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-25
Characteristic dimensions (auto, from its own voice profile): ▲ S_MONO ▲ S_NARR ▲ WARM ▲ RCQL ▼ S_RANT ▼ EXPL ▼ DARC ▼ DFLU · 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 117/585.
Best 5 of the kept top-20%
R 5.41 · genu 0.18 · blend 0.27 · qual 0.71 · WER 0.00
R 5.10 · genu 0.13 · blend 0.53 · qual 0.71 · WER 0.21
R 5.00 · genu 0.15 · blend 0.94 · qual 0.65 · WER 0.69
R 4.53 · genu 0.21 · blend 0.86 · qual 0.58 · WER 0.83
R 4.43 · genu 0.17 · blend 0.37 · qual 0.65 · WER 0.03
Weakest 5 still inside the top-20% (the selection boundary)
R 1.41 · genu 0.08 · blend 0.09 · qual 0.68 · WER 0.03
R 1.39 · genu 0.07 · blend 0.10 · qual 0.70 · WER 0.00
R 1.38 · genu 0.00 · blend 0.12 · qual 0.72 · WER 0.00
R 1.38 · genu 0.15 · blend 0.13 · qual 0.61 · WER 0.00
R 1.37 · genu 0.02 · blend 0.00 · qual 0.78 · 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.”