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-07
Characteristic dimensions (auto, from its own voice profile): ▲ S_MONO ▲ GEND ▲ S_NARR ▲ S_STRY ▼ S_RANT ▼ EXPL ▼ 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 128/640.
Best 5 of the kept top-20%
R 5.20 · genu 0.09 · blend 0.64 · qual 0.66 · WER 0.08
R 5.10 · genu 0.12 · blend 0.57 · qual 0.68 · WER 0.00
R 4.77 · genu 0.13 · blend 0.88 · qual 0.46 · WER 0.04
R 4.58 · genu 0.15 · blend 0.67 · qual 0.50 · WER 0.00
R 4.45 · genu 0.10 · blend 0.54 · qual 0.62 · WER 0.00
Weakest 5 still inside the top-20% (the selection boundary)
R 1.40 · genu 0.03 · blend 0.20 · qual 0.65 · WER 0.00
R 1.40 · genu 0.04 · blend 0.10 · qual 0.68 · WER 0.00
R 1.40 · genu 0.09 · blend 0.03 · qual 0.60 · WER 0.00
R 1.40 · genu 0.05 · blend 0.28 · qual 0.59 · WER 0.03
R 1.35 · genu 0.04 · blend 0.21 · qual 0.60 · WER 0.04
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.”