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-01
Characteristic dimensions (auto, from its own voice profile): ▲ GEND ▲ S_MONO ▲ S_STRY ▲ S_NARR ▼ S_RANT ▼ EXPL ▼ DARC ▼ S_ASMR · 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 126/634.
Best 5 of the kept top-20%
R 8.23 · genu 0.47 · blend 0.44 · qual 0.52 · WER 0.00
R 7.02 · genu 0.40 · blend 0.31 · qual 0.51 · WER 0.00
R 6.85 · genu 0.36 · blend 0.26 · qual 0.60 · WER 0.00
R 5.51 · genu 0.13 · blend 0.66 · qual 0.65 · WER 0.00
R 5.10 · genu 0.17 · blend 0.77 · qual 0.52 · WER 0.10
Weakest 5 still inside the top-20% (the selection boundary)
R 1.36 · genu 0.04 · blend 0.21 · qual 0.58 · WER 0.00
R 1.36 · genu 0.11 · blend 0.01 · qual 0.59 · WER 0.00
R 1.34 · genu 0.14 · blend 0.60 · qual 0.36 · WER 0.00
R 1.31 · genu 0.00 · blend 0.00 · qual 0.73 · WER 0.00
R 1.30 · genu 0.18 · blend 0.00 · qual 0.52 · 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.”