# Qurmtessence voice profile

Qurmtessence now supports an explicit delivery profile and redacted continuity memory.

## Hosted voice

The hosted site uses an allowlisted built-in speech voice. Choose a delivery style in the Qurmtessence **Voice & adaptation profile** panel:

- **Natural & expressive** — human cadence, warmth, and clear emphasis.
- **Warm & encouraging** — patient and supportive without flattery.
- **Elevated & commanding** — calm, decisive, and deliberate.
- **Crisp & concise** — high-signal and efficient.

The selected profile is stored in this browser and sent as bounded preferences. It cannot grant access, execute an action, or bypass QI approval.

## Continuity memory

When enabled, signed-in turns may be redacted and embedded for continuity. Secrets and credential-like values are filtered before storage. Turn the memory checkbox off to stop syncing new turns; use **Clear conversation** to remove the local transcript. Server retention remains subject to the configured QI memory policy.

## Christopher voice (private companion only)

A custom voice is not uploaded or cloned by the public browser page. To use Christopher’s voice, the approved local companion must provide a `speakLocal` bridge and read a private reference file named `voice/christopher-reference.wav`.

1. Record a clean, single-speaker reference sample (about 10 seconds, no music or other voices).
2. Keep the WAV on the approved local generation machine; never commit it or place it in `public/`.
3. Keep the owner consent record with the private asset. The browser checkbox only records the owner’s explicit choice to use that local profile; it does not upload audio.
4. Configure the local companion to reject the request unless the reference file and consent record are present. If the bridge is unavailable, Qurmtessence safely falls back to the selected hosted voice.

This separation keeps a personal voice private while still allowing the conversation style and continuity features to work on the hosted site.

## Start the private companion

The workspace includes a loopback-only reference implementation in the
private sibling folder `qurmmius-voice-companion/` (outside the Netlify site
directory). On the approved local generation machine:

```powershell
cd ..\qurmmius-voice-companion
python -m pip install chatterbox-tts soundfile torch
Copy-Item consent.example.json consent.json
# Edit consent.json so it records the owner authorization and signing date.
# Place the private recording at .\christopher-reference.wav.
$env:QURMTESSENCE_VOICE_CONSENT = 'owner-consented-local-reference-v1'
python companion.py
```

Keep the terminal running while using the hosted Qurmtessence page. In the
**Voice & adaptation profile** panel, choose **Check private companion**. A
ready result means the local model, recording, and consent are present. The
browser sends only the text to `127.0.0.1:8790`; the reference recording stays
on this machine. If the companion is stopped or unavailable, the page falls
back to the selected hosted voice.

The first Chatterbox run downloads its model locally (about 1 GB) and can take
a few minutes. Generated audio is cached under `qurmmius-voice-companion/cache/`
and stays outside the site deployment.