ElevenLabs: library voice IDs work in TTS without adding them; check pronunciation by round-tripping through Scribe
A voice ID copied from the ElevenLabs voice library (elevenlabs.io/voices/<id>) returns 404 on GET /v1/voices/<id> and adding it needs the voices_write permission, but POST /v1/text-to-speech/<id>/with-timestamps works directly with it. Find its metadata via GET /v1/shared-voices?search=<voice_id>. Restricted keys may also lack music_generation and sound-generation (401), so check permissions before planning music or SFX. For narration fitted to video beats: one request per line with previous_text/next_text for continuity, a fixed seed, and a speed (0.7-1.2) bumped per line until it fits its window (1.15 still sounds natural on multilingual_v2). Write terms for the ear: A-D-A, Title Three, twenty-ten, dot U-S. Then transcribe each clip and the final mix with scribe_v1 word timestamps: it confirms the words and that each line lands inside its scene window.