← all dimensions · markdown (agent skill) · manual home

Attack VoiceNet dimension guide

Physical Productioncode vn_ATCKslot 60
This dimension measures the onset character as sound initiates from silence. 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.

Attack controls the onset character as sound initiates from silence. The __high LoRA shifts it +0.61 at 125% (worst cost: burst-blend -0.36); the __low LoRA shifts it -0.02 at 50% (worst cost: speech quality -0.10). The __high direction is the stronger, more reliable knob here.
▲ HIGH — Attack __high
Dose-response
dosetargetΔtargetΔgenuΔblendΔqual
baseline (0%)0.079————
25%0.151+0.07-0.04-0.15+0.01
50%0.510+0.43-0.08-0.34-0.07
75%0.492+0.41-0.05-0.33-0.05
100%0.596+0.52-0.06-0.34-0.17
125% ★0.686+0.61-0.05-0.36-0.21
EmoNet emotion correlation
Impatience and Irritability+0.62
Teasing+0.55
Anger+0.55
Contemplation-0.74
Contentment-0.69
Relief-0.69
VoiceNet dimension correlation
Arousal+0.96
Tension+0.95
Metallic Character+0.94
Monologue Style-0.84
Warmth-0.83
Whisper-Talk Style-0.79
Listen — 3 samples per dose
25%measured target: 0.04 · 0.03 · 0.15
50%measured target: 0.18 · 0.74 · 0.82
75%measured target: 0.68 · 0.71 · 0.31
100%measured target: 0.49 · 0.22 · 0.57
125% ★measured target: 0.72 · 0.67 · 0.75
Merging Attack __high lifts the dimension: at the recommended 125% dose the measured Attack score moves from a 0.08 baseline to 0.69 (a +0.61 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 (-0.36); across the board genuineness -0.05, burst-blend -0.36, speech quality -0.21. Emotionally it reads as Impatience and Irritability (r 0.62), Teasing (0.55) and Anger (0.55), and it moves away from Contemplation, Contentment and Relief. On other voice dimensions it co-moves most with Arousal (0.96), Tension (0.95) and Metallic Character (0.94), while pulling down Monologue Style, Warmth and Whisper-Talk Style. Practical call: merge __high at 100–125%; expect Impatience and Irritability and Arousal to rise and burst-blend to fall. If you don’t want Impatience and Irritability bleeding in, offset it with a prompt cue toward Contemplation to keep the attack shift without the side-effect.
▼ LOW — Attack __low
Dose-response
dosetargetΔtargetΔgenuΔblendΔqual
baseline (0%)0.079————
25%0.127+0.05-0.02+0.06-0.01
50% ★0.058-0.02+0.03+0.04-0.10
75%0.063-0.02-0.04-0.06-0.03
100%0.159+0.08-0.07-0.02+0.11
125%0.213+0.13-0.09-0.06+0.08
EmoNet emotion correlation
Pride+0.34
Malevolence/Malice+0.30
Teasing+0.23
Disgust-0.42
Relief-0.39
Contentment-0.22
VoiceNet dimension correlation
Stance+0.89
Arousal+0.88
Metallic Character+0.87
Vulnerability-0.40
Genuineness-0.40
Casual Style-0.37
Listen — 3 samples per dose
25%measured target: 0.00 · 0.11 · 0.01
50% ★measured target: 0.00 · 0.02 · 0.00
75%measured target: 0.00 · 0.00 · 0.00
100%measured target: 0.07 · 0.01 · 0.10
125%measured target: 0.15 · 0.23 · 0.25
Merging Attack __low pulls down the dimension: at the recommended 50% dose the measured Attack score moves from a 0.08 baseline to 0.06 (a -0.02 shift). The dose-response is noisy and non-monotonic — the merge is an unreliable knob here. That reach is a modest cost — the biggest hit is to speech quality (-0.10); across the board genuineness +0.03, burst-blend +0.04, speech quality -0.10. Emotionally it reads as Pride (r 0.34), Malevolence/Malice (0.30) and Teasing (0.23), and it moves away from Disgust, Relief and Contentment. On other voice dimensions it co-moves most with Stance (0.89), Arousal (0.88) and Metallic Character (0.87), while pulling down Vulnerability, Genuineness and Casual Style. Practical call: merge __low at 25–75%; expect Pride and Stance to rise and speech quality to fall. If you don’t want Pride bleeding in, offset it with a prompt cue toward Disgust to keep the attack shift without the side-effect.
Data: MOSS-VA-v2 base + per-dimension LoRA, neutral text, no reference · 5 merge doses × 8 scored samples/dose · correlations are Pearson r over the sample pool · scaling grid · LoRA repo