# Chunking — VoiceNet dimension guide

*Category:* Rhythm & Timing · *code:* `vn_CHNK` · *slot:* 41

This dimension measures how words are grouped into breath units and how often the speaker pauses. Merge the `__high` / `__low` LoRA from `laion/moss-voicenet-dimension-loras` at 25/50/75/100/125%. Costs are shifts vs. the 24-sample neutral baseline.

**TL;DR.** **Chunking** controls how words are grouped into breath units and how often the speaker pauses. The **__high** LoRA shifts it +0.28 at 125% (worst cost: burst-blend -0.22); the **__low** LoRA shifts it +0.00 at 50% (worst cost: burst-blend -0.15). The **__high** direction is the stronger, more reliable knob here.

## ▲ HIGH — Chunking __high

| dose | target | Δtarget | Δgenu | Δblend | Δqual |
|---|---|---|---|---|---|
| baseline (0%) | 0.158 | — | — | — | — |
| 25% | 0.188 | +0.03 | -0.02 | -0.06 | +0.02 |
| 50% | 0.274 | +0.12 | -0.04 | +0.01 | +0.06 |
| 75% | 0.348 | +0.19 | -0.07 | -0.10 | +0.08 |
| 100% | 0.417 | +0.26 | -0.12 | -0.22 | +0.06 |
| 125% ★ | 0.434 | +0.28 | -0.06 | -0.22 | +0.04 |

- **EmoNet emotion** — ↑ Hope/Enthusiasm/Optimism (+0.40), Elation (+0.27), Amusement (+0.18); ↓ Helplessness (-0.39), Awe (-0.39), Fear (-0.38)
- **VoiceNet dimension** — ↑ Brightness (+0.90), Arousal Shift (+0.80), Smoothness (+0.79); ↓ Cognitive Load (-0.61), Disfluency (-0.57), Vulnerability (-0.55)

Merging **Chunking __high** raises the dimension: at the recommended **125%** dose the measured Chunking score moves from a 0.16 baseline to 0.43 (a +0.28 shift). The dose-response is clean and monotonic — each step pushes the dimension further. That reach is a real cost — the biggest hit is to **burst-blend** (-0.22); across the board genuineness -0.06, burst-blend -0.22, speech quality +0.04. Emotionally it reads as Hope/Enthusiasm/Optimism (r 0.40), Elation (0.27) and Amusement (0.18), and it moves away from Helplessness, Awe and Fear. On other voice dimensions it co-moves most with Brightness (0.90), Arousal Shift (0.80) and Smoothness (0.79), while pulling down Cognitive Load, Disfluency and Vulnerability. Practical call: merge __high at **100–125%**; expect Hope/Enthusiasm/Optimism and Brightness to rise and burst-blend to fall. If you don’t want Hope/Enthusiasm/Optimism bleeding in, offset it with a prompt cue toward Helplessness to keep the chunking shift without the side-effect.

## ▼ LOW — Chunking __low

| dose | target | Δtarget | Δgenu | Δblend | Δqual |
|---|---|---|---|---|---|
| baseline (0%) | 0.158 | — | — | — | — |
| 25% | 0.198 | +0.04 | -0.06 | -0.15 | +0.03 |
| 50% ★ | 0.156 | +0.00 | +0.00 | -0.15 | -0.10 |
| 75% | 0.172 | +0.01 | +0.00 | -0.18 | -0.16 |
| 100% | 0.281 | +0.12 | -0.10 | -0.20 | -0.01 |
| 125% | 0.265 | +0.11 | -0.05 | -0.01 | +0.05 |

- **EmoNet emotion** — ↑ Concentration (+0.28), Sourness (+0.25), Bitterness (+0.25); ↓ Pain (-0.39), Fatigue/Exhaustion (-0.37), Sadness (-0.36)
- **VoiceNet dimension** — ↑ Brightness (+0.76), Arousal Shift (+0.75), Structure (+0.71); ↓ Disfluency (-0.68), Vulnerability (-0.59), Genuineness (-0.51)

Merging **Chunking __low** lowers the dimension: at the recommended **50%** dose the measured Chunking score moves from a 0.16 baseline to 0.16 (a +0.00 shift). The dose-response is mostly ordered but noisy between doses. That reach is a modest cost — the biggest hit is to **burst-blend** (-0.15); across the board genuineness +0.00, burst-blend -0.15, speech quality -0.10. Emotionally it reads as Concentration (r 0.28), Sourness (0.25) and Bitterness (0.25), and it moves away from Pain, Fatigue/Exhaustion and Sadness. On other voice dimensions it co-moves most with Brightness (0.76), Arousal Shift (0.75) and Structure (0.71), while pulling down Disfluency, Vulnerability and Genuineness. Practical call: merge __low at **25–75%**; expect Concentration and Brightness to rise and burst-blend to fall. If you don’t want Concentration bleeding in, offset it with a prompt cue toward Pain to keep the chunking shift without the side-effect.

---
[← all dimensions](index.md) · [human/audio version](../../site/dims/CHNK.html)