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.
voice-31
Characteristic dimensions (auto, from its own voice profile): ▲ S_STRY ▲ GEND ▲ S_DRAM ▲ ROUG ▼ S_TECH ▼ REGS ▼ R_HEAD ▼ EXPL · 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 61/306.
Best 5 of the kept top-20%
R 4.47 · genu 0.37 · blend 0.44 · qual 0.43 · WER 0.03
R 3.42 · genu 0.16 · blend 0.86 · qual 0.53 · WER 0.37
R 3.41 · genu 0.12 · blend 0.68 · qual 0.49 · WER 0.00
R 3.38 · genu 0.14 · blend 0.77 · qual 0.53 · WER 0.00
R 3.36 · genu 0.16 · blend 0.89 · qual 0.34 · WER 0.06
Weakest 5 still inside the top-20% (the selection boundary)
R 1.61 · genu 0.10 · blend 0.05 · qual 0.64 · WER 0.07
R 1.56 · genu 0.16 · blend 0.21 · qual 0.42 · WER 0.03
R 1.52 · genu 0.09 · blend 0.45 · qual 0.56 · WER 0.03
R 1.49 · genu 0.08 · blend 0.43 · qual 0.58 · WER 0.03
R 1.45 · genu 0.52 · blend 0.50 · qual 0.17 · WER 1.00
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.”