AI Short Video Voiceover

The Specific Problem It Solves

Converting scripts into natural voiceover audio for short videos involves multiple challenges. Scripts often need adaptation for read-aloud format, matching platform requirements, audience tone, and language pronunciation. Traditional dubbing workflows suffer from opaque pricing, model compatibility issues, and inefficient voice selection, leading to repeated trials and delays.

Core Capabilities and Key Steps

This skill automates the dubbing process using AI models, driven by a production card for controlled execution.

• Determining Minimal Production Requirements: Reuses user-provided scripts and background, supplementing only essential choices like final script, platform use, audience tone, target language (BCP-47 code), and pronunciation notes. Scripts are adapted for clear presentation and read-aloud suitability, with user approval required for meaning changes.
• Voice and Model Selection: Calls beatra.voices.list to fetch voice candidates, recommending based on preview_url and compatible_models. Verifies model online status, language support, and compatibility via beatra.models.list with capability: "text_to_speech". Defaults to model: "auto" but resolves all candidates and confirms language support.
• Pricing and Confirmation: Applies beatra_weighted_characters rule (weight 2 for Chinese characters, 1 for others) with real-time pricing to generate a production card showing exact script, voice_id, model, language, format, sample rate, speed, volume, pitch, and estimated credits. Paid operation beatra.speech.synthesize is executed only after user approval.
• Execution and Polling: Uses bundled scripts/mcp_client.py client via CLI arguments and stdin JSON to invoke the synthesis API. Creates an opaque client_request_id for each approved request, records task_id immediately, and polls with beatra.tasks.get until terminal state.
• Recovery and Duplicate Avoidance: If interrupted, retrieves tasks via beatra.tasks.get or beatra.tasks.list by capability and time window, preventing duplicate paid audio generation. Replays same client_request_id only when original result is unknown due to transmission errors, with identical JSON.
• Delivering Facts: On success, returns facts like task_id, audio URL, artifact_id, duration_seconds, actual mime_type, sample rate, etc. Compares duration targets and reports deviations; delivery is audio-only without video, subtitles, or digital human synthesis.

Applicable Boundaries and Considerations

• Delivery Scope: Focuses on audio generation, not video creation, lip-sync, publishing, or voice cloning. For full videos, retain requirements as a sub-task or hand off to video workflows.
• Duration and Pricing: Duration targets are review references, not guarantees. Pricing is based on weighted characters and real-time rules, requiring user confirmation in the production card.
• Operation Constraints: All Beatra operations must use specified scripts/mcp_client.py, not configuring other connectors. Paid calls require user approval, with no automatic retries or replays.
• Task Management: Tasks are canceled only upon user request with beatra.tasks.cancel. Recovery relies on factual matching, not claiming remote ID correspondence.
• Update Mechanism: Auto-updates run silently in the background without disrupting operations, rolling back on failure while continuing user requests.