# AutoShorts Chatterbox premium voice testing protocol

Session learning from ERF-017 voice QA.

## Trigger

Use this when the user says the narration sounds unnatural, especially when sentence-ending emphasis/cadence sounds wrong.

## What worked

The accepted default voice for the user's AutoShorts/Everyday Red Flags videos is:

```json
{
  "voice_mode": "clone",
  "reference_voice": "Gianna.wav",
  "label": "Gianna clone premium v1",
  "temperature": 0.80,
  "exaggeration": 0.52,
  "cfg_weight": 0.50,
  "speed_factor": 1.0,
  "pause_ms": 500,
  "chunking": "warm narrator chunks; avoid over-hard sentence endings"
}
```

## Audio-only test workflow

Before re-rendering a full short:

1. Render short audio-only probes from the same script text.
2. Include at least:
   - Taylor continuity: lower exaggeration, longer chunks.
   - Taylor warm narrator: warmer but not overdramatic.
   - Alternate voice sanity check.
   - Premium/reference clone test.
3. Deliver the `.ogg` probes for user selection.
4. Only after approval, re-render the video and create fresh approval hashes.

## Pitfalls

- Do not fix awkward sentence endings only by slowing the voice down.
- Avoid too many short sentence chunks; they make Chatterbox perform each sentence ending like a hard stop.
- A reference clone is not automatically better; compare clone vs predefined when testing a new voice.
- Once the user explicitly accepts a voice as default, update both persistent memory and the rendering skill/reference notes.