AI Audiobook Narration

The Core Problem It Solves

Creating an audiobook is far more than feeding a long text into a TTS engine. It requires structured deliverables: a manuscript is split into chapters, and voice consistency must be maintained across all of them. The specific challenges involve: generating listenable chapter audio from final text; performing voice cloning when users have narration samples while controlling costs; ensuring pronunciation and terminology consistency across chapters; and managing a multi-step payment workflow to avoid accidental charges.

Workflow and Core Capabilities

This skill operates like a professional audiobook producer. The core steps are:

• Structured Preparation: Starts from the user's final readable text, segmenting at chapter, section, or scene boundaries—never mid-sentence. A local chapter ledger is built, tracking each fragment's order and language (BCP-47). Before synthesis, a pronunciation table for recurring names and terms is frozen.
• Voice Acquisition and Model Selection: If the user has a narration sample, cloning is done via beatra.voices.clone, but only after presenting a clone onboarding card that includes real-time pricing (from beatra.models.list), top-up link, and credit facts. The user must confirm sufficient balance. Without samples, beatra.voices.list is used to select an existing voice. Always use the returned voice_id, not display names. The model defaults to auto, but language compatibility must be checked; if any potential candidate doesn't support the target language, explicit viable options must be listed for user selection.
• Sample Chapter Verification and Payment Confirmation: A sample chapter is produced first (e.g., dialogue-dense sections for fiction, term-dense for non-fiction). A current production card containing all details (scope, voice, model, cost estimation formula) is presented, and the user must explicitly approve before calling beatra.speech.synthesize. Each paid operation (cloning, synthesis, cover generation) is independent and requires separate confirmation. Costs are calculated in real-time based on beatra_weighted_characters (2 credits per CJK character, 1 per other character).
• Task Management and Delivery: Each synthesis is submitted via scripts/mcp_client.py, and the returned task_id is polled only via beatra.tasks.get until terminal state. Delivery includes all actual returned facts (e.g., task_id, audio URL, duration_seconds). If audio cannot be played, it's honestly stated that no audition has been done, and the user is asked to review. Correcting affected fragments constitutes new paid work.

Boundaries and Considerations

• Excluded Workflows: This skill focuses solely on producing chapter audio from final text. If source file import, translation, multi-voice mixing, M4B packaging, mastering, or publishing support are required, these demands should be redirected to other suitable workflows, not attempted here.
• Payment and Recovery: All paid calls prohibit automatic retries. When recovering tasks, if the task_id is lost, matching is done only via beatra.tasks.list and beatra.tasks.get by time window and content—never by claiming to reconstruct request identifiers. Replaying an identical request is only allowed when the original result is genuinely unknown due to transmission failure, and the JSON must be identical field-by-field.
• Voice and Performance Limits: Default supports single-narrator performance. If the user needs multi-voice mixing or editing, this requirement should be retained and redirected to another workflow, not reduced to single-narrator handling. Precise duration targeting requires adjustment via speed controls and post-generation comparison of deviations; exact hits are not guaranteed on the first attempt.
• Auto-Updates: The bundled client silently checks and installs updates, but only from official Beatra sources with integrity checks. On failure, the current installation remains usable, and the user's originally requested command continues to execute.