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.
turkey
Characteristic dimensions (auto, from its own voice profile): ▲ S_DRAM ▲ AROU ▲ VOLT ▲ S_CASU ▼ EXPL ▼ S_FORM ▼ S_TECH ▼ S_ASMR · emo: Amusement Elation Pleasure_Ecstasy
Training set = top-20% by reward = z(genu)+z(blend)+z(quality)−z(WER)+0.7·z(profile-match). Kept 23/118.
Best 5 of the kept top-20%
R 3.57 · genu 0.69 · blend 0.67 · qual 0.17 · WER 1.00
R 2.75 · genu 0.50 · blend 0.83 · qual 0.18 · WER 1.00
R 2.34 · genu 0.11 · blend 0.74 · qual 0.62 · WER 0.80
R 2.25 · genu 0.26 · blend 0.60 · qual 0.36 · WER 0.66
R 2.11 · genu 0.36 · blend 0.68 · qual 0.33 · WER 0.80
Weakest 5 still inside the top-20% (the selection boundary)
R 1.10 · genu 0.30 · blend 0.97 · qual 0.12 · WER 1.00
R 1.04 · genu 0.33 · blend 0.82 · qual 0.20 · WER 1.00
R 0.99 · genu 0.34 · blend 0.81 · qual 0.14 · WER 1.00
R 0.96 · genu 0.45 · blend 0.83 · qual 0.11 · WER 1.00
R 0.95 · genu 0.38 · blend 0.78 · qual 0.21 · WER 1.00
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.”