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-18
Characteristic dimensions (auto, from its own voice profile): ▲ S_STRY ▲ AGEV ▲ S_MONO ▲ S_NARR ▼ S_RANT ▼ DARC ▼ EXPL ▼ S_NEWS · emo: Concentration Awe 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 5.00 · genu 0.15 · blend 0.47 · qual 0.73 · WER 0.00
R 4.31 · genu 0.25 · blend 0.28 · qual 0.66 · WER 0.00
R 4.25 · genu 0.12 · blend 0.85 · qual 0.61 · WER 0.00
R 4.20 · genu 0.19 · blend 0.55 · qual 0.53 · WER 0.00
R 4.20 · genu 0.17 · blend 0.44 · qual 0.60 · WER 0.00
Weakest 5 still inside the top-20% (the selection boundary)
R 1.38 · genu 0.14 · blend 0.37 · qual 0.57 · WER 0.00
R 1.36 · genu 0.11 · blend 0.29 · qual 0.66 · WER 0.00
R 1.36 · genu 0.02 · blend 0.15 · qual 0.76 · WER 0.00
R 1.36 · genu 0.20 · blend 0.55 · qual 0.64 · WER 0.60
R 1.36 · genu 0.09 · blend 0.16 · qual 0.72 · WER 0.08
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.”