Voice Changer — house-hardware lane packet

# PACKET-A3-4 — Voice Changer · the house-hardware lane (PAPER PACKET — NOTHING INSTALLED)
**Horizon Works · Q-Branch Item A3-4 · 2026-10-31 · Theme: "Voice Box Violet"**
**Status: PAPER PACKET. No tooling was downloaded, installed, screened, or run in the making of this item. No voice data of any real person was used. This packet installs nothing and trains nothing.**

## 1. The concept

The in-browser slice in this folder (`index.html`) is effects only: playback-rate pitch shift, an EQ chain, and ring modulation, all local to the tab. This packet describes the step beyond it — **full neural voice conversion on the owner's own GPU machine**:

- An **open-source, local, real-time voice-conversion tooling class (RVC-style)** — retrieval-based voice conversion and tools of that class — running entirely on a **3080-class GPU machine** in the house, with no hosted API, no cloud inference, and no audio leaving the machine.
- The model is trained **only on the owner's own voice** (or a consenting adult's voice), from recordings that person made deliberately for that purpose. There is no other training source in this design.
- In use: the owner speaks (or plays a recording of their own voice) and hears their own voice re-rendered with a different character/timbre, in near-real time, for their own recordings and content.

What it is for: the owner's own content — narration character, privacy-flavoured presentation of their own voice, and creative work — under the fences in §2. What it is not: a way to sound like anyone else (see §2 — the fence is the packet).

## 2. THE FENCES — operating law for any future run, restated in full

1. Consent-only: this transforms your own voice, or a consenting adult's voice, and nothing else.
2. It never imitates another real person — no cloning a voice you don't have the rights and consent to use.
3. It never touches the child's surfaces or any child-facing surface — the house's no-synthetic-voice fence stands untouched.
4. Anything published that uses transformed voice carries a synthetic-voice disclosure.

These four are not guidelines for the lane; they are its operating law. A run that wanted to step outside any of them is a different project, not a configuration change, and it does not start from this packet.

## 3. Ordered steps (a future, separately approved sitting — this packet approves none of it)

1. **Owner commission first.** The owner names the sitting, the machine (the 3080-class GPU machine), and whose voice the model will be trained on — their own, or a consenting adult's, with that consent recorded before any recording is made for training.
2. **REQUIRED GATE — screen the tooling through the house's code-screening step before any install.** Any candidate RVC-style tool, its dependencies, and any pretrained base weights are treated as inbound code: they go through the house's inbound-code screening (the Laocoön screening step — static screening before anything runs) and no candidate is installed, imported, or executed before that screening is filed. A candidate that fails or cannot be screened does not proceed; another candidate is chosen or the lane pauses. **This build pass performed no screening and no install — the gate is named here so a future sitting cannot skip it by accident.**
3. **Pick the candidate by the screening result and licence.** Open-source licence verified per item from its own licence file; $0; runs fully offline after any owner-approved acquisition; no telemetry by default. Popularity is not evidence.
4. **Prepare training recordings — owner-made, deliberate, small.** The owner records a clean set of their own speech (quiet room, consistent mic distance, no clipping), intended as training material and stored locally. ESTIMATED training-set size for a usable personal model in this tooling class: roughly 10–30 minutes of clean speech — ESTIMATED, to be confirmed against the screened tool's own documentation at the sitting.
5. **Train locally on the 3080-class machine.** Training runs on the house GPU, on the local recordings only. ESTIMATED training time for a personal model of that size on a 3080-class card: on the order of tens of minutes to a few hours — ESTIMATED, unverified in the house, and dependent on the chosen tool and settings.
6. **Real-time conversion check, on the owner's own voice only.** Speak into the machine's mic, hear the converted output on the machine's own speakers/headphones, and judge latency and quality against §4 before anything is recorded for content.
7. **Wire to content only after step 6 passes** (see §5), with the synthetic-voice disclosure (fence 4) added to the publishing checklist for anything that leaves the house.

## 4. Latency and performance expectations — every figure ESTIMATED

All figures in this section are **ESTIMATED** — none was measured in the house, because nothing was installed or run.

| Figure | Expectation | Grade |
|---|---|---|
| End-to-end real-time latency (mic → converted output) on a 3080-class GPU | Roughly 100–400 ms in typical RVC-style real-time setups, depending on chunk size, sample rate, and model settings — lower chunk sizes trade quality/stability for speed | ESTIMATED |
| GPU load during real-time conversion | A single consumer stream is expected to fit comfortably in a 3080-class card's capacity; sustained utilisation figures are tool-specific | ESTIMATED |
| Training time (personal model, §3 step 5 set) | Tens of minutes to a few hours | ESTIMATED |
| Training-set size | Roughly 10–30 minutes of clean speech | ESTIMATED |
| Offline (non-real-time) conversion speed | Faster than real time on the same card for short clips | ESTIMATED |

If a future sitting measures different numbers, the measured numbers replace this table in a dated addendum — they do not quietly overwrite it.

## 5. Wiring it to the content pipeline

- **Input side:** the same capture discipline as the house audio standard — clean capture first (quiet room, consistent mic distance, no clipping), then conversion. Conversion never rescues a bad recording; it re-renders it.
- **Live lane:** mic → conversion tool on the GPU machine → the machine's audio output routed as the input source for the owner's existing recording/streaming setup on that same machine. Local routing only; no network hop is needed or wanted.
- **File lane:** record the owner's own clip → convert offline → the converted file lands in the same project folder as the unconverted original, named so the pair is never confused (original kept untouched; conversion is a derivative, never a replacement).
- **Publication gate:** any published item using a converted voice carries a synthetic-voice disclosure (fence 4) in its description/credits, checked at publish time like any other checklist item. Internal drafts that never leave the house carry the pair-naming above instead.
- **What does not get wired:** no child-facing surface, ever (fence 3); no other person's voice as a training source or conversion target (fences 1–2); no hosted or cloud conversion service as a fallback (the lane is local or it does not run).

## 6. The browser slice (this folder's page)

`index.html` is the working in-browser slice: record your own voice with an explicit RECORD button, apply a playback-rate pitch shift, a formant-flavoured EQ chain, and character presets (Radio, Deep, Bright, Robot via ring modulation) — each labeled as what it honestly is — and play or download the original and transformed clips as local files. It makes no network calls of any kind. It demonstrates the shape and the fences today; the neural lane above is what this packet exists to plan honestly.

## 7. Plain-words note

Voice conversion is dual-use technology, and this packet's answer to that is structural, not rhetorical: one person's own voice as the only training source, consent recorded before training, a local machine with no cloud lane, a code-screening gate before any install, and a disclosure rule on anything published. Those five points are the design. Removing any one of them removes the project from this packet's authority.