← all dimensions · markdown (agent skill) · manual home
Roughness VoiceNet dimension guide
Spectral & Timbralcode vn_ROUGslot 62
This dimension measures the grit / rapid amplitude modulation in the vocal-fold vibration. Below: the merge-scaling dose-response for the __high and __low LoRA (laion/moss-voicenet-dimension-loras) at 25/50/75/100/125%, the measured target value at each dose, the top-3 ± correlated EmoNet emotions and VoiceNet dimensions, three audio samples per dose, and a grounded prompting tip. Costs are shifts vs. the 24-sample neutral baseline in genuineness, vocal-burst blend and speech quality.
Roughness controls the grit / rapid amplitude modulation in the vocal-fold vibration. The __high LoRA shifts it +0.31 at 125% (worst cost: burst-blend -0.41); the __low LoRA shifts it -0.10 at 50% (worst cost: genuineness -0.03). The __high direction is the stronger, more reliable knob here.
▲ HIGH — Roughness __high
Dose-response
| dose | target | Δtarget | Δgenu | Δblend | Δqual |
|---|
| baseline (0%) | 0.430 | — | — | — | — |
| 25% | 0.358 | -0.07 | -0.01 | -0.22 | -0.07 |
| 50% | 0.491 | +0.06 | -0.04 | +0.04 | +0.05 |
| 75% | 0.726 | +0.30 | -0.06 | -0.28 | -0.14 |
| 100% | 0.688 | +0.26 | -0.08 | -0.36 | -0.17 |
| 125% ★ | 0.740 | +0.31 | -0.07 | -0.41 | -0.21 |
EmoNet emotion correlation
Malevolence/Malice+0.66
Anger+0.57
Triumph+0.52
Contentment-0.73
Affection-0.57
Relief-0.47
VoiceNet dimension correlation
Cartoonish Style+0.92
Dramatic Style+0.76
Tension+0.75
Recording Quality-0.50
Warmth-0.48
Register-0.48
Listen — 3 samples per dose
25%measured target: 0.31 · 0.42 · 0.28
50%measured target: 0.30 · 0.53 · 0.69
75%measured target: 0.77 · 0.68 · 0.77
100%measured target: 0.47 · 0.76 · 0.76
125% ★measured target: 0.54 · 0.71 · 0.84
Merging Roughness __high raises the dimension: at the recommended 125% dose the measured Roughness score moves from a 0.43 baseline to 0.74 (a +0.31 shift). The dose-response is clean and monotonic — each step pushes the dimension further. That reach is a steep cost — the biggest hit is to burst-blend (-0.41); across the board genuineness -0.07, burst-blend -0.41, speech quality -0.21. Emotionally it reads as Malevolence/Malice (r 0.66), Anger (0.57) and Triumph (0.52), and it moves away from Contentment, Affection and Relief. On other voice dimensions it co-moves most with Cartoonish Style (0.92), Dramatic Style (0.76) and Tension (0.75), while pulling down Recording Quality, Warmth and Register. Practical call: merge __high at 100–125%; expect Malevolence/Malice and Cartoonish Style to rise and burst-blend to fall. If you don’t want Malevolence/Malice bleeding in, offset it with a prompt cue toward Contentment to keep the roughness shift without the side-effect.
▼ LOW — Roughness __low
Dose-response
| dose | target | Δtarget | Δgenu | Δblend | Δqual |
|---|
| baseline (0%) | 0.430 | — | — | — | — |
| 25% | 0.395 | -0.03 | -0.06 | -0.02 | +0.10 |
| 50% ★ | 0.327 | -0.10 | -0.03 | +0.06 | +0.03 |
| 75% | 0.344 | -0.09 | -0.07 | -0.14 | +0.08 |
| 100% | 0.413 | -0.02 | -0.06 | -0.07 | +0.17 |
| 125% | 0.333 | -0.10 | -0.03 | -0.18 | +0.10 |
EmoNet emotion correlation
Impatience and Irritability+0.44
Malevolence/Malice+0.35
Bitterness+0.27
Contentment-0.30
Contemplation-0.17
Pleasure/Ecstasy-0.12
VoiceNet dimension correlation
Cartoonish Style+0.77
Chest Resonance+0.72
Dramatic Style+0.70
Register-0.74
Valence-0.55
Oral Resonance-0.53
Listen — 3 samples per dose
25%measured target: 0.24 · 0.45 · 0.18
50% ★measured target: 0.15 · 0.20 · 0.30
75%measured target: 0.34 · 0.52 · 0.44
100%measured target: 0.29 · 0.34 · 0.36
125%measured target: 0.27 · 0.37 · 0.32
Merging Roughness __low suppresses the dimension: at the recommended 50% dose the measured Roughness score moves from a 0.43 baseline to 0.33 (a -0.10 shift). The dose-response is mostly ordered but noisy between doses. That reach is nearly free — the biggest hit is to genuineness (-0.03); across the board genuineness -0.03, burst-blend +0.06, speech quality +0.03. Emotionally it reads as Impatience and Irritability (r 0.44), Malevolence/Malice (0.35) and Bitterness (0.27), and it moves away from Contentment, Contemplation and Pleasure/Ecstasy. On other voice dimensions it co-moves most with Cartoonish Style (0.77), Chest Resonance (0.72) and Dramatic Style (0.70), while pulling down Register, Valence and Oral Resonance. Practical call: merge __low at 25–75%; expect Impatience and Irritability and Cartoonish Style to rise and genuineness to fall. If you don’t want Impatience and Irritability bleeding in, offset it with a prompt cue toward Contentment to keep the roughness shift without the side-effect.