← all dimensions · markdown (agent skill) · manual home
Disfluency VoiceNet dimension guide
Rhythm & Timingcode vn_DFLUslot 46
This dimension measures the presence of fillers, false starts and self-corrections. 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.
Disfluency controls the presence of fillers, false starts and self-corrections. The __high LoRA shifts it +0.06 at 125% (worst cost: burst-blend -0.19); the __low LoRA shifts it -0.24 at 125% (worst cost: burst-blend -0.31). The __low direction is the stronger, more reliable knob here.
▲ HIGH — Disfluency __high
Dose-response
| dose | target | Δtarget | Δgenu | Δblend | Δqual |
|---|
| baseline (0%) | 0.320 | — | — | — | — |
| 25% | 0.317 | +0.00 | -0.01 | -0.03 | -0.03 |
| 50% | 0.305 | -0.01 | +0.02 | -0.02 | -0.03 |
| 75% | 0.359 | +0.04 | +0.01 | -0.07 | -0.03 |
| 100% | 0.302 | -0.02 | +0.03 | -0.05 | -0.03 |
| 125% ★ | 0.377 | +0.06 | -0.01 | -0.19 | -0.07 |
EmoNet emotion correlation
Fatigue/Exhaustion+0.51
Impatience and Irritability+0.45
Fear+0.45
Contentment-0.53
Contemplation-0.52
Affection-0.37
VoiceNet dimension correlation
Cognitive Load+0.62
Respiration+0.52
Tension+0.50
Recording Quality-0.87
Background Noise-0.74
Smoothness-0.73
Listen — 3 samples per dose
25%measured target: 0.54 · 0.38 · 0.47
50%measured target: 0.41 · 0.35 · 0.34
75%measured target: 0.24 · 0.27 · 0.21
100%measured target: 0.53 · 0.21 · 0.41
125% ★measured target: 0.31 · 0.55 · 0.19
Merging Disfluency __high drives the dimension: at the recommended 125% dose the measured Disfluency score moves from a 0.32 baseline to 0.38 (a +0.06 shift). The dose-response is mostly ordered but noisy between doses. That reach is a real cost — the biggest hit is to burst-blend (-0.19); across the board genuineness -0.01, burst-blend -0.19, speech quality -0.07. Emotionally it reads as Fatigue/Exhaustion (r 0.51), Impatience and Irritability (0.45) and Fear (0.45), and it moves away from Contentment, Contemplation and Affection. On other voice dimensions it co-moves most with Cognitive Load (0.62), Respiration (0.52) and Tension (0.50), while pulling down Recording Quality, Background Noise and Smoothness. Practical call: merge __high at 100–125%; expect Fatigue/Exhaustion and Cognitive Load to rise and burst-blend to fall. If you don’t want Fatigue/Exhaustion bleeding in, offset it with a prompt cue toward Contentment to keep the disfluency shift without the side-effect.
▼ LOW — Disfluency __low
Dose-response
| dose | target | Δtarget | Δgenu | Δblend | Δqual |
|---|
| baseline (0%) | 0.320 | — | — | — | — |
| 25% | 0.231 | -0.09 | -0.06 | -0.08 | +0.08 |
| 50% | 0.205 | -0.11 | -0.06 | -0.06 | +0.01 |
| 75% | 0.095 | -0.22 | -0.12 | -0.30 | +0.07 |
| 100% | 0.084 | -0.24 | -0.11 | -0.28 | +0.10 |
| 125% ★ | 0.083 | -0.24 | -0.14 | -0.31 | +0.10 |
EmoNet emotion correlation
Fatigue/Exhaustion+0.66
Relief+0.57
Longing+0.48
Emotional Numbness-0.42
Concentration-0.37
Embarrassment-0.27
VoiceNet dimension correlation
Genuineness+0.83
Cognitive Load+0.80
Perceived Gender+0.61
Structure-0.84
Background Noise-0.83
Smoothness-0.82
Listen — 3 samples per dose
25%measured target: 0.30 · 0.14 · 0.16
50%measured target: 0.14 · 0.17 · 0.21
75%measured target: 0.08 · 0.10 · 0.07
100%measured target: 0.08 · 0.10 · 0.12
125% ★measured target: 0.12 · 0.13 · 0.08
Merging Disfluency __low suppresses the dimension: at the recommended 125% dose the measured Disfluency score moves from a 0.32 baseline to 0.08 (a -0.24 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 (-0.31); across the board genuineness -0.14, burst-blend -0.31, speech quality +0.10. Emotionally it reads as Fatigue/Exhaustion (r 0.66), Relief (0.57) and Longing (0.48), and it moves away from Emotional Numbness, Concentration and Embarrassment. On other voice dimensions it co-moves most with Genuineness (0.83), Cognitive Load (0.80) and Perceived Gender (0.61), while pulling down Structure, Background Noise and Smoothness. Practical call: merge __low at 100–125%; expect Fatigue/Exhaustion and Genuineness to rise and burst-blend to fall. If you don’t want Fatigue/Exhaustion bleeding in, offset it with a prompt cue toward Emotional Numbness to keep the disfluency shift without the side-effect.