Generates an AI voiceover from a script, entirely on-device. The WAV lands in the project's media library as a normal audio asset (with waveform and analysis), the script is stored as asset metadata, and a script-corrected word-level transcript is written so captions work immediately. CALL THIS BEFORE YOU CUT THE PICTURE. A script does not tell you how long it takes to SAY, so a narrated edit assembled first is built on guessed cut points that the real audio then breaks — every section, graphic and transition has to be retimed. Generate the narration, read its word timings with get_transcript, then lay picture against those frames. GENERATE IN SEGMENTS, NOT ONE BLOCK — call this 2-4 times (hook / body / close, or one per section) and lay the parts on the timeline. A single long generation comes out FLAT and evenly paced, everything pressed together at one energy; segmenting gives each part its own intent, puts natural air at the joins, and lets you redo one section instead of all of it. Write for the ear: short sentences, one idea each, LINE BREAKS between them — the strongest signal the model has to actually stop. See read_skill {topic:'voiceover'}. Async: returns {jobId, fileName} — fileName is the audio file's final name in the project's media/ folder, decided up front (WAV for local providers, MP3 for ElevenLabs). The job runs on the analysis MASTER tab and its request is persisted on disk: it survives tab closes and browser restarts, and get_job_status answers from disk even in a fresh session. Poll until done; the finished job also carries the asset's mediaRef. Voices/languages come from list_voiceover_models; a local model must be downloaded first (phase 'ready'), otherwise this returns PERMISSION_REQUIRED. PREMIUM (provider 'elevenlabs' — SPENDS THE USER'S CREDITS): check list_voiceover_models' elevenlabs status object first. Structured refusals you must handle: NO_API_KEY (no key stored — ask the USER to add their key in the Voiceover generator's ElevenLabs onboarding; never ask for, or pass, the key yourself), INSUFFICIENT_CREDITS (carries needed vs remaining — shorten the script or ask the user), CONSENT_REQUIRED (paid features are set to 'ask': a consent dialog with the cost is now open in Frapea and the error carries an actionId — TELL THE USER in your visible reply to confirm it, call await_user_action with that actionId, and on 'granted' retry this call with consentActionId set to it; 'denied' means drop it). AUDIO TAGS (ElevenLabs): inline bracket tags like [calm], [excited], [whispers] direct the delivery ONLY on models whose list_voiceover_models entry says supportsAudioTags: true (the v3 family). Every other model reads them ALOUD — so only write tags when the chosen model supports them. As a safety net, tags sent to a non-supporting model are stripped before synthesis and the success message says so.