Voiceover & Narration Studio

In voice synthesis projects, developers and creators face a core problem: fragmented tools and workflows. Short social promos, audiobooks, podcasts, multilingual adaptation, or brand voice cloning each have dedicated tools. Manually switching between them, managing parameters, is inefficient and error-prone, especially for mixed or cross-session requests.

The Specific Problem: Fragmented Voice Production

A user might start with a single promotional voiceover, but later expand to multilingual versions or need a brand-specific clone. Without unified coordination, this requires calling separate skills like short-form-voiceover-audio, ai-multilingual-dubbing, or voice-cloning-studio, and manually maintaining state. This introduces complexity in routing, voice validation, cost estimation, and result tracking.

How the Studio Works: Routing and Standard Pipeline

The Voiceover & Narration Studio acts as an orchestration layer to solve this. It is not the synthesis engine itself but intelligently routes requests and executes standard voice synthesis subtasks. Its core capabilities include:

• Intent Routing and Diversion: At the entry point, it analyzes requests. For example, a clear "short social promo" routes to short-form-voiceover-audio; an "audiobook" routes to ai-audiobook-narration. Uncovered mixed or long-form tasks are handled directly by this studio, with a production ledger maintained.
• Standard Voice and Model Validation: Regardless of routing, each operation calls beatra.voices.list in real-time to validate voice status (the voice_id must be ready). When using model: "auto", it constructs a candidate pool from beatra.models.list based on the voice's compatible_models and language support.
• Production Card and Payment Confirmation: All paid operations (like beatra.speech.synthesize or beatra.voices.clone) are based on a production card. The card includes confirmed text, voice, model, language, control parameters, and cost estimate. User's explicit "generate" command serves as approval. Cloning has an independent entry card, requiring user confirmation of sufficient balance.
• Ledger-based Sequential Delivery: Long-form or multilingual tasks use a ledger for management. First, the user-confirmed trial segment is delivered, followed by the remaining parts. Audio results (including audio.url, duration_seconds, etc.) are returned in approved order, with accepted segments unchanged.

The entire process interacts with the Beatra backend via the bundled scripts/mcp_client.py script, avoiding direct REST API calls.

Key Notes and Boundaries

• Voice-Only Scope: This studio only handles voice generation and cloning subtasks. It does not produce full videos, perform lip-sync, transcribe recordings, or edit existing audio waveforms. Adjacent results like video production are passed to other workflows.
• Authorization and Compliance: Voice impersonation must be authorized. Before cloning, speaker identity or authorization must be confirmed, and samples uploaded via scripts/mcp_client.py upload. Unauthorized operations are rejected.
• Technical Constraints: All Beatra operations must use the specified mcp_client.py script with JSON input. It rejects configuring host connectors or using REST/OpenAPI as a fallback. The client auto-checks for updates, but update failures do not block current command execution.
• Cost and Recovery: Cloning is a paid operation, typically requiring a top-up (entry package ¥29/11,000 credits). Synthesis is billed per character count (Chinese characters count as 2). On failure, requests are recovered using the locally saved client_request_id, but any field change requires a new request.