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-23
Characteristic dimensions (auto, from its own voice profile): ▲ S_MONO ▲ S_NARR ▲ S_STRY ▲ AGEV ▼ 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 123/616.
Best 5 of the kept top-20%
R 5.72 · genu 0.17 · blend 0.55 · qual 0.66 · WER 0.11
R 4.44 · genu 0.10 · blend 0.36 · qual 0.68 · WER 0.00
R 4.00 · genu 0.29 · blend 0.28 · qual 0.45 · WER 0.03
R 3.90 · genu 0.15 · blend 0.26 · qual 0.68 · WER 0.00
R 3.83 · genu 0.09 · blend 0.32 · qual 0.68 · WER 0.00
Weakest 5 still inside the top-20% (the selection boundary)
R 1.51 · genu 0.06 · blend 0.16 · qual 0.63 · WER 0.03
R 1.50 · genu 0.03 · blend 0.17 · qual 0.68 · WER 0.03
R 1.47 · genu 0.00 · blend 0.11 · qual 0.72 · WER 0.00
R 1.46 · genu 0.00 · blend 0.25 · qual 0.65 · WER 0.00
R 1.45 · genu 0.11 · blend 0.29 · qual 0.49 · 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.”