original corpus recordings vs generated winner B · 7 voices × 4 lines
Round 07 — SOURCE VERIFICATION. Rows SRC_* are the ORIGINAL corpus recordings, untouched — this is what the model learned from. Bottom row = the generated winner B on the same sentences.
The theory under test: shigure's recordings have hot sibilants + a hiss floor (measured S-harshness 2.92 vs ~1.5 others, floor ~9x) — listen to SRC_shigure's sib1/sib2 against SRC_ami_pun / SRC_aoba on the same sentences and judge for yourself.
Also compare SRC rows vs GEN row: how much of the S noise is inherited from the source (present in SRC_shigure) vs added by the vocoder (present in GEN but not in clean SRC rows).
voice
sib1ディスカッションを進める。RECITATION324_208 — densest sibilants (and it contains 進める)