German Mediathek HQ LoRA — 7 emotions, German & English
Does the Mediathek high-quality adapter carry emotion on its own, or does it need
the per-emotion adapter stacked on top? Both arms get the same emotion-describing GENERAL
caption, the same delivery cue and the same inline vocal burst — the only difference
is whether the emotion adapter is merged. 104 clips, 4 candidates per cell, none discarded.
Configuration. Mediathek HQ adapter
r64 / epoch 3 at
1.0 in every cell (it is the thing under test, not a variable). Emotion adapters from
moss-emotion-loras-v3 at
0.5. Burst adapters from
vocal-burst-lora-adapters at
0.5 — the corrected dose: 0.75–1.0 raises burst probability but eats the words after the
burst (tail coverage 0.90 at 0.5 vs 0.45 at 1.0).
Duration by cell
| emotion | EN · HQ only | EN · HQ+emotion |
DE · HQ only | DE · HQ+emotion |
|---|
| neutral | 10.2 s | | 3.8 s | |
| anger | 14.4 s | 16.9 s | 3.8 s | 4.2 s |
| sadness | 8.2 s | 17.6 s | 3.8 s | 9.9 s |
| joy | 10.0 s | 22.3 s | 3.9 s | 4.8 s |
| surprise | 5.0 s | 4.8 s | 3.7 s | 3.7 s |
| amusement | 11.0 s | 16.6 s | 3.7 s | 4.0 s |
| fear | 21.4 s | 21.7 s | 3.8 s | 3.7 s |
The English takes run long; the German ones do not. German lands at
3.7–4.8 s for a 9-word line (~2.4 words/s, about right for the language). English reaches 10–22 s
for a 13-word line, which is far past natural pacing — the model is stretching and padding rather
than speaking. Adding the emotion adapter makes it worse in English (joy 10.0 → 22.3 s, sadness
8.2 → 17.6 s) and barely moves German.
That asymmetry is expected in direction — this adapter was trained on German broadcast
audio only — but the size of it is worth knowing before using it for English. Judge the English
cells on register rather than on timing.
Listen
Left two columns: HQ LoRA alone. Right two: HQ + the emotion adapter. Neutral has
no emotion adapter, so it has only the two "HQ only" columns.
neutral
inline burst: — @0.5 · emotion adapter: —@0.5 (right-hand columns only) · Mediathek HQ adapter @1.0 in every cell
EN: I had the whole thing planned out and then the call came through.
DE: Ich hatte alles genau geplant und dann kam dieser Anruf.
anger
inline burst: displeased grunt @0.5 · emotion adapter: Anger@0.5 (right-hand columns only) · Mediathek HQ adapter @1.0 in every cell
EN: I had the whole thing planned out and then the call came through.
DE: Ich hatte alles genau geplant und dann kam dieser Anruf.
sadness
inline burst: quiet sob @0.5 · emotion adapter: Sadness@0.5 (right-hand columns only) · Mediathek HQ adapter @1.0 in every cell
EN: I had the whole thing planned out and then the call came through.
DE: Ich hatte alles genau geplant und dann kam dieser Anruf.
English — HQ + Sadness@0.5
German — HQ + Sadness@0.5
joy
inline burst: guffaw @0.5 · emotion adapter: Elation@0.5 (right-hand columns only) · Mediathek HQ adapter @1.0 in every cell
EN: I had the whole thing planned out and then the call came through.
DE: Ich hatte alles genau geplant und dann kam dieser Anruf.
English — HQ + Elation@0.5
German — HQ + Elation@0.5
surprise
inline burst: surprised gasp @0.5 · emotion adapter: Astonishment_Surprise@0.5 (right-hand columns only) · Mediathek HQ adapter @1.0 in every cell
EN: I had the whole thing planned out and then the call came through.
DE: Ich hatte alles genau geplant und dann kam dieser Anruf.
English — HQ + Astonishment_Surprise@0.5
German — HQ + Astonishment_Surprise@0.5
amusement
inline burst: chuckle @0.5 · emotion adapter: Amusement@0.5 (right-hand columns only) · Mediathek HQ adapter @1.0 in every cell
EN: I had the whole thing planned out and then the call came through.
DE: Ich hatte alles genau geplant und dann kam dieser Anruf.
English — HQ + Amusement@0.5
German — HQ + Amusement@0.5
fear
inline burst: fearful gasp @0.5 · emotion adapter: Fear@0.5 (right-hand columns only) · Mediathek HQ adapter @1.0 in every cell
EN: I had the whole thing planned out and then the call came through.
DE: Ich hatte alles genau geplant und dann kam dieser Anruf.
Adapters:
laion/moss-mediathek-hq-lora ·
base moss-tts-local-transformer-4.55b-voice-acting-v2 ·
dataset TTS-AGI/german-mediathek-hq-expressive (private) ·
prompting manual.