AI Voiceover for Short Videos

Turn a final short-form script into ready-to-edit narration audio. Work as a voiceover director: preserve the intended claim and hook, make the words speakable, freeze one viable voice and model path, and generate only the exact read the user approves.

Establish the smallest complete brief

Reuse everything already supplied. Obtain only missing choices that materially change the audio:

• the final script, or permission to make a clearly shown speakable edit;
• the platform or use, audience, and tone when they guide the voice direction;
• the target BCP-47 language or dialect and important pronunciations;
• a duration target when useful, treating it as a review goal rather than a guarantee; and
• the output format or other delivery requirement when the default MP3 is not suitable.

Platform context guides the voiceover; it is not permission to create or publish a video. If the user wants only a material script rewrite, show the spoken draft, obtain approval when its meaning or claims changed, and stop at the approved script without creating paid audio. Follow voiceover workflow for script preparation, timing, casting, pricing, recovery, and handoff.

By default, prepare one logical short read and keep one synthesis input at or below 50,000 characters. Split longer text only at natural sentence or section boundaries. Do not invent extra hooks, alternate takes, or A/B variants. Each variant the user requests is a separate paid request whose exact text and count must appear in the production card.

Freeze a real voice and viable model path

If no voice is frozen, call beatra.voices.list and offer a small set of relevant voices with their returned previews. Treat each returned voice_id as an opaque value and record its preview_url and compatible_models; never use a display name or prose preference as synthesis voice.

Call beatra.models.list with capability: "text_to_speech" before deciding a model, checking language support, or estimating cost. An explicit model must be live, available, in the frozen voice's compatible_models, and support the requested language. For auto, first form the complete set of live available voice-compatible candidates. Language is validated after voice-first auto resolution, so compare the valid BCP-47 request's primary language with every candidate's constraints.supported_languages, using only documented aliases. Keep auto only when every potential candidate supports the language; otherwise present viable explicit models and ask the user to confirm one.

Unless the user or destination requires another supported setting, use model: "auto", format: "mp3", speed: 1.0, volume: 1.0, pitch: 0, no emotion, and no explicit sample rate.

Price and confirm the exact read

Planning, voice previews, script preparation, and estimates are free. beatra.speech.synthesize is paid. Count the exact final text using the live beatra_weighted_characters rule: each Han ideograph weighs 2 and every other character weighs 1. Apply each applicable live card's estimate_formula, unit_price_credits, and scale. If auto can resolve to more than one candidate, show the full range or a clearly labeled upper bound.

Before any paid call, present one production card containing:

• exact approved script or segment and every approved variant;
• opaque voice_id, model or auto candidate range, BCP-47 language, format, sample rate if set, speed, volume, pitch, and emotion if set;
• weighted-character count, live formula, unit price or range, and estimated credits;
• exact paid-request count; and
• the requested duration target, clearly labeled as not guaranteed.

One explicit approval of that exact current card is sufficient; do not ask for a duplicate confirmation. Any change to script, segment, variant count, voice, model, language, format, or another control creates a new card and requires new approval and a new request identity. Never automatically retry a paid call.

Execute each approved request once

Use only the bundled scripts/mcp_client.py for Beatra operations. Pass the tool name as the CLI argument and its JSON arguments on stdin. Do not configure or call a host Beatra Connector and do not use REST/OpenAPI as a fallback. For exact commands and troubleshooting, use Bundled MCP Client diagnostics.

For each newly approved execution, create a new opaque client_request_id and call python3 scripts/mcp_client.py call beatra.speech.synthesize once with the frozen JSON on stdin:

{
  "voice": "<opaque voice_id>",
  "input": "<exact approved script or segment>",
  "client_request_id": "<new opaque id>",
  "model": "auto",
  "language": "<BCP-47 tag>",
  "format": "mp3",
  "speed": 1.0,
  "volume": 1.0,
  "pitch": 0
}

Record the returned task_id immediately and poll only that task with beatra.tasks.get until it is terminal. A queued or running task is not a reason to submit again.

Recover without duplicating paid audio

When task_id is known, use beatra.tasks.get. If it was lost, use beatra.tasks.list with capability: "text_to_speech", follow next_cursor through every relevant page, and inspect plausible candidates with beatra.tasks.get. The remote list/get envelope does not expose the locally stored client_request_id; match by capability, time window, returned input, settings, and other returned task facts without claiming remote ID proof.

Only when transport or task creation left the original outcome genuinely unknown may the same client_request_id be replayed, and then only with JSON that is field-for-field identical. Any changed field is a new paid execution with a new card, approval, and ID. A known task, slow polling, authorization or update trouble, and terminal failure never authorize an automatic replacement.

Use beatra.tasks.cancel only when the user requests cancellation. If the request conflicts, continue reconciling the same task and do not promise a stop or refund.

Deliver only returned facts

On success, return every available actual fact: task_id, audio URL, artifact_id, duration_seconds, actual mime_type, actual sample rate, task.links.assets, resolved model, usage, and billing. Compare returned duration_seconds with any target and report drift. If the host can play the audio, review pronunciation, pauses, clarity, energy, and format; otherwise say it was not auditioned and ask the user to review it.

On failure or cancellation, report only returned provider-neutral TaskError facts such as code, message, retryable, param, and details. Billing fields may be absent or unsettled: distinguish missing values from numeric zero and never invent a charge, refund, or net result. Any adjustment is separately confirmed paid work.

The deliverable is audio only: not video creation, captions, avatars, lip-sync, or publishing. Do not promise exact duration. For a complete video, preserve the platform, duration, and script requirements and offer this voiceover as the audio subtask or route to an appropriate video workflow. For a custom cloned voice, preserve the sample and consent requirements and route to the voice cloning workflow.

Installation and account operations

For first use and shared operations, follow installation and authentication, installation registration, tasks and results, billing, errors, and recovery, and uninstall and disconnect.

Updates and security

This Skill runs only the files installed with it and leaves them unchanged while it runs; it never downloads or executes other code. It talks to Beatra's official services over HTTPS, and a file you choose to upload goes only to the one-time upload address Beatra issues for it.

Once a day at most, the bundled client reads Beatra's public version list. When the store or tool you installed from has a newer version, it prints one line with the command that updates it there. Updating stays your choice, through that store or tool. See updates and security.