Skip to content

beatra-ai/beatra-skills / plugins/ai-podcast-voiceover/skills/ai-podcast-voiceover

ai-podcast-voiceover

Verified

Turn an article, notes, or a finished script into a listener-ready solo podcast episode with a consistent host voice. This AI podcast voice generator and AI podcast narration service adapts supplied material into a speakable podcast script, sets names and specialist terms for clear pronunciation, and creates MP3 podcast audio with natural pacing. Use this podcast voiceover AI and text-to-speech podcast service for article-to-podcast audio, news briefings, expert commentary, and knowledge shows, then carry the host direction into the next episode.

references/episode-script.md

View on GitHub
# Episode script

## Start from supplied material

This workflow turns an article, notes, an outline, a finished script, or a
named topic into a single-host episode. Reuse material and show decisions
already present in the conversation before asking for more.

| What the user supplied | What to do |
| --- | --- |
| An article, notes, an outline, or a finished script | Every fact, opinion, and number comes from that material or from an edit the user confirmed. |
| Only an evergreen topic: a method, a concept, a practice, or an explainer | Write a complete draft from general, widely accepted knowledge. Do not put a specific number, date, personal name, company name, quotation, or case in the spoken script. Where the point needs one of those facts, put the fact on the please-check list. |
| Only a time-sensitive topic: news, policy, a price, an event, or a recent development about a person or company | Do not write facts. When the host can search the web, list 3-5 candidate sources and wait for the user to choose. The host does not choose. Treat the chosen sources as supplied material. When the host cannot search, ask the user to paste the material. In the same turn, write the opening, the structure, the transitions, and the close. |

For a health, parenting, legal, or personal-finance topic, write only cautious
general guidance, and add one sentence that the listener's own situation needs
a professional opinion.

If a supplied link or file cannot be read, stop, say that it could not be
read, and ask the user to paste the text. Do not change the topic and do not
substitute a source. A time-sensitive topic does not override this stop: a
link or file the user already supplied is not a reason to search for a
replacement.

Do not leave a blank, a placeholder, or a bracketed gap in the spoken script.
In the review copy, mark every specific fact in the script with its origin,
supplied material or confirmed by the user, and mark every sentence of general
background as general knowledge to check. The synthesis text itself carries
no marks, so nothing extra is read aloud. Keep every unconfirmed specific fact
on the please-check list, outside the spoken script.

## Set the listener focus

Use the intended listener from the show profile when one exists. If supplied
material could credibly become different episodes for different listeners,
ask which listener and takeaway should guide the spoken emphasis. A topic
alone never waits on this question: infer the listener and focus from the
request and the show, write the draft in the same turn, and mark the inference
for approval; ask first only when the user's own words name two different
audiences. Otherwise infer one short episode focus from the supplied material
and mark it for approval with the draft. A listener
focus can change ordering, transitions, and how much context the host gives;
it never authorizes new facts, claims, examples, or positions. How much
background to give also follows `listener_level` on the show profile.

## Shape a solo episode

Build the episode in this order:

1. A cold-open hook.
2. A one-sentence route that says what this episode covers.
3. Two or three body sections, with a hook back or a short recap about every
   3-5 minutes.
4. A recap of what the episode already said.
5. One action the listener can take.
6. A next-episode tease when the show profile already has a next episode to
   name. The tease adds no new fact.

Speak to the listener directly. Use a signpost sentence at each new point,
such as "The second point is...". Keep each sentence short enough to say in
one breath. For Chinese, keep a sentence to about 40 characters, one
speakable line. Write numbers in the form the host should say aloud.

The recap only restates points the episode has already made. It does not draw
a further conclusion, forecast, or introduce a new fact.

## Make prose easy to hear

Preserve the meaning while adapting the material for a listener. When the
draft comes from a topic, apply the same spoken shape to that draft.

- split long written sentences into short, natural spoken thoughts;
- replace page-only constructions with spoken phrasing;
- write numbers, dates, units, URLs, citations, and abbreviations in the form
  the host should say aloud;
- add a short transition wherever the listener needs the connection between
  sections; and
- shape the user's material into an opening hook, a clear sequence of sections,
  and the show's recurring closer when it has one.

