03
R 1.38 · genu 0.07 · blend 0.06 · qual 0.58 · WER 0.03
R 1.37 · genu 0.14 · blend 0.00 · qual 0.50 · WER 0.00
R 1.37 · genu 0.13 · blend 0.01 · qual 0.50 · WER 0.00
R 1.36 · genu 0.27 · blend 0.09 · qual 0.51 · WER 0.42
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.”