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-13
Characteristic dimensions (auto, from its own voice profile): ▲ S_STRY ▲ S_PLAY ▲ S_CASU ▲ AROU ▼ S_NEWS ▼ EXPL ▼ S_WHIS ▼ COGL · emo: Concentration Teasing Amusement
Training set = top-20% by reward = z(genu)+z(blend)+z(quality)−z(WER)+0.7·z(profile-match). Kept 117/587.
Best 5 of the kept top-20%
R 6.07 · genu 0.33 · blend 0.41 · qual 0.40 · WER 0.00
R 5.60 · genu 0.35 · blend 0.87 · qual 0.35 · WER 0.90
R 5.35 · genu 0.29 · blend 0.28 · qual 0.47 · WER 0.00
R 5.22 · genu 0.17 · blend 0.74 · qual 0.49 · WER 0.77
R 5.18 · genu 0.17 · blend 0.33 · qual 0.55 · WER 0.00
Weakest 5 still inside the top-20% (the selection boundary)
R 1.53 · genu 0.13 · blend 0.01 · qual 0.45 · WER 0.00
R 1.52 · genu 0.21 · blend 0.22 · qual 0.39 · WER 0.00
R 1.50 · genu 0.08 · blend 0.19 · qual 0.37 · WER 0.00
R 1.50 · genu 0.16 · blend 0.07 · qual 0.42 · WER 0.06
R 1.49 · genu 0.01 · blend 0.04 · qual 0.55 · 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.”