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-24
Characteristic dimensions (auto, from its own voice profile): ▲ S_STRY ▲ S_MONO ▲ S_NARR ▲ GEND ▼ EXPL ▼ S_CONV ▼ S_ASMR ▼ REGS · 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 126/631.
Best 5 of the kept top-20%
R 6.41 · genu 0.43 · blend 0.28 · qual 0.26 · WER 0.00
R 5.88 · genu 0.26 · blend 0.81 · qual 0.55 · WER 0.83
R 5.72 · genu 0.15 · blend 0.62 · qual 0.44 · WER 0.00
R 4.77 · genu 0.48 · blend 0.34 · qual 0.20 · WER 0.09
R 4.47 · genu 0.39 · blend 0.29 · qual 0.18 · WER 0.03
Weakest 5 still inside the top-20% (the selection boundary)
R 1.11 · genu 0.11 · blend 0.00 · qual 0.57 · WER 0.06
R 1.10 · genu 0.10 · blend 0.00 · qual 0.58 · WER 0.00
R 1.09 · genu 0.00 · blend 0.00 · qual 0.69 · WER 0.00
R 1.09 · genu 0.08 · blend 0.07 · qual 0.52 · WER 0.00
R 1.08 · genu 0.20 · blend 0.02 · qual 0.32 · 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.”