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-22
Characteristic dimensions (auto, from its own voice profile): ▲ S_MONO ▲ S_NARR ▲ S_STRY ▲ STRU ▼ 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 126/633.
Best 5 of the kept top-20%
R 5.57 · genu 0.11 · blend 0.45 · qual 0.74 · WER 0.00
R 5.36 · genu 0.11 · blend 0.55 · qual 0.55 · WER 0.00
R 5.15 · genu 0.23 · blend 0.30 · qual 0.60 · WER 0.00
R 4.96 · genu 0.09 · blend 0.38 · qual 0.69 · WER 0.00
R 4.52 · genu 0.11 · blend 0.21 · qual 0.71 · WER 0.00
Weakest 5 still inside the top-20% (the selection boundary)
R 1.26 · genu 0.01 · blend 0.06 · qual 0.70 · WER 0.00
R 1.24 · genu 0.00 · blend 0.21 · qual 0.65 · WER 0.07
R 1.24 · genu 0.08 · blend 0.00 · qual 0.64 · WER 0.00
R 1.24 · genu 0.00 · blend 0.02 · qual 0.76 · WER 0.00
R 1.24 · genu 0.03 · blend 0.09 · qual 0.68 · 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.”