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-21
Characteristic dimensions (auto, from its own voice profile): ▲ FULL ▲ S_MONO ▲ STRU ▲ S_FORM ▼ S_ASMR ▼ S_RANT ▼ S_WHIS ▼ S_TECH · 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 127/638.
Best 5 of the kept top-20%
R 8.01 · genu 0.14 · blend 0.51 · qual 0.48 · WER 0.00
R 6.32 · genu 0.21 · blend 0.65 · qual 0.59 · WER 0.71
R 5.46 · genu 0.11 · blend 0.40 · qual 0.52 · WER 0.00
R 5.46 · genu 0.17 · blend 0.21 · qual 0.49 · WER 0.04
R 5.38 · genu 0.14 · blend 0.21 · qual 0.55 · WER 0.03
Weakest 5 still inside the top-20% (the selection boundary)
R 1.02 · genu 0.00 · blend 0.00 · qual 0.64 · WER 0.00
R 1.01 · genu 0.00 · blend 0.00 · qual 0.63 · WER 0.00
R 1.01 · genu 0.01 · blend 0.01 · qual 0.64 · WER 0.03
R 1.00 · genu 0.04 · blend 0.02 · qual 0.56 · WER 0.00
R 1.00 · genu 0.01 · blend 0.00 · qual 0.66 · WER 0.03
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.”