Does the vocal-burst LoRA help inside an emotional sentence?
The A/B the merge sweep was missing. Every clip below is the same kind of
sentence — ordinary speech with one burst asked for mid-utterance, e.g.
(amused) "I was just about to leave when" (Chuckle) "and then everything made sense."
— generated four ways with everything else held fixed: same carriers, same seeds, same sampling.
MOSS performs inline burst markers from the caption on its own to some extent, so the
no-LoRA column is the baseline the adapter has to beat.
Headline: the burst adapter changes
burst presence from 17 % to 71 % (+54 pp) when stacked
on the emotion adapter, holding prompt and emotion dose fixed. The trade is blend — see the table.
The four cells, pooled
| configuration | burst present | blend |
genuineness | burst duration | n |
|---|
| no LoRA at all Z_base
| 23.6 % | 5.98 | 1.00 | 0.36 s | 360 |
| emotion LoRA @0.5 only Z_emo
| 16.7 % | 6.27 | 1.15 | 0.40 s | 360 |
| burst LoRA @1.0 only Z_burst
| 71.9 % | 3.93 | 1.94 | 0.56 s | 360 |
| burst @1.0 + emotion @0.5 Z_both
| 70.8 % | 3.82 | 1.84 | 0.54 s | 360 |
Per emotion family
| family | no LoRA at all | emotion LoRA @0.5 only | burst LoRA @1.0 only | burst @1.0 + emotion @0.5 |
burst LoRA effect |
|---|
| Anger | 19 % | 29 % | 74 % | 69 % | +40 pp |
| Fatigue_Exhaustion | 33 % | 12 % | 62 % | 60 % | +48 pp |
| Fear | 15 % | 29 % | 90 % | 85 % | +56 pp |
| Malevolence_Malice | 15 % | 25 % | 69 % | 65 % | +40 pp |
| Pain | 17 % | 7 % | 81 % | 83 % | +76 pp |
| Sadness | 31 % | 8 % | 81 % | 77 % | +69 pp |
| Sexual_Lust | 33 % | 5 % | 43 % | 55 % | +50 pp |
| Teasing | 27 % | 17 % | 73 % | 71 % | +54 pp |
The last column is Z_both − Z_emo: what the burst adapter adds once the
emotion adapter and the prompt are already in place.
Listen
For each family, the two burst classes where the adapter moves the needle most.
3 candidates per cell, unselected (candidates 0–2 of 6, not the best ones).
Anger
Clears Throat
no LoRA at all
burst in 17 % · blend 5.72
emotion LoRA @0.5 only
burst in 17 % · blend 5.44
burst LoRA @1.0 only
burst in 100 % · blend 4.08
burst @1.0 + emotion @0.5
burst in 100 % · blend 3.61
Exasperated Sigh
no LoRA at all
burst in 33 % · blend 6.98
emotion LoRA @0.5 only
burst in 33 % · blend 6.51
burst LoRA @1.0 only
burst in 50 % · blend 3.71
burst @1.0 + emotion @0.5
burst in 83 % · blend 4.64
Fatigue_Exhaustion
Relief Sigh
no LoRA at all
burst in 17 % · blend 6.01
emotion LoRA @0.5 only
burst in 0 % · blend 7.41
burst LoRA @1.0 only
burst in 100 % · blend 6.41
burst @1.0 + emotion @0.5
burst in 100 % · blend 5.91
Exhausted Groan
no LoRA at all
burst in 17 % · blend 6.46
emotion LoRA @0.5 only
burst in 0 % · blend 6.65
burst LoRA @1.0 only
burst in 100 % · blend 4.72
burst @1.0 + emotion @0.5
burst in 83 % · blend 5.95
Fear
Fearful Gasp
no LoRA at all
burst in 17 % · blend 6.94
emotion LoRA @0.5 only
burst in 17 % · blend 5.82
burst LoRA @1.0 only
burst in 100 % · blend 3.56
burst @1.0 + emotion @0.5
burst in 100 % · blend 4.32
Scream
no LoRA at all
burst in 50 % · blend 6.41
emotion LoRA @0.5 only
burst in 17 % · blend 6.00
burst LoRA @1.0 only
burst in 67 % · blend 2.49
burst @1.0 + emotion @0.5
burst in 100 % · blend 2.44
Malevolence_Malice
Low Mumble
no LoRA at all
burst in 33 % · blend 7.13
emotion LoRA @0.5 only
burst in 17 % · blend 8.15
burst LoRA @1.0 only
burst in 67 % · blend 3.17
burst @1.0 + emotion @0.5
burst in 83 % · blend 3.22
Resonant Hum
no LoRA at all
burst in 0 % · blend 6.31
emotion LoRA @0.5 only
burst in 17 % · blend 6.80
burst LoRA @1.0 only
burst in 83 % · blend 2.86
burst @1.0 + emotion @0.5
burst in 83 % · blend 2.43
Pain
Coughing
no LoRA at all
burst in 33 % · blend 5.24
emotion LoRA @0.5 only
burst in 0 % · blend 4.42
burst LoRA @1.0 only
burst in 67 % · blend 3.49
burst @1.0 + emotion @0.5
burst in 100 % · blend 2.33
Cough
no LoRA at all
burst in 0 % · blend 6.37
emotion LoRA @0.5 only
burst in 0 % · blend 5.47
burst LoRA @1.0 only
burst in 83 % · blend 2.88
burst @1.0 + emotion @0.5
burst in 83 % · blend 3.48
Sadness
Quiet Sob
no LoRA at all
burst in 50 % · blend 6.20
emotion LoRA @0.5 only
burst in 0 % · blend 4.76
burst LoRA @1.0 only
burst in 83 % · blend 4.57
burst @1.0 + emotion @0.5
burst in 100 % · blend 4.07
Sobs
no LoRA at all
burst in 0 % · blend 5.84
emotion LoRA @0.5 only
burst in 0 % · blend 7.42
burst LoRA @1.0 only
burst in 67 % · blend 3.88
burst @1.0 + emotion @0.5
burst in 100 % · blend 1.92
Sexual_Lust
Breathy Giggle
no LoRA at all
burst in 17 % · blend 5.60
emotion LoRA @0.5 only
burst in 0 % · blend 6.71
burst LoRA @1.0 only
burst in 100 % · blend 6.07
burst @1.0 + emotion @0.5
burst in 100 % · blend 4.28
Pleasure Moan
no LoRA at all
burst in 67 % · blend 5.18
emotion LoRA @0.5 only
burst in 0 % · blend 6.38
burst LoRA @1.0 only
burst in 83 % · blend 6.12
burst @1.0 + emotion @0.5
burst in 100 % · blend 5.02
Teasing
Chuckle
no LoRA at all
burst in 50 % · blend 5.32
emotion LoRA @0.5 only
burst in 17 % · blend 6.13
burst LoRA @1.0 only
burst in 83 % · blend 5.08
burst @1.0 + emotion @0.5
burst in 100 % · blend 4.10
Guffaw
no LoRA at all
burst in 0 % · blend 7.06
emotion LoRA @0.5 only
burst in 17 % · blend 5.95
burst LoRA @1.0 only
burst in 100 % · blend 1.84
burst @1.0 + emotion @0.5
burst in 100 % · blend 1.63
Adapters:
laion/vocal-burst-lora-adapters ·
TTS-AGI/moss-emotion-loras-v3 ·
base moss-tts-local-transformer-4.55b-voice-acting-v2.
Method and dose curves:
manual page.