Names, terms, figures, and stated positions stay intact unless the user
approves a change. A target duration is a review goal rather than a promise:
the returned audio duration is authoritative.

### Length tiers

The character ranges below are estimates, from about 225 characters per
minute, which is also an estimate. Use that rate on the first pass only.
After a sample returns, estimate later episodes with `words_per_minute` on
the show profile. Measuring that rate is in
[voice, delivery, and recovery](voice-and-delivery.md).

| Tier | Duration | Estimated characters |
| --- | --- | --- |
| Bulletin | 3-5 minutes | 700-1100 |
| Standard | 8-12 minutes | 1800-2700 |
| In-depth | 15-20 minutes | 3400-4500 |

When the episode language is Chinese, say the tier by the name in this list:

```text
Bulletin: 快讯
Standard: 标准
In-depth: 深度
```

When the user does not name a tier, size a supplied-material episode to the
material. A topic-only request defaults to Bulletin, an estimated 3-5
minutes, and say that a longer tier is available.

A duration written into the show's name is not a tier choice. Only the user's
own request sets the tier.

When the material cannot carry the chosen tier, say first that the material
on hand comes to about N minutes, then offer three choices: accept the
shorter cut, add material, or add general background marked in the review
copy as general knowledge to check.

## Titles, description, notes, and chapters

Deliver these with the script review and again with the audio. They add no
charge:

- three title options;
- an episode description of two or three sentences, and show notes as a list
  of the points, with the chosen sources when the topic is time-sensitive;
- chapter timestamps. At draft time, estimate them in proportion to each
  section's character count and label them as estimates. After the audio
  returns, re-time them in proportion to the returned duration.

## Approve the spoken draft

Present the listener focus and episode script section by section, with the
titles, description, show notes, and estimated chapters. For multi-host shows or
guest conversations, attribute every turn explicitly (`speaker`, spoken text,
and optional delivery cue) so scenes can pass `beatra.speech.validate_dialogue`
cleanly. Confirm the title, opening, closer, substantive claims, and any
pronunciations that affect the read before proposing paid synthesis. On the
speech production card, state how many specific facts are still on the
please-check list. When the user wants only a speakable script, stop at the
approved script and create no audio.

Build a pronunciation list for proper nouns, foreign words, acronyms, and
polyphonic characters. Record each written term and the spoken form the user
confirms, then carry it forward in the show profile.

If `beatra.speech.validate` or synthesis refuses the text as too long, follow the
refusal message: split at natural topic or section boundaries and submit the parts
in order. If a dialogue request is refused as too long, split it at scene or turn
boundaries as the refusal message says. Never split mid-sentence or force a
listener through an abrupt cut.

## Smart Voice Director and free aligned subtitles

When natural phrasing, pause holds, dramatic beats, or pronunciation
compilation needs active performance direction, run paid
`beatra.speech.prepare_script` (`{"capability": "speech_script"}`) on plain
spoken text plus `instructions` rather than guessing or
hand-writing braced performance markup (never invent braced tags). Smart Voice
Director preserves spoken words untouched and returns `directed` plus
`scripts[]`; verify every ready `scripts[].synthesis` payload with free
`beatra.speech.validate` before confirming synthesis.

Keep `subtitle_enable: true` (default; never pass `subtitles`) on
`beatra.speech.synthesize` and `beatra.speech.synthesize_dialogue`. Beatra
returns `output.subtitles` (JSON, SRT, VTT, and dialogue `segments[]` with
`turn_index`, `speaker`, `start_seconds`, `end_seconds`) free of charge, which
supports exact chapter re-timing from returned durations and segment timing
without paying for a separate transcription call.
slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack."Stars: 198kUpdated Claude CodeCodexVerifiedInstall
ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read.Stars: 196kUpdated Claude CodeCodexVerifiedInstall
ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report.Stars: 196kUpdated Claude CodeCodexVerifiedInstall
ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt.Stars: 196kUpdated Claude CodeCodexVerifiedInstall