Voice profiles — listen

One reference voice, rendered through all 832 conditions, three different ways. Every clip below is the take that won its group, with the scores that made it win.

What a "voice profile" is. Take one reference speaker. Ask the model to say something as that speaker in every combination of a large grid: 40 emotions × {intense, moderate} × {freely expressed, held in}, 57 voice-quality dimensions × 4 levels, plus screams, laughter, crying, a sports commentator and more — each in English and German, from the same sentence. That grid is 832 groups. A profile is all of them for one voice.

Profile A — translation-corpus sentences → The baseline. Sentences drawn from the FineTranslations corpus and used as-is. No Mediathek LoRA. This is what the pipeline produced before the text-rewrite experiment. Profile B — the same, plus 25 % Mediathek LoRA → Identical sentences and conditions, with a real-speech adapter merged at λ=0.25. A later dose sweep found this to be a null result — kept published so you can hear that for yourself rather than take it on trust. Profile RW — sentences rewritten to carry the emotion → Same voice, same 832 conditions, but every sentence rewritten so its content gives the speaker a reason to feel the target emotion. Emotion strength +22 %, genuineness +27 %, blend +39 %, audio quality flat, word error rate down 12 %.

The comparison worth making

Open A and RW side by side and listen to Bitterness, Contempt or Disappointment in the "intense · free" column. Same voice, same adapter, same settings — only the words differ.

emotion	A (corpus sentences)	RW (rewritten)
Bitterness	−0.005	0.613
Contempt	0.041	0.842
Disappointment	0.028	1.499
Hope / Enthusiasm	−0.001	1.390
Astonishment / Surprise	0.039	2.347

Emotion-head score on the winning take. Full analysis: Does the text carry the emotion?

Audio quality

Profile RW is now 48 kHz at 160 kbps — regenerated at the model's native rate. Profiles A and B remain 16 kHz mono at ~97 kbps: the pipeline used to resample to 16 kHz before storing, band-limiting everything to 8 kHz, and that ceiling is baked into their data.

Because of that, do not A/B profile RW against profile A on this page to judge the text rewrite — you would be hearing the bandwidth as much as the words. The side-by-side demo grid deliberately keeps both columns at 16 kHz so the only difference there is the text.