explainer-audio-cast-base

Files (4)

• SKILL.md
• pair-formal.md · references/
• pair-natural.md · references/
• pair-zundamon.md · references/

SKILL.md (first lines)

# Audio cast base

Owns the **cast pairs**: one self-contained file per pair under `references/`,
holding its two speakers, their voices, and how they behave toward each other.
There is no roster of casts anywhere else, and a voice id appears nowhere but in
its own pair file.

## The model

A **cast pair** names two characters and assigns the roles. The roles are fixed
for every pair:

- the **learner** is speaker `A` — the listener who has not understood yet;
- the **explainer** is speaker `B` — the one who has.

They differ in *understanding*, and that asymmetry is why a script is a dialogue
rather than a narration. A pair chooses voices, register, and the relationship
between the two. **It does not choose whether the learner can be wrong.** Every
pattern in
`explainer-audio-dialogue/references/conversational-act-patterns.md` requires the
learner's moves to expose an inference, doubt, demand, paraphrase, or revised
judgment; a cast that cannot do that cannot hold the learner role, whatever its
voice. Each pair file records which patterns it performs most naturally — a tilt,
not a restriction: the pattern is chosen per act from that act's obstacle.

## Pairs

| Pair | Register | Fits |
|---|---|---|
| `zundamon` *(default)* | 常体・キャラクター調 | 入門的な原典、広く共有された誤解が障害になる題材、短めのガイド |
| `natural` | 敬体・番組的な距離感 | 実務的な原典、practitioner 向け、長さのあるガイド |
| `formal` | 硬めの敬体・端正 | 規格や論文のように原典自身が厳密な題材、expert 向け |

Each pair's file — `references/pair-<id>.md` — is the complete definition:
speakers, personas, relationship, one worked conversational act, and the spoken
orientation in that register. **Read only the selected pair's file.** They are
deliberately self-contained so that writing one pair never drags another pair's
register into the draft.

Each file opens with the machine-readable binding that narration reads:

```text
cast-binding learner   <cast-id> <voicevox-style-id>
cast-binding explainer <cast-id> <voicevox-style-id>
```

## Choosing a pair

The Guide plan's `cast_pair` policy value records the choice. There is no rule
table mapping policy values to pairs, and none should be added: this is a
judgment. Weigh these.

- **Who is listening.** A `newcomer` audience is served by a learner who says the
  wrong thing plainly and is corrected without ceremony. An `expert` audience
  hears that as condescension; there the learner's error should be a matter of
  precision, not of basic misunderstanding.
- **What the source sounds like.** A standard, a specification, or a paper whose
  own register is exact loses credibility when explained through a casual pair.
  An introductory or practical source survives, and often benefits from, one.
- **How wrong the listener is likely to be.** When the obstacle is a widespread
  misconception, prefer a pair whose learner states things outright, so the wrong
  idea is built and visibly broken. When the material is subtle rather than
  counter-intuitive, prefer a precise learner, so the correction lands as a
  refinement.
- **How long the guide is.** A strong character register carries a short guide
  and tires over a long one; a plainer pair sustains length better.

When these pull apart, the audience decides. Record the reason in the Guide plan
so the next revision is not a coin flip.

Voices available to the local engine are listed by
`curl http://127.0.0.1:50021/speakers`.

From choplin/explainer-studio, MIT.

More from choplin/explainer-studio