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.
bear
Characteristic dimensions (auto, from its own voice profile): ▲ S_STRY ▲ GEND ▲ S_DRAM ▲ AROU ▼ S_TECH ▼ S_WHIS ▼ S_ASMR ▼ REGS · emo: Concentration Interest Malevolence_Malice
Training set = top-20% by reward = z(genu)+z(blend)+z(quality)−z(WER)+0.7·z(profile-match). Kept 122/614.
Best 5 of the kept top-20%
R 6.32 · genu 0.12 · blend 0.70 · qual 0.53 · WER 0.00
R 5.15 · genu 0.06 · blend 0.56 · qual 0.64 · WER 0.00
R 5.12 · genu 0.63 · blend 0.35 · qual 0.56 · WER 0.80
R 5.02 · genu 0.13 · blend 0.44 · qual 0.62 · WER 0.04
R 4.68 · genu 0.30 · blend 0.17 · qual 0.46 · WER 0.04
Weakest 5 still inside the top-20% (the selection boundary)
R 1.28 · genu 0.19 · blend 0.27 · qual 0.33 · WER 0.08
R 1.28 · genu 0.19 · blend 0.07 · qual 0.41 · WER 0.07
R 1.27 · genu 0.08 · blend 0.01 · qual 0.54 · WER 0.00
R 1.27 · genu 0.11 · blend 0.34 · qual 0.39 · WER 0.03
R 1.26 · genu 0.09 · blend 0.16 · qual 0.38 · 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.”