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-10
Characteristic dimensions (auto, from its own voice profile): ▲ S_MONO ▲ S_NARR ▲ S_STRY ▲ ESTH ▼ EXPL ▼ S_RANT ▼ DARC ▼ S_CONV · 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 127/637.
Best 5 of the kept top-20%
R 6.66 · genu 0.21 · blend 0.83 · qual 0.68 · WER 0.55
R 4.02 · genu 0.22 · blend 0.93 · qual 0.50 · WER 0.84
R 3.99 · genu 0.26 · blend 0.55 · qual 0.61 · WER 0.82
R 3.95 · genu 0.24 · blend 0.68 · qual 0.65 · WER 0.63
R 3.73 · genu 0.08 · blend 0.18 · qual 0.74 · WER 0.00
Weakest 5 still inside the top-20% (the selection boundary)
R 1.25 · genu 0.09 · blend 0.06 · qual 0.62 · WER 0.00
R 1.23 · genu 0.14 · blend 0.01 · qual 0.54 · WER 0.00
R 1.22 · genu 0.08 · blend 0.41 · qual 0.61 · WER 0.58
R 1.21 · genu 0.00 · blend 0.05 · qual 0.76 · WER 0.00
R 1.21 · genu 0.04 · blend 0.06 · qual 0.72 · 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.”