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.
imp
Characteristic dimensions (auto, from its own voice profile): ▲ S_STRY ▲ S_DRAM ▲ AROU ▲ EMPH ▼ EXPL ▼ S_NEWS ▼ S_TECH ▼ S_WHIS · emo: Concentration Interest Hope_Enthusiasm_Optimism
Training set = top-20% by reward = z(genu)+z(blend)+z(quality)−z(WER)+0.7·z(profile-match). Kept 125/629.
Best 5 of the kept top-20%
R 7.80 · genu 0.48 · blend 0.59 · qual 0.28 · WER 0.03
R 6.18 · genu 0.36 · blend 0.45 · qual 0.49 · WER 0.07
R 6.00 · genu 0.36 · blend 0.40 · qual 0.39 · WER 0.03
R 4.67 · genu 0.21 · blend 0.70 · qual 0.37 · WER 0.00
R 4.52 · genu 0.28 · blend 0.09 · qual 0.58 · WER 0.00
Weakest 5 still inside the top-20% (the selection boundary)
R 1.15 · genu 0.11 · blend 0.29 · qual 0.36 · WER 0.00
R 1.13 · genu 0.01 · blend 0.01 · qual 0.73 · WER 0.00
R 1.13 · genu 0.14 · blend 0.07 · qual 0.56 · WER 0.07
R 1.11 · genu 0.18 · blend 0.03 · qual 0.45 · WER 0.00
R 1.11 · genu 0.10 · blend 0.47 · qual 0.28 · WER 0.07
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.”