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.
gravelly-draconic-elder
Characteristic dimensions (auto, from its own voice profile): ▲ S_STRY ▲ S_MONO ▲ GEND ▲ S_NARR ▼ S_TECH ▼ EXPL ▼ S_RANT ▼ S_CONV · emo: Concentration Interest Emotional_Numbness
Training set = top-20% by reward = z(genu)+z(blend)+z(quality)−z(WER)+0.7·z(profile-match). Kept 126/633.
Best 5 of the kept top-20%
R 5.57 · genu 0.14 · blend 0.69 · qual 0.74 · WER 0.38
R 4.83 · genu 0.15 · blend 0.68 · qual 0.60 · WER 0.26
R 4.54 · genu 0.16 · blend 0.85 · qual 0.63 · WER 0.56
R 4.36 · genu 0.17 · blend 0.58 · qual 0.62 · WER 0.63
R 4.35 · genu 0.09 · blend 0.39 · qual 0.67 · WER 0.00
Weakest 5 still inside the top-20% (the selection boundary)
R 1.26 · genu 0.16 · blend 0.49 · qual 0.51 · WER 0.58
R 1.26 · genu 0.06 · blend 0.08 · qual 0.61 · WER 0.12
R 1.26 · genu 0.16 · blend 0.11 · qual 0.40 · WER 0.07
R 1.25 · genu 0.02 · blend 0.01 · qual 0.66 · WER 0.03
R 1.25 · genu 0.07 · blend 0.36 · qual 0.46 · 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.”